init

.gitignore

@@ -0,0 +1,3 @@
+**/__pycache__
+.DS_Store
+

README.md

@@ -0,0 +1,8 @@
+
+
+
+Linux
+```
+apt-get install freeglut3-dev
+apt-get install libosmesa6-dev
+```

__init__.py

@@ -0,0 +1,29 @@
+import os
+
+REPO_PATH = os.path.dirname(os.path.abspath(__file__))
+
+from folder_paths import models_dir
+
+SMPL_PATH = os.path.join(models_dir, 'smpl')
+os.makedirs(SMPL_PATH, exist_ok=True)
+
+
+from .nodes.process_humans_node import ProcessHumansNode
+from .nodes.load_detectron_node import LoadDetectronNode
+from .nodes.load_hmr_node import LoadHMRNode
+# from .nodes.select_human_node import SelectHumanNode
+
+
+NODE_CLASS_MAPPINGS = {
+    'ProcessHumans': ProcessHumansNode,
+    'LoadDetectron': LoadDetectronNode,
+    'LoadHMR': LoadHMRNode,
+    # 'SelectHuman': SelectHumanNode,
+}
+
+NODE_DISPLAY_NAME_MAPPINGS = {
+    'ProcessHumans': 'Process 4D Humans',
+    'LoadDetectron': 'Load Detectron Model',
+    'LoadHMR': 'Load HMR Model',
+    # 'SelectHuman': 'Select 4D Human'
+}

configs/cascade_mask_rcnn_vitdet_h_75ep.py

@@ -0,0 +1,129 @@
+## coco_loader_lsj.py
+
+import detectron2.data.transforms as T
+from detectron2 import model_zoo
+from detectron2.config import LazyCall as L
+
+# Data using LSJ
+image_size = 1024
+dataloader = model_zoo.get_config("common/data/coco.py").dataloader
+dataloader.train.mapper.augmentations = [
+    L(T.RandomFlip)(horizontal=True),  # flip first
+    L(T.ResizeScale)(
+        min_scale=0.1, max_scale=2.0, target_height=image_size, target_width=image_size
+    ),
+    L(T.FixedSizeCrop)(crop_size=(image_size, image_size), pad=False),
+]
+dataloader.train.mapper.image_format = "RGB"
+dataloader.train.total_batch_size = 64
+# recompute boxes due to cropping
+dataloader.train.mapper.recompute_boxes = True
+
+dataloader.test.mapper.augmentations = [
+    L(T.ResizeShortestEdge)(short_edge_length=image_size, max_size=image_size),
+]
+
+from functools import partial
+from fvcore.common.param_scheduler import MultiStepParamScheduler
+
+from detectron2 import model_zoo
+from detectron2.config import LazyCall as L
+from detectron2.solver import WarmupParamScheduler
+from detectron2.modeling.backbone.vit import get_vit_lr_decay_rate
+
+# mask_rcnn_vitdet_b_100ep.py
+
+model = model_zoo.get_config("common/models/mask_rcnn_vitdet.py").model
+
+# Initialization and trainer settings
+train = model_zoo.get_config("common/train.py").train
+train.amp.enabled = True
+train.ddp.fp16_compression = True
+train.init_checkpoint = "detectron2://ImageNetPretrained/MAE/mae_pretrain_vit_base.pth"
+
+
+# Schedule
+# 100 ep = 184375 iters * 64 images/iter / 118000 images/ep
+train.max_iter = 184375
+
+lr_multiplier = L(WarmupParamScheduler)(
+    scheduler=L(MultiStepParamScheduler)(
+        values=[1.0, 0.1, 0.01],
+        milestones=[163889, 177546],
+        num_updates=train.max_iter,
+    ),
+    warmup_length=250 / train.max_iter,
+    warmup_factor=0.001,
+)
+
+# Optimizer
+optimizer = model_zoo.get_config("common/optim.py").AdamW
+optimizer.params.lr_factor_func = partial(get_vit_lr_decay_rate, num_layers=12, lr_decay_rate=0.7)
+optimizer.params.overrides = {"pos_embed": {"weight_decay": 0.0}}
+
+# cascade_mask_rcnn_vitdet_b_100ep.py
+
+from detectron2.config import LazyCall as L
+from detectron2.layers import ShapeSpec
+from detectron2.modeling.box_regression import Box2BoxTransform
+from detectron2.modeling.matcher import Matcher
+from detectron2.modeling.roi_heads import (
+    FastRCNNOutputLayers,
+    FastRCNNConvFCHead,
+    CascadeROIHeads,
+)
+
+# arguments that don't exist for Cascade R-CNN
+[model.roi_heads.pop(k) for k in ["box_head", "box_predictor", "proposal_matcher"]]
+
+model.roi_heads.update(
+    _target_=CascadeROIHeads,
+    box_heads=[
+        L(FastRCNNConvFCHead)(
+            input_shape=ShapeSpec(channels=256, height=7, width=7),
+            conv_dims=[256, 256, 256, 256],
+            fc_dims=[1024],
+            conv_norm="LN",
+        )
+        for _ in range(3)
+    ],
+    box_predictors=[
+        L(FastRCNNOutputLayers)(
+            input_shape=ShapeSpec(channels=1024),
+            test_score_thresh=0.05,
+            box2box_transform=L(Box2BoxTransform)(weights=(w1, w1, w2, w2)),
+            cls_agnostic_bbox_reg=True,
+            num_classes="${...num_classes}",
+        )
+        for (w1, w2) in [(10, 5), (20, 10), (30, 15)]
+    ],
+    proposal_matchers=[
+        L(Matcher)(thresholds=[th], labels=[0, 1], allow_low_quality_matches=False)
+        for th in [0.5, 0.6, 0.7]
+    ],
+)
+
+# cascade_mask_rcnn_vitdet_h_75ep.py
+
+from functools import partial
+
+train.init_checkpoint = "detectron2://ImageNetPretrained/MAE/mae_pretrain_vit_huge_p14to16.pth"
+
+model.backbone.net.embed_dim = 1280
+model.backbone.net.depth = 32
+model.backbone.net.num_heads = 16
+model.backbone.net.drop_path_rate = 0.5
+# 7, 15, 23, 31 for global attention
+model.backbone.net.window_block_indexes = (
+    list(range(0, 7)) + list(range(8, 15)) + list(range(16, 23)) + list(range(24, 31))
+)
+
+optimizer.params.lr_factor_func = partial(get_vit_lr_decay_rate, lr_decay_rate=0.9, num_layers=32)
+optimizer.params.overrides = {}
+optimizer.params.weight_decay_norm = None
+
+train.max_iter = train.max_iter * 3 // 4  # 100ep -> 75ep
+lr_multiplier.scheduler.milestones = [
+    milestone * 3 // 4 for milestone in lr_multiplier.scheduler.milestones
+]
+lr_multiplier.scheduler.num_updates = train.max_iter

configs/model_config.yaml

@@ -0,0 +1,148 @@
+task_name: train
+tags:
+- dev
+train: true
+test: false
+ckpt_path: null
+seed: null
+DATASETS:
+  SUPPRESS_KP_CONF_THRESH: 0.3
+  FILTER_NUM_KP: 4
+  FILTER_NUM_KP_THRESH: 0.0
+  FILTER_REPROJ_THRESH: 31000
+  SUPPRESS_BETAS_THRESH: 3.0
+  SUPPRESS_BAD_POSES: true
+  POSES_BETAS_SIMULTANEOUS: true
+  FILTER_NO_POSES: false
+  TRAIN:
+    H36M-TRAIN-WMASK:
+      WEIGHT: 0.1
+    MPII-TRAIN-WMASK:
+      WEIGHT: 0.1
+    COCO-TRAIN-2014-WMASK-PRUNED:
+      WEIGHT: 0.1
+    COCO-TRAIN-2014-VITPOSE-REPLICATE-PRUNED12:
+      WEIGHT: 0.1
+    MPI-INF-TRAIN-PRUNED:
+      WEIGHT: 0.02
+    AVA-TRAIN-MIDFRAMES-1FPS-WMASK:
+      WEIGHT: 0.19
+    AIC-TRAIN-WMASK:
+      WEIGHT: 0.19
+    INSTA-TRAIN-WMASK:
+      WEIGHT: 0.2
+  VAL:
+    COCO-VAL:
+      WEIGHT: 1.0
+  MOCAP: CMU-MOCAP
+  CONFIG:
+    SCALE_FACTOR: 0.3
+    ROT_FACTOR: 30
+    TRANS_FACTOR: 0.02
+    COLOR_SCALE: 0.2
+    ROT_AUG_RATE: 0.6
+    TRANS_AUG_RATE: 0.5
+    DO_FLIP: true
+    FLIP_AUG_RATE: 0.5
+    EXTREME_CROP_AUG_RATE: 0.1
+    EXTREME_CROP_AUG_LEVEL: 1
+  BETAS_REG: true
+trainer:
+  _target_: pytorch_lightning.Trainer
+  default_root_dir: ${paths.output_dir}
+  accelerator: gpu
+  devices: 8
+  deterministic: false
+  num_sanity_val_steps: 0
+  log_every_n_steps: ${GENERAL.LOG_STEPS}
+  val_check_interval: ${GENERAL.VAL_STEPS}
+  precision: 16
+  max_steps: ${GENERAL.TOTAL_STEPS}
+  move_metrics_to_cpu: true
+  limit_val_batches: 1
+  track_grad_norm: -1
+  strategy: ddp
+  num_nodes: 1
+  sync_batchnorm: true
+paths:
+  root_dir: ${oc.env:PROJECT_ROOT}
+  data_dir: ${paths.root_dir}/data/
+  log_dir: /fsx/shubham/code/hmr2023/logs_hydra/
+  output_dir: ${hydra:runtime.output_dir}
+  work_dir: ${hydra:runtime.cwd}
+extras:
+  ignore_warnings: false
+  enforce_tags: true
+  print_config: true
+exp_name: hmr2
+SMPL:
+  MODEL_PATH: data/smpl
+  GENDER: neutral
+  NUM_BODY_JOINTS: 23
+  JOINT_REGRESSOR_EXTRA: data/SMPL_to_J19.pkl
+  MEAN_PARAMS: data/smpl_mean_params.npz
+EXTRA:
+  FOCAL_LENGTH: 5000
+  NUM_LOG_IMAGES: 4
+  NUM_LOG_SAMPLES_PER_IMAGE: 8
+  PELVIS_IND: 39
+MODEL:
+  IMAGE_SIZE: 256
+  IMAGE_MEAN:
+  - 0.485
+  - 0.456
+  - 0.406
+  IMAGE_STD:
+  - 0.229
+  - 0.224
+  - 0.225
+  BACKBONE:
+    TYPE: vit
+    NUM_LAYERS: 50
+    OUT_CHANNELS: 2048
+  ADD_NECK: false
+  FLOW:
+    DIM: 144
+    NUM_LAYERS: 4
+    CONTEXT_FEATURES: 2048
+    LAYER_HIDDEN_FEATURES: 1024
+    LAYER_DEPTH: 2
+  FC_HEAD:
+    NUM_FEATURES: 1024
+  SMPL_HEAD:
+    TYPE: transformer_decoder
+    IN_CHANNELS: 2048
+    TRANSFORMER_DECODER:
+      depth: 6
+      heads: 8
+      mlp_dim: 1024
+      dim_head: 64
+      dropout: 0.0
+      emb_dropout: 0.0
+      norm: layer
+      context_dim: 1280
+GENERAL:
+  TOTAL_STEPS: 1000000
+  LOG_STEPS: 1000
+  VAL_STEPS: 1000
+  CHECKPOINT_STEPS: 10000
+  CHECKPOINT_SAVE_TOP_K: 1
+  NUM_WORKERS: 6
+  PREFETCH_FACTOR: 2
+TRAIN:
+  LR: 1.0e-05
+  WEIGHT_DECAY: 0.0001
+  BATCH_SIZE: 48
+  LOSS_REDUCTION: mean
+  NUM_TRAIN_SAMPLES: 2
+  NUM_TEST_SAMPLES: 64
+  POSE_2D_NOISE_RATIO: 0.01
+  SMPL_PARAM_NOISE_RATIO: 0.005
+LOSS_WEIGHTS:
+  KEYPOINTS_3D: 0.05
+  KEYPOINTS_2D: 0.01
+  GLOBAL_ORIENT: 0.001
+  BODY_POSE: 0.001
+  BETAS: 0.0005
+  ADVERSARIAL: 0.0005
+local: {}

humans4d/.gitignore

@@ -0,0 +1,148 @@
+# Specific
+/logs*/
+/results/
+/sandbox/
+*.lock
+*.pt
+*.npy
+/example_data/downloaded*
+*.tar
+*.tar.gz
+/discord_sandbox/
+/demo_out/
+token_channel.csv
+/hmr2_training_data/
+/hmr2_evaluation_data/
+/outputs/
+
+# Byte-compiled / optimized / DLL files
+__pycache__/
+*.py[cod]
+*$py.class
+
+# C extensions
+*.so
+
+# Distribution / packaging
+.Python
+build/
+develop-eggs/
+dist/
+downloads/
+eggs/
+.eggs/
+lib/
+lib64/
+parts/
+sdist/
+var/
+wheels/
+pip-wheel-metadata/
+share/python-wheels/
+*.egg-info/
+.installed.cfg
+*.egg
+MANIFEST
+
+# PyInstaller
+#  Usually these files are written by a python script from a template
+#  before PyInstaller builds the exe, so as to inject date/other infos into it.
+*.manifest
+*.spec
+
+# Installer logs
+pip-log.txt
+pip-delete-this-directory.txt
+
+# Unit test / coverage reports
+htmlcov/
+.tox/
+.nox/
+.coverage
+.coverage.*
+.cache
+nosetests.xml
+coverage.xml
+*.cover
+*.py,cover
+.hypothesis/
+.pytest_cache/
+
+# Translations
+*.mo
+*.pot
+
+# Django stuff:
+*.log
+local_settings.py
+db.sqlite3
+db.sqlite3-journal
+
+# Flask stuff:
+instance/
+.webassets-cache
+
+# Scrapy stuff:
+.scrapy
+
+# Sphinx documentation
+docs/_build/
+
+# PyBuilder
+target/
+
+# Jupyter Notebook
+.ipynb_checkpoints
+
+# IPython
+profile_default/
+ipython_config.py
+
+# pyenv
+.python-version
+
+# pipenv
+#   According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.
+#   However, in case of collaboration, if having platform-specific dependencies or dependencies
+#   having no cross-platform support, pipenv may install dependencies that don't work, or not
+#   install all needed dependencies.
+#Pipfile.lock
+
+# PEP 582; used by e.g. github.com/David-OConnor/pyflow
+__pypackages__/
+
+# Celery stuff
+celerybeat-schedule
+celerybeat.pid
+
+# SageMath parsed files
+*.sage.py
+
+# Environments
+.env
+.venv
+env/
+venv/
+ENV/
+env.bak/
+venv.bak/
+
+# Spyder project settings
+.spyderproject
+.spyproject
+
+# Rope project settings
+.ropeproject
+
+# mkdocs documentation
+/site
+
+# mypy
+.mypy_cache/
+.dmypy.json
+dmypy.json
+
+# Pyre type checker
+.pyre/
+/checkpoints/
+/data/

humans4d/LICENSE.md

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2023 UC Regents, Shubham Goel
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

humans4d/README.md

@@ -0,0 +1,104 @@
+# 4DHumans: Reconstructing and Tracking Humans with Transformers
+Code repository for the paper:
+**Humans in 4D: Reconstructing and Tracking Humans with Transformers**
+[Shubham Goel](https://people.eecs.berkeley.edu/~shubham-goel/), [Georgios Pavlakos](https://geopavlakos.github.io/), [Jathushan Rajasegaran](http://people.eecs.berkeley.edu/~jathushan/), [Angjoo Kanazawa](https://people.eecs.berkeley.edu/~kanazawa/)<sup>\*</sup>, [Jitendra Malik](http://people.eecs.berkeley.edu/~malik/)<sup>\*</sup>
+
+[![arXiv](https://img.shields.io/badge/arXiv-2305.20091-00ff00.svg)](https://arxiv.org/pdf/2305.20091.pdf)  [![Website shields.io](https://img.shields.io/website-up-down-green-red/http/shields.io.svg)](https://shubham-goel.github.io/4dhumans/)     [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1Ex4gE5v1bPR3evfhtG7sDHxQGsWwNwby?usp=sharing)  [![Hugging Face Spaces](https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Spaces-blue)](https://huggingface.co/spaces/brjathu/HMR2.0)
+
+
+![teaser](assets/teaser.png)
+
+## Installation and Setup
+First, clone the repo. Then, we recommend creating a clean [conda](https://docs.conda.io/) environment, installing all dependencies, and finally activating the environment, as follows:
+```bash
+git clone https://github.com/shubham-goel/4D-Humans.git
+cd 4D-Humans
+conda env create -f environment.yml
+conda activate 4D-humans
+```
+
+If conda is too slow, you can use pip:
+```bash
+conda create --name 4D-humans python=3.10
+conda activate 4D-humans
+pip install torch
+pip install -e .[all]
+```
+
+All checkpoints and data will automatically be downloaded to `$HOME/.cache/4DHumans` the first time you run the demo code.
+
+Besides these files, you also need to download the *SMPL* model. You will need the [neutral model](http://smplify.is.tue.mpg.de) for training and running the demo code. Please go to the corresponding website and register to get access to the downloads section. Download the model and place `basicModel_neutral_lbs_10_207_0_v1.0.0.pkl` in `./data/`.
+
+## Run demo on images
+The following command will run ViTDet and HMR2.0 on all images in the specified `--img_folder`, and save renderings of the reconstructions in `--out_folder`. `--batch_size` batches the images together for faster processing. The `--side_view` flags additionally renders the side view of the reconstructed mesh, `--full_frame` renders all people together in front view, `--save_mesh` saves meshes as `.obj`s.
+```bash
+python demo.py \
+    --img_folder example_data/images \
+    --out_folder demo_out \
+    --batch_size=48 --side_view --save_mesh --full_frame
+```
+
+## Run tracking demo on videos
+Our tracker builds on PHALP, please install that first:
+```bash
+pip install git+https://github.com/brjathu/PHALP.git
+```
+
+Now, run `track.py` to reconstruct and track humans in any video. Input video source may be a video file, a folder of frames, or a youtube link:
+```bash
+# Run on video file
+python track.py video.source="example_data/videos/gymnasts.mp4"
+
+# Run on extracted frames
+python track.py video.source="/path/to/frames_folder/"
+
+# Run on a youtube link (depends on pytube working properly)
+python track.py video.source=\'"https://www.youtube.com/watch?v=xEH_5T9jMVU"\'
+```
+The output directory (`./outputs` by default) will contain a video rendering of the tracklets and a `.pkl` file containing the tracklets with 3D pose and shape. Please see the [PHALP](https://github.com/brjathu/PHALP) repository for details.
+
+## Training
+Download the [training data](https://www.dropbox.com/sh/mjdwu59fxuhls5h/AACQ6FCGSrggUXmRzuubRHXIa) to `./hmr2_training_data/`, then start training using the following command:
+```
+bash fetch_training_data.sh
+python train.py exp_name=hmr2 data=mix_all experiment=hmr_vit_transformer trainer=gpu launcher=local
+```
+Checkpoints and logs will be saved to `./logs/`. We trained on 8 A100 GPUs for 7 days using PyTorch 1.13.1 and PyTorch-Lightning 1.8.1 with CUDA 11.6 on a Linux system. You may adjust batch size and number of GPUs per your convenience.
+
+## Evaluation
+Download the [evaluation metadata](https://www.dropbox.com/scl/fi/kl79djemdgqcl6d691er7/hmr2_evaluation_data.tar.gz?rlkey=ttmbdu3x5etxwqqyzwk581zjl) to `./hmr2_evaluation_data/`. Additionally, download the Human3.6M, 3DPW, LSP-Extended, COCO, and PoseTrack dataset images and update the corresponding paths in  `hmr2/configs/datasets_eval.yaml`.
+
+Run evaluation on multiple datasets as follows, results are stored in `results/eval_regression.csv`. 
+```bash
+python eval.py --dataset 'H36M-VAL-P2,3DPW-TEST,LSP-EXTENDED,POSETRACK-VAL,COCO-VAL' 
+```
+
+By default, our code uses the released checkpoint (mentioned as HMR2.0b in the paper). To use the HMR2.0a checkpoint, you may download and untar from [here](https://people.eecs.berkeley.edu/~jathushan/projects/4dhumans/hmr2a_model.tar.gz)
+
+## Preprocess code
+To preprocess LSP Extended and Posetrack into metadata zip files for evaluation, see `hmr2/datasets/preprocess`.
+
+Training data preprocessing coming soon.
+
+## Acknowledgements
+Parts of the code are taken or adapted from the following repos:
+- [ProHMR](https://github.com/nkolot/ProHMR)
+- [SPIN](https://github.com/nkolot/SPIN)
+- [SMPLify-X](https://github.com/vchoutas/smplify-x)
+- [HMR](https://github.com/akanazawa/hmr)
+- [ViTPose](https://github.com/ViTAE-Transformer/ViTPose)
+- [Detectron2](https://github.com/facebookresearch/detectron2)
+
+Additionally, we thank [StabilityAI](https://stability.ai/) for a generous compute grant that enabled this work.
+
+## Citing
+If you find this code useful for your research, please consider citing the following paper:
+
+```bibtex
+@inproceedings{goel2023humans,
+    title={Humans in 4{D}: Reconstructing and Tracking Humans with Transformers},
+    author={Goel, Shubham and Pavlakos, Georgios and Rajasegaran, Jathushan and Kanazawa, Angjoo and Malik, Jitendra},
+    booktitle={ICCV},
+    year={2023}
+}
+```

humans4d/demo.py

@@ -0,0 +1,171 @@
+from pathlib import Path
+import torch
+import argparse
+import os
+import cv2
+import numpy as np
+
+from hmr2.configs import CACHE_DIR_4DHUMANS
+from hmr2.models import HMR2, download_models, load_hmr2, DEFAULT_CHECKPOINT
+from hmr2.utils import recursive_to
+from hmr2.datasets.vitdet_dataset import ViTDetDataset, DEFAULT_MEAN, DEFAULT_STD
+from hmr2.utils.renderer import Renderer, cam_crop_to_full
+
+LIGHT_BLUE=(0.65098039,  0.74117647,  0.85882353)
+
+def main():
+    import time
+    start = time.time()
+    parser = argparse.ArgumentParser(description='HMR2 demo code')
+    parser.add_argument('--checkpoint', type=str, default=DEFAULT_CHECKPOINT, help='Path to pretrained model checkpoint')
+    parser.add_argument('--img_folder', type=str, default='example_data/images', help='Folder with input images')
+    parser.add_argument('--out_folder', type=str, default='demo_out', help='Output folder to save rendered results')
+    parser.add_argument('--side_view', dest='side_view', action='store_true', default=False, help='If set, render side view also')
+    parser.add_argument('--top_view', dest='top_view', action='store_true', default=False, help='If set, render top view also')
+    parser.add_argument('--full_frame', dest='full_frame', action='store_true', default=False, help='If set, render all people together also')
+    parser.add_argument('--save_mesh', dest='save_mesh', action='store_true', default=False, help='If set, save meshes to disk also')
+    parser.add_argument('--detector', type=str, default='vitdet', choices=['vitdet', 'regnety'], help='Using regnety improves runtime')
+    parser.add_argument('--batch_size', type=int, default=1, help='Batch size for inference/fitting')
+    parser.add_argument('--file_type', nargs='+', default=['*.jpg', '*.png'], help='List of file extensions to consider')
+
+    args = parser.parse_args()
+
+    # Download and load checkpoints
+    download_models(CACHE_DIR_4DHUMANS)
+    model, model_cfg = load_hmr2(args.checkpoint)
+
+    # Setup HMR2.0 model
+    device = torch.device('cuda') if torch.cuda.is_available() else torch.device('cpu')
+    model = model.to(device)
+    model.eval()
+
+    # Load detector
+    from hmr2.utils.utils_detectron2 import DefaultPredictor_Lazy
+    if args.detector == 'vitdet':
+        from detectron2.config import LazyConfig
+        import hmr2
+        cfg_path = Path(hmr2.__file__).parent/'configs'/'cascade_mask_rcnn_vitdet_h_75ep.py'
+        detectron2_cfg = LazyConfig.load(str(cfg_path))
+        detectron2_cfg.train.init_checkpoint = "https://dl.fbaipublicfiles.com/detectron2/ViTDet/COCO/cascade_mask_rcnn_vitdet_h/f328730692/model_final_f05665.pkl"
+        for i in range(3):
+            detectron2_cfg.model.roi_heads.box_predictors[i].test_score_thresh = 0.25
+        detector = DefaultPredictor_Lazy(detectron2_cfg)
+    elif args.detector == 'regnety':
+        from detectron2 import model_zoo
+        from detectron2.config import get_cfg
+        detectron2_cfg = model_zoo.get_config('new_baselines/mask_rcnn_regnety_4gf_dds_FPN_400ep_LSJ.py', trained=True)
+        detectron2_cfg.model.roi_heads.box_predictor.test_score_thresh = 0.5
+        detectron2_cfg.model.roi_heads.box_predictor.test_nms_thresh   = 0.4
+        detector       = DefaultPredictor_Lazy(detectron2_cfg)
+
+    # Setup the renderer
+    renderer = Renderer(model_cfg, faces=model.smpl.faces)
+
+    # Make output directory if it does not exist
+    os.makedirs(args.out_folder, exist_ok=True)
+
+    # Get all demo images that end with .jpg or .png
+    img_paths = [img for end in args.file_type for img in Path(args.img_folder).glob(end)]
+
+    # Iterate over all images in folder
+    for img_path in img_paths:
+        img_cv2 = cv2.imread(str(img_path))
+
+        # Detect humans in image
+        det_out = detector(img_cv2)
+
+        det_instances = det_out['instances']
+        valid_idx = (det_instances.pred_classes==0) & (det_instances.scores > 0.5)
+        boxes=det_instances.pred_boxes.tensor[valid_idx].cpu().numpy()
+
+        # Run HMR2.0 on all detected humans
+        dataset = ViTDetDataset(model_cfg, img_cv2, boxes)
+        dataloader = torch.utils.data.DataLoader(dataset, batch_size=8, shuffle=False, num_workers=0)
+
+        all_verts = []
+        all_cam_t = []
+        
+        for batch in dataloader:
+            batch = recursive_to(batch, device)
+            with torch.no_grad():
+                out = model(batch)
+
+            pred_cam = out['pred_cam']
+            box_center = batch["box_center"].float()
+            box_size = batch["box_size"].float()
+            img_size = batch["img_size"].float()
+            scaled_focal_length = model_cfg.EXTRA.FOCAL_LENGTH / model_cfg.MODEL.IMAGE_SIZE * img_size.max()
+            pred_cam_t_full = cam_crop_to_full(pred_cam, box_center, box_size, img_size, scaled_focal_length).detach().cpu().numpy()
+
+            # Render the result
+            batch_size = batch['img'].shape[0]
+            for n in range(batch_size):
+                # Get filename from path img_path
+                img_fn, _ = os.path.splitext(os.path.basename(img_path))
+                person_id = int(batch['personid'][n])
+                white_img = (torch.ones_like(batch['img'][n]).cpu() - DEFAULT_MEAN[:,None,None]/255) / (DEFAULT_STD[:,None,None]/255)
+                input_patch = batch['img'][n].cpu() * (DEFAULT_STD[:,None,None]/255) + (DEFAULT_MEAN[:,None,None]/255)
+                input_patch = input_patch.permute(1,2,0).numpy()
+
+                regression_img = renderer(out['pred_vertices'][n].detach().cpu().numpy(),
+                                        out['pred_cam_t'][n].detach().cpu().numpy(),
+                                        batch['img'][n],
+                                        mesh_base_color=LIGHT_BLUE,
+                                        scene_bg_color=(1, 1, 1),
+                                        )
+
+                final_img = np.concatenate([input_patch, regression_img], axis=1)
+
+                if args.side_view:
+                    side_img = renderer(out['pred_vertices'][n].detach().cpu().numpy(),
+                                            out['pred_cam_t'][n].detach().cpu().numpy(),
+                                            white_img,
+                                            mesh_base_color=LIGHT_BLUE,
+                                            scene_bg_color=(1, 1, 1),
+                                            side_view=True)
+                    final_img = np.concatenate([final_img, side_img], axis=1)
+
+                if args.top_view:
+                    top_img = renderer(out['pred_vertices'][n].detach().cpu().numpy(),
+                                            out['pred_cam_t'][n].detach().cpu().numpy(),
+                                            white_img,
+                                            mesh_base_color=LIGHT_BLUE,
+                                            scene_bg_color=(1, 1, 1),
+                                            top_view=True)
+                    final_img = np.concatenate([final_img, top_img], axis=1)
+
+                cv2.imwrite(os.path.join(args.out_folder, f'{img_fn}_{person_id}.png'), 255*final_img[:, :, ::-1])
+
+                # Add all verts and cams to list
+                verts = out['pred_vertices'][n].detach().cpu().numpy()
+                cam_t = pred_cam_t_full[n]
+                all_verts.append(verts)
+                all_cam_t.append(cam_t)
+
+                # Save all meshes to disk
+                if args.save_mesh:
+                    camera_translation = cam_t.copy()
+                    tmesh = renderer.vertices_to_trimesh(verts, camera_translation, LIGHT_BLUE)
+                    tmesh.export(os.path.join(args.out_folder, f'{img_fn}_{person_id}.obj'))
+
+        # Render front view
+        if args.full_frame and len(all_verts) > 0:
+            misc_args = dict(
+                mesh_base_color=LIGHT_BLUE,
+                scene_bg_color=(1, 1, 1),
+                focal_length=scaled_focal_length,
+            )
+            cam_view = renderer.render_rgba_multiple(all_verts, cam_t=all_cam_t, render_res=img_size[n], **misc_args)
+
+            # Overlay image
+            input_img = img_cv2.astype(np.float32)[:,:,::-1]/255.0
+            input_img = np.concatenate([input_img, np.ones_like(input_img[:,:,:1])], axis=2) # Add alpha channel
+            input_img_overlay = input_img[:,:,:3] * (1-cam_view[:,:,3:]) + cam_view[:,:,:3] * cam_view[:,:,3:]
+
+            cv2.imwrite(os.path.join(args.out_folder, f'{img_fn}_all.png'), 255*input_img_overlay[:, :, ::-1])
+
+        end = time.time()
+        print(end - start)
+
+if __name__ == '__main__':
+    main()

humans4d/environment.yml

@@ -0,0 +1,33 @@
+name: 4D-humans
+channels:
+  - pytorch
+  - nvidia
+  - conda-forge
+dependencies:
+  - python=3.10
+  - numpy
+  - pytorch
+  - pytorch-cuda=11.8
+  - torchvision
+  - pip
+  - pip:
+    - pytorch-lightning
+    - smplx==0.1.28
+    - pyrender
+    - opencv-python
+    - yacs
+    - scikit-image
+    - einops
+    - timm
+    - dill
+    - pandas
+    - git+https://github.com/facebookresearch/detectron2
+    - git+https://github.com/mattloper/chumpy
+
+    # For training
+    - rich
+    - hydra-core
+    - hydra-submitit-launcher
+    - hydra-colorlog
+    - pyrootutils
+    - webdataset

humans4d/eval.py

@@ -0,0 +1,131 @@
+import argparse
+import os
+from pathlib import Path
+import traceback
+from typing import Optional
+
+import pandas as pd
+import torch
+from filelock import FileLock
+from hmr2.configs import dataset_eval_config
+from hmr2.datasets import create_dataset
+from hmr2.utils import Evaluator, recursive_to
+from tqdm import tqdm
+
+from hmr2.configs import CACHE_DIR_4DHUMANS
+from hmr2.models import HMR2, download_models, load_hmr2, DEFAULT_CHECKPOINT
+
+def main():
+    parser = argparse.ArgumentParser(description='Evaluate trained models')
+    parser.add_argument('--checkpoint', type=str, default=DEFAULT_CHECKPOINT, help='Path to pretrained model checkpoint')
+    parser.add_argument('--results_file', type=str, default='results/eval_regression.csv', help='Path to results file.')
+    parser.add_argument('--dataset', type=str, default='H36M-VAL-P2,3DPW-TEST,LSP-EXTENDED,POSETRACK-VAL,COCO-VAL', help='Dataset to evaluate') # choices=['H36M-VAL-P2', '3DPW-TEST', 'MPI-INF-TEST']
+    parser.add_argument('--batch_size', type=int, default=16, help='Batch size for inference')
+    parser.add_argument('--num_samples', type=int, default=1, help='Number of test samples to draw')
+    parser.add_argument('--num_workers', type=int, default=8, help='Number of workers used for data loading')
+    parser.add_argument('--log_freq', type=int, default=10, help='How often to log results')
+    parser.add_argument('--shuffle', dest='shuffle', action='store_true', default=False, help='Shuffle the dataset during evaluation')
+    parser.add_argument('--exp_name', type=str, default=None, help='Experiment name')
+
+    args = parser.parse_args()
+
+    # Download and load checkpoints
+    download_models(CACHE_DIR_4DHUMANS)
+    model, model_cfg = load_hmr2(args.checkpoint)
+
+    # Setup HMR2.0 model
+    device = torch.device('cuda') if torch.cuda.is_available() else torch.device('cpu')
+    model = model.to(device)
+    model.eval()
+
+    # Load config and run eval, one dataset at a time
+    print('Evaluating on datasets: {}'.format(args.dataset), flush=True)
+    for dataset in args.dataset.split(','):
+        dataset_cfg = dataset_eval_config()[dataset]
+        args.dataset = dataset
+        run_eval(model, model_cfg, dataset_cfg, device, args)
+
+def run_eval(model, model_cfg, dataset_cfg, device, args):
+    # Create dataset and data loader
+    dataset = create_dataset(model_cfg, dataset_cfg, train=False)
+    dataloader = torch.utils.data.DataLoader(dataset, args.batch_size, shuffle=args.shuffle, num_workers=args.num_workers)
+
+    # List of metrics to log
+    if args.dataset in ['H36M-VAL-P2','3DPW-TEST']:
+        metrics = ['mode_re', 'mode_mpjpe']
+        pck_thresholds = None
+    if args.dataset in ['LSP-EXTENDED', 'POSETRACK-VAL', 'COCO-VAL']:
+        metrics = ['mode_kpl2']
+        pck_thresholds = [0.05, 0.1]
+
+    # Setup evaluator object
+    evaluator = Evaluator(
+        dataset_length=int(1e8), 
+        keypoint_list=dataset_cfg.KEYPOINT_LIST, 
+        pelvis_ind=model_cfg.EXTRA.PELVIS_IND, 
+        metrics=metrics,
+        pck_thresholds=pck_thresholds,
+    )
+
+    # Go over the images in the dataset.
+    try:
+        for i, batch in enumerate(tqdm(dataloader)):
+            batch = recursive_to(batch, device)
+            with torch.no_grad():
+                out = model(batch)
+            evaluator(out, batch)
+            if i % args.log_freq == args.log_freq - 1:
+                evaluator.log()
+        evaluator.log()
+        error = None
+    except (Exception, KeyboardInterrupt) as e:
+        traceback.print_exc()
+        error = repr(e)
+        i = 0
+
+    # Append results to file
+    metrics_dict = evaluator.get_metrics_dict()
+    save_eval_result(args.results_file, metrics_dict, args.checkpoint, args.dataset, error=error, iters_done=i, exp_name=args.exp_name)
+
+
+def save_eval_result(
+    csv_path: str,
+    metric_dict: float,
+    checkpoint_path: str,
+    dataset_name: str,
+    # start_time: pd.Timestamp,
+    error: Optional[str] = None,
+    iters_done=None,
+    exp_name=None,
+) -> None:
+    """Save evaluation results for a single scene file to a common CSV file."""
+
+    timestamp = pd.Timestamp.now()
+    exists: bool = os.path.exists(csv_path)
+    exp_name = exp_name or Path(checkpoint_path).parent.parent.name
+
+    # save each metric as different row to the csv path
+    metric_names = list(metric_dict.keys())
+    metric_values = list(metric_dict.values())
+    N = len(metric_names)
+    df = pd.DataFrame(
+        dict(
+            timestamp=[timestamp] * N,
+            checkpoint_path=[checkpoint_path] * N,
+            exp_name=[exp_name] * N,
+            dataset=[dataset_name] * N,
+            metric_name=metric_names,
+            metric_value=metric_values,
+            error=[error] * N,
+            iters_done=[iters_done] * N,
+        ),
+        index=list(range(N)),
+    )
+
+    # Lock the file to prevent multiple processes from writing to it at the same time.
+    lock = FileLock(f"{csv_path}.lock", timeout=10)
+    with lock:
+        df.to_csv(csv_path, mode="a", header=not exists, index=False)
+
+if __name__ == '__main__':
+    main()

humans4d/fetch_training_data.sh

@@ -0,0 +1,21 @@
+# Google drive links to download the training data
+gdown https://drive.google.com/uc?id=19e5w1x8JIJkYUT9TIPcFhQF1SXny_Qwz # hmr2_training_data_part1.tar
+gdown https://drive.google.com/uc?id=10glv5LKLhqPHHpaWmvkg4wT3wSmL6E2f # hmr2_training_data_part2.tar
+gdown https://drive.google.com/uc?id=19AJMbE0nKD6Lao22aAaPLngiuvrnxF5V # hmr2_training_data_part3.tar
+gdown https://drive.google.com/uc?id=1ntyjT3-JbtDxoHqmCJzJ8JemMKxM78Xj # hmr2_training_data_part4a.tar
+gdown https://drive.google.com/uc?id=1H1xBQqePHcayR65oRn3cnxdgau_WLrLD # hmr2_training_data_part4b.tar
+gdown https://drive.google.com/uc?id=1TnYqqG-wQMfMjd0EHbMEqxDt8DXG3vbW # hmr2_training_data_part4c.tar
+gdown https://drive.google.com/uc?id=1tY6N8RDtgWfHSnTLgUNfbGhPI5X9fMg9 # hmr2_training_data_part4d.tar
+
+# Alternatively, consider using the dropbox links:
+#wget https://www.dropbox.com/sh/4y90ay6a6nki9v7/AACJetGWdtee9TP_ejmBMdKOa/hmr2_training_data_part1.tar
+#wget https://www.dropbox.com/sh/4y90ay6a6nki9v7/AAA7vh5e4jubZ3cYRyOtYYQZa/hmr2_training_data_part2.tar
+#wget https://www.dropbox.com/sh/4y90ay6a6nki9v7/AACtCMwexHgDCMxh5CYP1VEya/hmr2_training_data_part3.tar
+#wget https://www.dropbox.com/sh/4y90ay6a6nki9v7/AAB7C9Y33MEX5G9_UCHQ2KyLa/hmr2_training_data_part4a.tar
+#wget https://www.dropbox.com/sh/4y90ay6a6nki9v7/AABkok2N00PHFGHiKV8-RjpNa/hmr2_training_data_part4b.tar
+#wget https://www.dropbox.com/sh/4y90ay6a6nki9v7/AADYg_LKlGbjpjvrOQWoOgrna/hmr2_training_data_part4c.tar
+#wget https://www.dropbox.com/sh/4y90ay6a6nki9v7/AACARYZJMoxq9nS6cDMAL8Pya/hmr2_training_data_part4d.tar
+
+for f in hmr2_training_data_part*.tar; do
+    tar --warning=no-unknown-keyword --exclude=".*" -xvf $f
+done

humans4d/gradio_app.py

@@ -0,0 +1,187 @@
+from pathlib import Path
+import tempfile
+import torch
+import argparse
+import os
+import numpy as np
+import gradio as gr
+from PIL import Image
+
+from hmr2.configs import CACHE_DIR_4DHUMANS
+from hmr2.models import HMR2, download_models, load_hmr2, DEFAULT_CHECKPOINT
+from hmr2.utils import recursive_to
+from hmr2.datasets.vitdet_dataset import ViTDetDataset, DEFAULT_MEAN, DEFAULT_STD
+from hmr2.utils.renderer import Renderer, cam_crop_to_full
+
+LIGHT_BLUE=(0.65098039,  0.74117647,  0.85882353)
+
+def main():
+    parser = argparse.ArgumentParser(description='HMR2 demo code')
+    parser.add_argument('--checkpoint', type=str, default=DEFAULT_CHECKPOINT, help='Path to pretrained model checkpoint')
+    # parser.add_argument('--img_folder', type=str, default='example_data/images', help='Folder with input images')
+    parser.add_argument('--out_folder', type=str, default='demo_out', help='Output folder to save rendered results')
+    # parser.add_argument('--side_view', dest='side_view', action='store_true', default=False, help='If set, render side view also')
+    # parser.add_argument('--full_frame', dest='full_frame', action='store_true', default=False, help='If set, render all people together also')
+    # parser.add_argument('--save_mesh', dest='save_mesh', action='store_true', default=False, help='If set, save meshes to disk also')
+    parser.add_argument('--batch_size', type=int, default=64, help='Batch size for inference/fitting')
+
+    args = parser.parse_args()
+
+    # Download and load checkpoints
+    download_models(CACHE_DIR_4DHUMANS)
+    model, model_cfg = load_hmr2(args.checkpoint)
+
+    # Setup HMR2.0 model
+    device = torch.device('cuda') if torch.cuda.is_available() else torch.device('cpu')
+    model = model.to(device)
+    model.eval()
+
+    # Load detector
+    from hmr2.utils.utils_detectron2 import DefaultPredictor_Lazy
+    from detectron2.config import LazyConfig
+    import hmr2
+    cfg_path = Path(hmr2.__file__).parent/'configs'/'cascade_mask_rcnn_vitdet_h_75ep.py'
+    detectron2_cfg = LazyConfig.load(str(cfg_path))
+    detectron2_cfg.train.init_checkpoint = "https://dl.fbaipublicfiles.com/detectron2/ViTDet/COCO/cascade_mask_rcnn_vitdet_h/f328730692/model_final_f05665.pkl"
+    for i in range(3):
+        detectron2_cfg.model.roi_heads.box_predictors[i].test_score_thresh = 0.25
+    detector = DefaultPredictor_Lazy(detectron2_cfg)
+
+    # Setup the renderer
+    renderer = Renderer(model_cfg, faces=model.smpl.faces)
+
+    # Make output directory if it does not exist
+    os.makedirs(args.out_folder, exist_ok=True)
+
+    # # Iterate over all images in folder
+    # for img_path in Path(args.img_folder).glob('*.jpg'):
+    #     img_cv2 = cv2.imread(str(img_path))
+    def infer(in_pil_img, in_threshold=0.8, out_pil_img=None):
+        # Convert RGB to BGR
+        img_cv2 = np.array(in_pil_img)[:, :, ::-1].copy()
+
+        # Detect humans in image
+        det_out = detector(img_cv2)
+
+        det_instances = det_out['instances']
+        valid_idx = (det_instances.pred_classes==0) & (det_instances.scores > 0.5)
+        boxes=det_instances.pred_boxes.tensor[valid_idx].cpu().numpy()
+
+        # Run HMR2.0 on all detected humans
+        dataset = ViTDetDataset(model_cfg, img_cv2, boxes)
+        dataloader = torch.utils.data.DataLoader(dataset, batch_size=8, shuffle=False, num_workers=0)
+
+        all_verts = []
+        all_cam_t = []
+        all_mesh_paths = []
+        
+        temp_name = next(tempfile._get_candidate_names())
+
+        for batch in dataloader:
+            batch = recursive_to(batch, device)
+            with torch.no_grad():
+                out = model(batch)
+
+            pred_cam = out['pred_cam']
+            box_center = batch["box_center"].float()
+            box_size = batch["box_size"].float()
+            img_size = batch["img_size"].float()
+            scaled_focal_length = model_cfg.EXTRA.FOCAL_LENGTH / model_cfg.MODEL.IMAGE_SIZE * img_size.max()
+            pred_cam_t_full = cam_crop_to_full(pred_cam, box_center, box_size, img_size, scaled_focal_length).detach().cpu().numpy()
+
+            # Render the result
+            batch_size = batch['img'].shape[0]
+            for n in range(batch_size):
+                # Get filename from path img_path
+                # img_fn, _ = os.path.splitext(os.path.basename(img_path))
+                person_id = int(batch['personid'][n])
+                white_img = (torch.ones_like(batch['img'][n]).cpu() - DEFAULT_MEAN[:,None,None]/255) / (DEFAULT_STD[:,None,None]/255)
+                input_patch = batch['img'][n].cpu() * (DEFAULT_STD[:,None,None]/255) + (DEFAULT_MEAN[:,None,None]/255)
+                input_patch = input_patch.permute(1,2,0).numpy()
+
+                # regression_img = renderer(out['pred_vertices'][n].detach().cpu().numpy(),
+                #                         out['pred_cam_t'][n].detach().cpu().numpy(),
+                #                         batch['img'][n],
+                #                         mesh_base_color=LIGHT_BLUE,
+                #                         scene_bg_color=(1, 1, 1),
+                #                         )
+
+                # if args.side_view:
+                #     side_img = renderer(out['pred_vertices'][n].detach().cpu().numpy(),
+                #                             out['pred_cam_t'][n].detach().cpu().numpy(),
+                #                             white_img,
+                #                             mesh_base_color=LIGHT_BLUE,
+                #                             scene_bg_color=(1, 1, 1),
+                #                             side_view=True)
+                #     final_img = np.concatenate([input_patch, regression_img, side_img], axis=1)
+                # else:
+                #     final_img = np.concatenate([input_patch, regression_img], axis=1)
+
+                # cv2.imwrite(os.path.join(args.out_folder, f'{img_fn}_{person_id}.png'), 255*final_img[:, :, ::-1])
+
+                # Add all verts and cams to list
+                verts = out['pred_vertices'][n].detach().cpu().numpy()
+                cam_t = pred_cam_t_full[n]
+                all_verts.append(verts)
+                all_cam_t.append(cam_t)
+
+                # Save all meshes to disk
+                # if args.save_mesh:
+                if True:
+                    camera_translation = cam_t.copy()
+                    tmesh = renderer.vertices_to_trimesh(verts, camera_translation, LIGHT_BLUE)
+
+                    temp_path = os.path.join(args.out_folder, f'{temp_name}_{person_id}.obj')
+                    tmesh.export(temp_path)
+                    all_mesh_paths.append(temp_path)
+
+        # Render front view
+        if len(all_verts) > 0:
+            misc_args = dict(
+                mesh_base_color=LIGHT_BLUE,
+                scene_bg_color=(1, 1, 1),
+                focal_length=scaled_focal_length,
+            )
+            cam_view = renderer.render_rgba_multiple(all_verts, cam_t=all_cam_t, render_res=img_size[n], **misc_args)
+
+            # Overlay image
+            input_img = img_cv2.astype(np.float32)[:,:,::-1]/255.0
+            input_img = np.concatenate([input_img, np.ones_like(input_img[:,:,:1])], axis=2) # Add alpha channel
+            input_img_overlay = input_img[:,:,:3] * (1-cam_view[:,:,3:]) + cam_view[:,:,:3] * cam_view[:,:,3:]
+
+            # cv2.imwrite(os.path.join(args.out_folder, f'{img_fn}_all.png'), 255*input_img_overlay[:, :, ::-1])
+
+            # convert to PIL image
+            out_pil_img =  Image.fromarray((input_img_overlay*255).astype(np.uint8))
+            return out_pil_img, all_mesh_paths
+        else:
+            return None, []
+
+    with gr.Blocks(title="4DHumans", css=".gradio-container") as demo:
+
+        gr.HTML("""<div style="font-weight:bold; text-align:center; color:royalblue;">HMR 2.0</div>""")
+
+        with gr.Row():
+            with gr.Column():
+                input_image = gr.Image(label="Input image", type="pil")
+            with gr.Column():
+                output_image = gr.Image(label="Reconstructions", type="pil")
+                output_meshes = gr.File(label="3D meshes")
+
+        gr.HTML("""<br/>""")
+
+        with gr.Row():
+            threshold = gr.Slider(0, 1.0, value=0.6, label='Detection Threshold')
+            send_btn = gr.Button("Infer")
+            send_btn.click(fn=infer, inputs=[input_image, threshold], outputs=[output_image, output_meshes])
+
+        gr.Examples([
+                ['example_data/images/pexels-anete-lusina-4793258.jpg'], 
+                ['example_data/images/skates.png'],
+            ], 
+            inputs=[input_image, 0.6])
+
+    demo.launch(debug=True)
+
+if __name__ == '__main__':
+    main()

humans4d/hmr2/configs/__init__.py

@@ -0,0 +1,115 @@
+import os
+from typing import Dict
+from yacs.config import CfgNode as CN
+
+CACHE_DIR = os.path.join(os.environ.get("HOME"), ".cache")
+CACHE_DIR_4DHUMANS = os.path.join(CACHE_DIR, "4DHumans")
+
+def to_lower(x: Dict) -> Dict:
+    """
+    Convert all dictionary keys to lowercase
+    Args:
+      x (dict): Input dictionary
+    Returns:
+      dict: Output dictionary with all keys converted to lowercase
+    """
+    return {k.lower(): v for k, v in x.items()}
+
+_C = CN(new_allowed=True)
+
+_C.GENERAL = CN(new_allowed=True)
+_C.GENERAL.RESUME = True
+_C.GENERAL.TIME_TO_RUN = 3300
+_C.GENERAL.VAL_STEPS = 100
+_C.GENERAL.LOG_STEPS = 100
+_C.GENERAL.CHECKPOINT_STEPS = 20000
+_C.GENERAL.CHECKPOINT_DIR = "checkpoints"
+_C.GENERAL.SUMMARY_DIR = "tensorboard"
+_C.GENERAL.NUM_GPUS = 1
+_C.GENERAL.NUM_WORKERS = 4
+_C.GENERAL.MIXED_PRECISION = True
+_C.GENERAL.ALLOW_CUDA = True
+_C.GENERAL.PIN_MEMORY = False
+_C.GENERAL.DISTRIBUTED = False
+_C.GENERAL.LOCAL_RANK = 0
+_C.GENERAL.USE_SYNCBN = False
+_C.GENERAL.WORLD_SIZE = 1
+
+_C.TRAIN = CN(new_allowed=True)
+_C.TRAIN.NUM_EPOCHS = 100
+_C.TRAIN.BATCH_SIZE = 32
+_C.TRAIN.SHUFFLE = True
+_C.TRAIN.WARMUP = False
+_C.TRAIN.NORMALIZE_PER_IMAGE = False
+_C.TRAIN.CLIP_GRAD = False
+_C.TRAIN.CLIP_GRAD_VALUE = 1.0
+_C.LOSS_WEIGHTS = CN(new_allowed=True)
+
+_C.DATASETS = CN(new_allowed=True)
+
+_C.MODEL = CN(new_allowed=True)
+_C.MODEL.IMAGE_SIZE = 224
+
+_C.EXTRA = CN(new_allowed=True)
+_C.EXTRA.FOCAL_LENGTH = 5000
+
+_C.DATASETS.CONFIG = CN(new_allowed=True)
+_C.DATASETS.CONFIG.SCALE_FACTOR = 0.3
+_C.DATASETS.CONFIG.ROT_FACTOR = 30
+_C.DATASETS.CONFIG.TRANS_FACTOR = 0.02
+_C.DATASETS.CONFIG.COLOR_SCALE = 0.2
+_C.DATASETS.CONFIG.ROT_AUG_RATE = 0.6
+_C.DATASETS.CONFIG.TRANS_AUG_RATE = 0.5
+_C.DATASETS.CONFIG.DO_FLIP = True
+_C.DATASETS.CONFIG.FLIP_AUG_RATE = 0.5
+_C.DATASETS.CONFIG.EXTREME_CROP_AUG_RATE = 0.10
+
+def default_config() -> CN:
+    """
+    Get a yacs CfgNode object with the default config values.
+    """
+    # Return a clone so that the defaults will not be altered
+    # This is for the "local variable" use pattern
+    return _C.clone()
+
+def dataset_config(name='datasets_tar.yaml') -> CN:
+    """
+    Get dataset config file
+    Returns:
+      CfgNode: Dataset config as a yacs CfgNode object.
+    """
+    cfg = CN(new_allowed=True)
+    config_file = os.path.join(os.path.dirname(os.path.realpath(__file__)), name)
+    cfg.merge_from_file(config_file)
+    cfg.freeze()
+    return cfg
+
+def dataset_eval_config() -> CN:
+    return dataset_config('datasets_eval.yaml')
+
+def get_config(config_file: str, merge: bool = True, update_cachedir: bool = False) -> CN:
+    """
+    Read a config file and optionally merge it with the default config file.
+    Args:
+      config_file (str): Path to config file.
+      merge (bool): Whether to merge with the default config or not.
+    Returns:
+      CfgNode: Config as a yacs CfgNode object.
+    """
+    if merge:
+      cfg = default_config()
+    else:
+      cfg = CN(new_allowed=True)
+    cfg.merge_from_file(config_file)
+
+    if update_cachedir:
+      def update_path(path: str) -> str:
+        if os.path.isabs(path):
+          return path
+        return os.path.join(CACHE_DIR_4DHUMANS, path)
+
+    #   cfg.SMPL.MODEL_PATH = update_path(cfg.SMPL.MODEL_PATH)
+    #   cfg.SMPL.JOINT_REGRESSOR_EXTRA = update_path(cfg.SMPL.JOINT_REGRESSOR_EXTRA)
+    #   cfg.SMPL.MEAN_PARAMS = update_path(cfg.SMPL.MEAN_PARAMS)
+
+    return cfg

humans4d/hmr2/configs/cascade_mask_rcnn_vitdet_h_75ep.py

@@ -0,0 +1,129 @@
+## coco_loader_lsj.py
+
+import detectron2.data.transforms as T
+from detectron2 import model_zoo
+from detectron2.config import LazyCall as L
+
+# Data using LSJ
+image_size = 1024
+dataloader = model_zoo.get_config("common/data/coco.py").dataloader
+dataloader.train.mapper.augmentations = [
+    L(T.RandomFlip)(horizontal=True),  # flip first
+    L(T.ResizeScale)(
+        min_scale=0.1, max_scale=2.0, target_height=image_size, target_width=image_size
+    ),
+    L(T.FixedSizeCrop)(crop_size=(image_size, image_size), pad=False),
+]
+dataloader.train.mapper.image_format = "RGB"
+dataloader.train.total_batch_size = 64
+# recompute boxes due to cropping
+dataloader.train.mapper.recompute_boxes = True
+
+dataloader.test.mapper.augmentations = [
+    L(T.ResizeShortestEdge)(short_edge_length=image_size, max_size=image_size),
+]
+
+from functools import partial
+from fvcore.common.param_scheduler import MultiStepParamScheduler
+
+from detectron2 import model_zoo
+from detectron2.config import LazyCall as L
+from detectron2.solver import WarmupParamScheduler
+from detectron2.modeling.backbone.vit import get_vit_lr_decay_rate
+
+# mask_rcnn_vitdet_b_100ep.py
+
+model = model_zoo.get_config("common/models/mask_rcnn_vitdet.py").model
+
+# Initialization and trainer settings
+train = model_zoo.get_config("common/train.py").train
+train.amp.enabled = True
+train.ddp.fp16_compression = True
+train.init_checkpoint = "detectron2://ImageNetPretrained/MAE/mae_pretrain_vit_base.pth"
+
+
+# Schedule
+# 100 ep = 184375 iters * 64 images/iter / 118000 images/ep
+train.max_iter = 184375
+
+lr_multiplier = L(WarmupParamScheduler)(
+    scheduler=L(MultiStepParamScheduler)(
+        values=[1.0, 0.1, 0.01],
+        milestones=[163889, 177546],
+        num_updates=train.max_iter,
+    ),
+    warmup_length=250 / train.max_iter,
+    warmup_factor=0.001,
+)
+
+# Optimizer
+optimizer = model_zoo.get_config("common/optim.py").AdamW
+optimizer.params.lr_factor_func = partial(get_vit_lr_decay_rate, num_layers=12, lr_decay_rate=0.7)
+optimizer.params.overrides = {"pos_embed": {"weight_decay": 0.0}}
+
+# cascade_mask_rcnn_vitdet_b_100ep.py
+
+from detectron2.config import LazyCall as L
+from detectron2.layers import ShapeSpec
+from detectron2.modeling.box_regression import Box2BoxTransform
+from detectron2.modeling.matcher import Matcher
+from detectron2.modeling.roi_heads import (
+    FastRCNNOutputLayers,
+    FastRCNNConvFCHead,
+    CascadeROIHeads,
+)
+
+# arguments that don't exist for Cascade R-CNN
+[model.roi_heads.pop(k) for k in ["box_head", "box_predictor", "proposal_matcher"]]
+
+model.roi_heads.update(
+    _target_=CascadeROIHeads,
+    box_heads=[
+        L(FastRCNNConvFCHead)(
+            input_shape=ShapeSpec(channels=256, height=7, width=7),
+            conv_dims=[256, 256, 256, 256],
+            fc_dims=[1024],
+            conv_norm="LN",
+        )
+        for _ in range(3)
+    ],
+    box_predictors=[
+        L(FastRCNNOutputLayers)(
+            input_shape=ShapeSpec(channels=1024),
+            test_score_thresh=0.05,
+            box2box_transform=L(Box2BoxTransform)(weights=(w1, w1, w2, w2)),
+            cls_agnostic_bbox_reg=True,
+            num_classes="${...num_classes}",
+        )
+        for (w1, w2) in [(10, 5), (20, 10), (30, 15)]
+    ],
+    proposal_matchers=[
+        L(Matcher)(thresholds=[th], labels=[0, 1], allow_low_quality_matches=False)
+        for th in [0.5, 0.6, 0.7]
+    ],
+)
+
+# cascade_mask_rcnn_vitdet_h_75ep.py
+
+from functools import partial
+
+train.init_checkpoint = "detectron2://ImageNetPretrained/MAE/mae_pretrain_vit_huge_p14to16.pth"
+
+model.backbone.net.embed_dim = 1280
+model.backbone.net.depth = 32
+model.backbone.net.num_heads = 16
+model.backbone.net.drop_path_rate = 0.5
+# 7, 15, 23, 31 for global attention
+model.backbone.net.window_block_indexes = (
+    list(range(0, 7)) + list(range(8, 15)) + list(range(16, 23)) + list(range(24, 31))
+)
+
+optimizer.params.lr_factor_func = partial(get_vit_lr_decay_rate, lr_decay_rate=0.9, num_layers=32)
+optimizer.params.overrides = {}
+optimizer.params.weight_decay_norm = None
+
+train.max_iter = train.max_iter * 3 // 4  # 100ep -> 75ep
+lr_multiplier.scheduler.milestones = [
+    milestone * 3 // 4 for milestone in lr_multiplier.scheduler.milestones
+]
+lr_multiplier.scheduler.num_updates = train.max_iter

humans4d/hmr2/configs/datasets_eval.yaml

@@ -0,0 +1,31 @@
+H36M-VAL-P2:
+    TYPE: ImageDataset
+    DATASET_FILE: hmr2_evaluation_data/h36m_val_p2.npz
+    IMG_DIR: /shared/pavlakos/datasets/h36m/images/
+    KEYPOINT_LIST: [25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 43]
+    USE_HIPS: True
+
+3DPW-TEST:
+    TYPE: ImageDataset
+    DATASET_FILE: hmr2_evaluation_data/3dpw_test.npz
+    IMG_DIR: /shared/pavlakos/datasets/3DPW/
+    KEYPOINT_LIST: [25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 43]
+    USE_HIPS: False
+
+POSETRACK-VAL:
+    TYPE: ImageDataset
+    DATASET_FILE: hmr2_evaluation_data/posetrack_2018_val.npz
+    IMG_DIR:      /shared/pavlakos/datasets/posetrack/posetrack2018/posetrack_data/
+    KEYPOINT_LIST: [0] # Dummy
+
+LSP-EXTENDED:
+    TYPE: ImageDataset
+    DATASET_FILE:   hmr2_evaluation_data/hr-lspet_train.npz
+    IMG_DIR:        /shared/pavlakos/datasets/hr-lspet/
+    KEYPOINT_LIST: [0] # Dummy
+
+COCO-VAL:
+    TYPE: ImageDataset
+    DATASET_FILE: hmr2_evaluation_data/coco_val.npz
+    IMG_DIR: /shared/pavlakos/datasets/coco/
+    KEYPOINT_LIST: [0] # Dummy

humans4d/hmr2/configs/datasets_tar.yaml

@@ -0,0 +1,38 @@
+MPI-INF-TRAIN-PRUNED:
+    TYPE: ImageDataset
+    URLS: hmr2_training_data/dataset_tars/mpi-inf-train-pruned/{000000..00006}.tar
+    epoch_size: 12_000
+H36M-TRAIN-WMASK:
+    TYPE: ImageDataset
+    URLS: hmr2_training_data/dataset_tars/h36m-train/{000000..000312}.tar
+    epoch_size: 314_000 # Can be changed arbitrarily. Enables resampling. 
+MPII-TRAIN-WMASK:
+    TYPE: ImageDataset
+    URLS: hmr2_training_data/dataset_tars/mpii-train/{000000..000009}.tar
+    epoch_size: 100_000 # Increase to ensure dataset doesn't end before epoch
+COCO-TRAIN-2014-WMASK-PRUNED:
+    TYPE: ImageDataset
+    URLS: hmr2_training_data/dataset_tars/coco-train-2014-pruned/{000000..000017}.tar
+    epoch_size: 18_000
+AVA-TRAIN-MIDFRAMES-1FPS-WMASK:
+    TYPE: ImageDataset
+    URLS: hmr2_training_data/dataset_tars/ava-train-midframes-1fps-vitpose/{000000..000092}.tar
+    epoch_size: 200_000
+AIC-TRAIN-WMASK:
+    TYPE: ImageDataset
+    URLS: hmr2_training_data/dataset_tars/aic-train-vitpose/{000000..000104}.tar
+    epoch_size: 200_000
+INSTA-TRAIN-WMASK:
+    TYPE: ImageDataset
+    URLS: hmr2_training_data/dataset_tars/insta-train-vitpose-replicate/{000000..003657}.tar
+    epoch_size: 4_000_000
+COCO-TRAIN-2014-VITPOSE-REPLICATE-PRUNED12:
+    TYPE: ImageDataset
+    URLS: hmr2_training_data/dataset_tars/coco-train-2014-vitpose-pruned/{000000..000044}.tar
+    epoch_size: 45_000
+COCO-VAL:
+    TYPE: ImageDataset
+    URLS: hmr2_training_data/dataset_tars/coco-val/{000000..000000}.tar
+    KEYPOINT_LIST: [0] # Dummy
+CMU-MOCAP:
+    DATASET_FILE: hmr2_training_data/cmu_mocap.npz

humans4d/hmr2/configs_hydra/data/mix_all.yaml

@@ -0,0 +1,27 @@
+# @package _global_
+defaults:
+  - /data_filtering: low1
+
+DATASETS:
+  TRAIN: 
+    H36M-TRAIN-WMASK:
+      WEIGHT: 0.1
+    MPII-TRAIN-WMASK:
+      WEIGHT: 0.1
+    COCO-TRAIN-2014-WMASK-PRUNED:
+      WEIGHT: 0.1
+    COCO-TRAIN-2014-VITPOSE-REPLICATE-PRUNED12:
+      WEIGHT: 0.1
+    MPI-INF-TRAIN-PRUNED:
+      WEIGHT: 0.02
+    AVA-TRAIN-MIDFRAMES-1FPS-WMASK:
+      WEIGHT: 0.19
+    AIC-TRAIN-WMASK:
+      WEIGHT: 0.19
+    INSTA-TRAIN-WMASK:
+      WEIGHT: 0.2
+  VAL:
+    COCO-VAL:
+      WEIGHT: 1.0
+
+  MOCAP: CMU-MOCAP

humans4d/hmr2/configs_hydra/data_filtering/low1.yaml

@@ -0,0 +1,13 @@
+# @package _global_
+
+DATASETS:
+  # Data filtering during training
+  SUPPRESS_KP_CONF_THRESH: 0.3
+  FILTER_NUM_KP: 4
+  FILTER_NUM_KP_THRESH: 0.0
+  FILTER_REPROJ_THRESH: 31000
+
+  SUPPRESS_BETAS_THRESH: 3.0
+  SUPPRESS_BAD_POSES: True
+  POSES_BETAS_SIMULTANEOUS: True
+  FILTER_NO_POSES: False # If True, filters images that don't have poses

humans4d/hmr2/configs_hydra/experiment/default.yaml

@@ -0,0 +1,29 @@
+# @package _global_
+
+SMPL:
+  DATA_DIR: ${oc.env:HOME}/.cache/4DHumans/data/
+  MODEL_PATH: ${SMPL.DATA_DIR}/smpl
+  GENDER: neutral
+  NUM_BODY_JOINTS: 23
+  JOINT_REGRESSOR_EXTRA: ${SMPL.DATA_DIR}/SMPL_to_J19.pkl
+  MEAN_PARAMS: ${SMPL.DATA_DIR}/smpl_mean_params.npz
+
+EXTRA:
+  FOCAL_LENGTH: 5000
+  NUM_LOG_IMAGES: 4
+  NUM_LOG_SAMPLES_PER_IMAGE: 8
+  PELVIS_IND: 39
+
+DATASETS:
+  BETAS_REG: True
+  CONFIG:
+    SCALE_FACTOR: 0.3
+    ROT_FACTOR: 30
+    TRANS_FACTOR: 0.02
+    COLOR_SCALE: 0.2
+    ROT_AUG_RATE: 0.6
+    TRANS_AUG_RATE: 0.5
+    DO_FLIP: True
+    FLIP_AUG_RATE: 0.5
+    EXTREME_CROP_AUG_RATE: 0.10
+    EXTREME_CROP_AUG_LEVEL: 1

humans4d/hmr2/configs_hydra/experiment/hmr_vit_transformer.yaml

@@ -0,0 +1,51 @@
+# @package _global_
+
+defaults:
+  - default.yaml
+
+GENERAL:
+  TOTAL_STEPS: 1_000_000
+  LOG_STEPS: 1000
+  VAL_STEPS: 1000
+  CHECKPOINT_STEPS: 10000
+  CHECKPOINT_SAVE_TOP_K: 1
+  NUM_WORKERS: 6
+  PREFETCH_FACTOR: 2
+
+TRAIN:
+  LR: 1e-5
+  WEIGHT_DECAY: 1e-4
+  BATCH_SIZE: 48
+  LOSS_REDUCTION: mean
+  NUM_TRAIN_SAMPLES: 2
+  NUM_TEST_SAMPLES: 64
+  POSE_2D_NOISE_RATIO: 0.01
+  SMPL_PARAM_NOISE_RATIO: 0.005
+
+MODEL:
+  IMAGE_SIZE: 256
+  IMAGE_MEAN: [0.485, 0.456, 0.406]
+  IMAGE_STD: [0.229, 0.224, 0.225]
+  BACKBONE:
+    TYPE: vit
+    PRETRAINED_WEIGHTS: hmr2_training_data/vitpose_backbone.pth
+  SMPL_HEAD:
+    TYPE: transformer_decoder
+    IN_CHANNELS: 2048
+    TRANSFORMER_DECODER:
+      depth: 6
+      heads: 8
+      mlp_dim: 1024
+      dim_head: 64
+      dropout: 0.0
+      emb_dropout: 0.0
+      norm: layer
+      context_dim: 1280 # from vitpose-H
+
+LOSS_WEIGHTS:
+  KEYPOINTS_3D: 0.05
+  KEYPOINTS_2D: 0.01
+  GLOBAL_ORIENT: 0.001
+  BODY_POSE: 0.001
+  BETAS: 0.0005
+  ADVERSARIAL: 0.0005

humans4d/hmr2/configs_hydra/extras/default.yaml

@@ -0,0 +1,8 @@
+# disable python warnings if they annoy you
+ignore_warnings: False
+
+# ask user for tags if none are provided in the config
+enforce_tags: True
+
+# pretty print config tree at the start of the run using Rich library
+print_config: True

humans4d/hmr2/configs_hydra/hydra/default.yaml

@@ -0,0 +1,26 @@
+# @package _global_
+# https://hydra.cc/docs/configure_hydra/intro/
+
+# enable color logging
+defaults:
+  - override /hydra/hydra_logging: colorlog
+  - override /hydra/job_logging: colorlog
+
+# exp_name: ovrd_${hydra:job.override_dirname}
+exp_name: ${now:%Y-%m-%d}_${now:%H-%M-%S}
+
+hydra:
+  run:
+    dir: ${paths.log_dir}/${task_name}/runs/${exp_name}
+  sweep:
+    dir: ${paths.log_dir}/${task_name}/multiruns/${exp_name}
+    subdir: ${hydra.job.num}
+  job:
+    config:
+      override_dirname:
+        exclude_keys:
+          - trainer
+          - trainer.devices
+          - trainer.num_nodes
+          - callbacks
+          - debug

humans4d/hmr2/configs_hydra/launcher/local.yaml

@@ -0,0 +1,13 @@
+# @package _global_
+
+defaults:
+  - override /hydra/launcher: submitit_local
+
+hydra:
+  launcher:
+    timeout_min: 10_080   # 7 days
+    nodes: 1
+    tasks_per_node: ${trainer.devices}
+    cpus_per_task: 6
+    gpus_per_node: ${trainer.devices}
+    name: hmr2

humans4d/hmr2/configs_hydra/launcher/slurm.yaml

@@ -0,0 +1,22 @@
+# @package _global_
+
+defaults:
+  - override /hydra/launcher: submitit_slurm
+
+hydra:
+  launcher:
+    timeout_min: 10_080   # 7 days
+    max_num_timeout: 3
+    partition: g40
+    qos: idle
+    nodes: 1
+    tasks_per_node: ${trainer.devices}
+    gpus_per_task: null
+    cpus_per_task: 12
+    gpus_per_node: ${trainer.devices}
+    cpus_per_gpu: null
+    comment: laion
+    name: hmr2
+    setup:
+      - module load cuda openmpi libfabric-aws
+      - export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7

humans4d/hmr2/configs_hydra/paths/default.yaml

@@ -0,0 +1,18 @@
+# path to root directory
+# this requires PROJECT_ROOT environment variable to exist
+# PROJECT_ROOT is inferred and set by pyrootutils package in `train.py` and `eval.py`
+root_dir: ${oc.env:PROJECT_ROOT}
+
+# path to data directory
+data_dir: ${paths.root_dir}/data/
+
+# path to logging directory
+log_dir: logs/
+
+# path to output directory, created dynamically by hydra
+# path generation pattern is specified in `configs/hydra/default.yaml`
+# use it to store all files generated during the run, like ckpts and metrics
+output_dir: ${hydra:runtime.output_dir}
+
+# path to working directory
+work_dir: ${hydra:runtime.cwd}

humans4d/hmr2/configs_hydra/train.yaml

@@ -0,0 +1,47 @@
+# @package _global_
+
+# specify here default configuration
+# order of defaults determines the order in which configs override each other
+defaults:
+  - _self_
+  - data: mix_all.yaml
+  - trainer: ddp.yaml
+  - paths: default.yaml
+  - extras: default.yaml
+  - hydra: default.yaml
+
+  # experiment configs allow for version control of specific hyperparameters
+  # e.g. best hyperparameters for given model and datamodule
+  - experiment: null
+  - texture_exp: null
+
+  # optional local config for machine/user specific settings
+  # it's optional since it doesn't need to exist and is excluded from version control
+  - optional launcher: local.yaml
+  # - optional launcher: slurm.yaml
+
+  # debugging config (enable through command line, e.g. `python train.py debug=default)
+  - debug: null
+
+# task name, determines output directory path
+task_name: "train"
+
+# tags to help you identify your experiments
+# you can overwrite this in experiment configs
+# overwrite from command line with `python train.py tags="[first_tag, second_tag]"`
+# appending lists from command line is currently not supported :(
+# https://github.com/facebookresearch/hydra/issues/1547
+tags: ["dev"]
+
+# set False to skip model training
+train: True
+
+# evaluate on test set, using best model weights achieved during training
+# lightning chooses best weights based on the metric specified in checkpoint callback
+test: False
+
+# simply provide checkpoint path to resume training
+ckpt_path: null
+
+# seed for random number generators in pytorch, numpy and python.random
+seed: null

humans4d/hmr2/configs_hydra/trainer/cpu.yaml

@@ -0,0 +1,6 @@
+defaults:
+  - default.yaml
+  - default_hmr.yaml
+
+accelerator: cpu
+devices: 1

humans4d/hmr2/configs_hydra/trainer/ddp.yaml

@@ -0,0 +1,14 @@
+defaults:
+  - default.yaml
+  - default_hmr.yaml
+
+# use "ddp_spawn" instead of "ddp",
+# it's slower but normal "ddp" currently doesn't work ideally with hydra
+# https://github.com/facebookresearch/hydra/issues/2070
+# https://pytorch-lightning.readthedocs.io/en/latest/accelerators/gpu_intermediate.html#distributed-data-parallel-spawn
+strategy: ddp
+
+accelerator: gpu
+devices: 8
+num_nodes: 1
+sync_batchnorm: True

humans4d/hmr2/configs_hydra/trainer/default.yaml

@@ -0,0 +1,10 @@
+_target_: pytorch_lightning.Trainer
+
+default_root_dir: ${paths.output_dir}
+
+accelerator: cpu
+devices: 1
+
+# set True to to ensure deterministic results
+# makes training slower but gives more reproducibility than just setting seeds
+deterministic: False

humans4d/hmr2/configs_hydra/trainer/default_hmr.yaml

@@ -0,0 +1,8 @@
+num_sanity_val_steps: 0
+log_every_n_steps: ${GENERAL.LOG_STEPS}
+val_check_interval: ${GENERAL.VAL_STEPS}
+precision: 16
+max_steps: ${GENERAL.TOTAL_STEPS}
+# move_metrics_to_cpu: True
+limit_val_batches: 1
+# track_grad_norm: -1

humans4d/hmr2/configs_hydra/trainer/gpu.yaml

@@ -0,0 +1,6 @@
+defaults:
+  - default.yaml
+  - default_hmr.yaml
+
+accelerator: gpu
+devices: 1

humans4d/hmr2/configs_hydra/trainer/mps.yaml

@@ -0,0 +1,6 @@
+defaults:
+  - default.yaml
+  - default_hmr.yaml
+
+accelerator: mps
+devices: 1

humans4d/hmr2/datasets/__init__.py

@@ -0,0 +1,88 @@
+from typing import Dict, Optional
+
+import torch
+import numpy as np
+import pytorch_lightning as pl
+from yacs.config import CfgNode
+
+import webdataset as wds
+from ..configs import to_lower
+from .dataset import Dataset
+from .image_dataset import ImageDataset
+from .mocap_dataset import MoCapDataset
+
+def create_dataset(cfg: CfgNode, dataset_cfg: CfgNode, train: bool = True, **kwargs) -> Dataset:
+    """
+    Instantiate a dataset from a config file.
+    Args:
+        cfg (CfgNode): Model configuration file.
+        dataset_cfg (CfgNode): Dataset configuration info.
+        train (bool): Variable to select between train and val datasets.
+    """
+
+    dataset_type = Dataset.registry[dataset_cfg.TYPE]
+    return dataset_type(cfg, **to_lower(dataset_cfg), train=train, **kwargs)
+
+def create_webdataset(cfg: CfgNode, dataset_cfg: CfgNode, train: bool = True) -> Dataset:
+    """
+    Like `create_dataset` but load data from tars.
+    """
+    dataset_type = Dataset.registry[dataset_cfg.TYPE]
+    return dataset_type.load_tars_as_webdataset(cfg, **to_lower(dataset_cfg), train=train)
+
+
+class MixedWebDataset(wds.WebDataset):
+    def __init__(self, cfg: CfgNode, dataset_cfg: CfgNode, train: bool = True) -> None:
+        super(wds.WebDataset, self).__init__()
+        dataset_list = cfg.DATASETS.TRAIN if train else cfg.DATASETS.VAL
+        datasets = [create_webdataset(cfg, dataset_cfg[dataset], train=train) for dataset, v in dataset_list.items()]
+        weights = np.array([v.WEIGHT for dataset, v in dataset_list.items()])
+        weights = weights / weights.sum()  # normalize
+        self.append(wds.RandomMix(datasets, weights))
+
+class HMR2DataModule(pl.LightningDataModule):
+
+    def __init__(self, cfg: CfgNode, dataset_cfg: CfgNode) -> None:
+        """
+        Initialize LightningDataModule for HMR2 training
+        Args:
+            cfg (CfgNode): Config file as a yacs CfgNode containing necessary dataset info.
+            dataset_cfg (CfgNode): Dataset configuration file
+        """
+        super().__init__()
+        self.cfg = cfg
+        self.dataset_cfg = dataset_cfg
+        self.train_dataset = None
+        self.val_dataset = None
+        self.test_dataset = None
+        self.mocap_dataset = None
+
+    def setup(self, stage: Optional[str] = None) -> None:
+        """
+        Load datasets necessary for training
+        Args:
+            cfg (CfgNode): Config file as a yacs CfgNode containing necessary dataset info.
+        """
+        if self.train_dataset == None:
+            self.train_dataset = MixedWebDataset(self.cfg, self.dataset_cfg, train=True).with_epoch(100_000).shuffle(4000)
+            self.val_dataset = MixedWebDataset(self.cfg, self.dataset_cfg, train=False).shuffle(4000)
+            self.mocap_dataset = MoCapDataset(**to_lower(self.dataset_cfg[self.cfg.DATASETS.MOCAP]))
+
+    def train_dataloader(self) -> Dict:
+        """
+        Setup training data loader.
+        Returns:
+            Dict: Dictionary containing image and mocap data dataloaders
+        """
+        train_dataloader = torch.utils.data.DataLoader(self.train_dataset, self.cfg.TRAIN.BATCH_SIZE, drop_last=True, num_workers=self.cfg.GENERAL.NUM_WORKERS, prefetch_factor=self.cfg.GENERAL.PREFETCH_FACTOR)
+        mocap_dataloader = torch.utils.data.DataLoader(self.mocap_dataset, self.cfg.TRAIN.NUM_TRAIN_SAMPLES * self.cfg.TRAIN.BATCH_SIZE, shuffle=True, drop_last=True, num_workers=1)
+        return {'img': train_dataloader, 'mocap': mocap_dataloader}
+
+    def val_dataloader(self) -> torch.utils.data.DataLoader:
+        """
+        Setup val data loader.
+        Returns:
+            torch.utils.data.DataLoader: Validation dataloader  
+        """
+        val_dataloader = torch.utils.data.DataLoader(self.val_dataset, self.cfg.TRAIN.BATCH_SIZE, drop_last=True, num_workers=self.cfg.GENERAL.NUM_WORKERS)
+        return val_dataloader

humans4d/hmr2/datasets/dataset.py

@@ -0,0 +1,27 @@
+"""
+This file contains the defition of the base Dataset class.
+"""
+
+class DatasetRegistration(type):
+    """
+    Metaclass for registering different datasets
+    """
+    def __init__(cls, name, bases, nmspc):
+        super().__init__(name, bases, nmspc)
+        if not hasattr(cls, 'registry'):
+            cls.registry = dict()
+        cls.registry[name] = cls
+
+    # Metamethods, called on class objects:
+    def __iter__(cls):
+        return iter(cls.registry)
+
+    def __str__(cls):
+        return str(cls.registry)
+
+class Dataset(metaclass=DatasetRegistration):
+    """
+    Base Dataset class
+    """
+    def __init__(self, *args, **kwargs):
+        pass

humans4d/hmr2/datasets/image_dataset.py

@@ -0,0 +1,454 @@
+import copy
+import os
+import numpy as np
+import torch
+from typing import Any, Dict, List
+from yacs.config import CfgNode
+import braceexpand
+import cv2
+
+from .dataset import Dataset
+from .utils import get_example, expand_to_aspect_ratio
+from .smplh_prob_filter import poses_check_probable, load_amass_hist_smooth
+
+def expand(s):
+    return os.path.expanduser(os.path.expandvars(s))
+def expand_urls(urls: str|List[str]):
+    if isinstance(urls, str):
+        urls = [urls]
+    urls = [u for url in urls for u in braceexpand.braceexpand(expand(url))]
+    return urls
+
+AIC_TRAIN_CORRUPT_KEYS = {
+    '0a047f0124ae48f8eee15a9506ce1449ee1ba669',
+    '1a703aa174450c02fbc9cfbf578a5435ef403689',
+    '0394e6dc4df78042929b891dbc24f0fd7ffb6b6d',
+    '5c032b9626e410441544c7669123ecc4ae077058',
+    'ca018a7b4c5f53494006ebeeff9b4c0917a55f07',
+    '4a77adb695bef75a5d34c04d589baf646fe2ba35',
+    'a0689017b1065c664daef4ae2d14ea03d543217e',
+    '39596a45cbd21bed4a5f9c2342505532f8ec5cbb',
+    '3d33283b40610d87db660b62982f797d50a7366b',
+}
+CORRUPT_KEYS = {
+    *{f'aic-train/{k}' for k in AIC_TRAIN_CORRUPT_KEYS},
+    *{f'aic-train-vitpose/{k}' for k in AIC_TRAIN_CORRUPT_KEYS},
+}
+
+body_permutation = [0, 1, 5, 6, 7, 2, 3, 4, 8, 12, 13, 14, 9, 10, 11, 16, 15, 18, 17, 22, 23, 24, 19, 20, 21]
+extra_permutation = [5, 4, 3, 2, 1, 0, 11, 10, 9, 8, 7, 6, 12, 13, 14, 15, 16, 17, 18]
+FLIP_KEYPOINT_PERMUTATION = body_permutation + [25 + i for i in extra_permutation]
+
+DEFAULT_MEAN = 255. * np.array([0.485, 0.456, 0.406])
+DEFAULT_STD = 255. * np.array([0.229, 0.224, 0.225])
+DEFAULT_IMG_SIZE = 256
+
+class ImageDataset(Dataset):
+
+    def __init__(self,
+                 cfg: CfgNode,
+                 dataset_file: str,
+                 img_dir: str,
+                 train: bool = True,
+                 prune: Dict[str, Any] = {},
+                 **kwargs):
+        """
+        Dataset class used for loading images and corresponding annotations.
+        Args:
+            cfg (CfgNode): Model config file.
+            dataset_file (str): Path to npz file containing dataset info.
+            img_dir (str): Path to image folder.
+            train (bool): Whether it is for training or not (enables data augmentation).
+        """
+        super(ImageDataset, self).__init__()
+        self.train = train
+        self.cfg = cfg
+
+        self.img_size = cfg.MODEL.IMAGE_SIZE
+        self.mean = 255. * np.array(self.cfg.MODEL.IMAGE_MEAN)
+        self.std = 255. * np.array(self.cfg.MODEL.IMAGE_STD)
+
+        self.img_dir = img_dir
+        self.data = np.load(dataset_file, allow_pickle=True)
+
+        self.imgname = self.data['imgname']
+        self.personid = np.zeros(len(self.imgname), dtype=np.int32)
+        self.extra_info = self.data.get('extra_info', [{} for _ in range(len(self.imgname))])
+
+        self.flip_keypoint_permutation = copy.copy(FLIP_KEYPOINT_PERMUTATION)
+
+        num_pose = 3 * (self.cfg.SMPL.NUM_BODY_JOINTS + 1)
+
+        # Bounding boxes are assumed to be in the center and scale format
+        self.center = self.data['center']
+        self.scale = self.data['scale'].reshape(len(self.center), -1) / 200.0
+        if self.scale.shape[1] == 1:
+            self.scale = np.tile(self.scale, (1, 2))
+        assert self.scale.shape == (len(self.center), 2)
+
+        # Get gt SMPLX parameters, if available
+        try:
+            self.body_pose = self.data['body_pose'].astype(np.float32)
+            self.has_body_pose = self.data['has_body_pose'].astype(np.float32)
+        except KeyError:
+            self.body_pose = np.zeros((len(self.imgname), num_pose), dtype=np.float32)
+            self.has_body_pose = np.zeros(len(self.imgname), dtype=np.float32)
+        try:
+            self.betas = self.data['betas'].astype(np.float32)
+            self.has_betas = self.data['has_betas'].astype(np.float32)
+        except KeyError:
+            self.betas = np.zeros((len(self.imgname), 10), dtype=np.float32)
+            self.has_betas = np.zeros(len(self.imgname), dtype=np.float32)
+
+        # Try to get 2d keypoints, if available
+        try:
+            body_keypoints_2d = self.data['body_keypoints_2d']
+        except KeyError:
+            body_keypoints_2d = np.zeros((len(self.center), 25, 3))
+        # Try to get extra 2d keypoints, if available
+        try:
+            extra_keypoints_2d = self.data['extra_keypoints_2d']
+        except KeyError:
+            extra_keypoints_2d = np.zeros((len(self.center), 19, 3))
+
+        self.keypoints_2d = np.concatenate((body_keypoints_2d, extra_keypoints_2d), axis=1).astype(np.float32)
+
+        # Try to get 3d keypoints, if available
+        try:
+            body_keypoints_3d = self.data['body_keypoints_3d'].astype(np.float32)
+        except KeyError:
+            body_keypoints_3d = np.zeros((len(self.center), 25, 4), dtype=np.float32)
+        # Try to get extra 3d keypoints, if available
+        try:
+            extra_keypoints_3d = self.data['extra_keypoints_3d'].astype(np.float32)
+        except KeyError:
+            extra_keypoints_3d = np.zeros((len(self.center), 19, 4), dtype=np.float32)
+
+        body_keypoints_3d[:, [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14], -1] = 0
+
+        self.keypoints_3d = np.concatenate((body_keypoints_3d, extra_keypoints_3d), axis=1).astype(np.float32)
+
+    def __len__(self) -> int:
+        return len(self.scale)
+
+    def __getitem__(self, idx: int) -> Dict:
+        """
+        Returns an example from the dataset.
+        """
+        try:
+            image_file_rel = self.imgname[idx].decode('utf-8')
+        except AttributeError:
+            image_file_rel = self.imgname[idx]
+        image_file = os.path.join(self.img_dir, image_file_rel)
+        keypoints_2d = self.keypoints_2d[idx].copy()
+        keypoints_3d = self.keypoints_3d[idx].copy()
+
+        center = self.center[idx].copy()
+        center_x = center[0]
+        center_y = center[1]
+        scale = self.scale[idx]
+        BBOX_SHAPE = self.cfg.MODEL.get('BBOX_SHAPE', None)
+        bbox_size = expand_to_aspect_ratio(scale*200, target_aspect_ratio=BBOX_SHAPE).max()
+        bbox_expand_factor = bbox_size / ((scale*200).max())
+        body_pose = self.body_pose[idx].copy().astype(np.float32)
+        betas = self.betas[idx].copy().astype(np.float32)
+
+        has_body_pose = self.has_body_pose[idx].copy()
+        has_betas = self.has_betas[idx].copy()
+
+        smpl_params = {'global_orient': body_pose[:3],
+                       'body_pose': body_pose[3:],
+                       'betas': betas
+                      }
+
+        has_smpl_params = {'global_orient': has_body_pose,
+                           'body_pose': has_body_pose,
+                           'betas': has_betas
+                           }
+
+        smpl_params_is_axis_angle = {'global_orient': True,
+                                     'body_pose': True,
+                                     'betas': False
+                                    }
+
+        augm_config = self.cfg.DATASETS.CONFIG
+        # Crop image and (possibly) perform data augmentation
+        img_patch, keypoints_2d, keypoints_3d, smpl_params, has_smpl_params, img_size = get_example(image_file,
+                                                                                                    center_x, center_y,
+                                                                                                    bbox_size, bbox_size,
+                                                                                                    keypoints_2d, keypoints_3d,
+                                                                                                    smpl_params, has_smpl_params,
+                                                                                                    self.flip_keypoint_permutation,
+                                                                                                    self.img_size, self.img_size,
+                                                                                                    self.mean, self.std, self.train, augm_config)
+
+        item = {}
+        # These are the keypoints in the original image coordinates (before cropping)
+        orig_keypoints_2d = self.keypoints_2d[idx].copy()
+
+        item['img'] = img_patch
+        item['keypoints_2d'] = keypoints_2d.astype(np.float32)
+        item['keypoints_3d'] = keypoints_3d.astype(np.float32)
+        item['orig_keypoints_2d'] = orig_keypoints_2d
+        item['box_center'] = self.center[idx].copy()
+        item['box_size'] = bbox_size
+        item['bbox_expand_factor'] = bbox_expand_factor
+        item['img_size'] = 1.0 * img_size[::-1].copy()
+        item['smpl_params'] = smpl_params
+        item['has_smpl_params'] = has_smpl_params
+        item['smpl_params_is_axis_angle'] = smpl_params_is_axis_angle
+        item['imgname'] = image_file
+        item['imgname_rel'] = image_file_rel
+        item['personid'] = int(self.personid[idx])
+        item['extra_info'] = copy.deepcopy(self.extra_info[idx])
+        item['idx'] = idx
+        item['_scale'] = scale
+        return item
+
+
+    @staticmethod
+    def load_tars_as_webdataset(cfg: CfgNode, urls: str|List[str], train: bool,
+            resampled=False,
+            epoch_size=None,
+            cache_dir=None,
+            **kwargs) -> Dataset:
+        """
+        Loads the dataset from a webdataset tar file.
+        """
+
+        IMG_SIZE = cfg.MODEL.IMAGE_SIZE
+        BBOX_SHAPE = cfg.MODEL.get('BBOX_SHAPE', None)
+        MEAN = 255. * np.array(cfg.MODEL.IMAGE_MEAN)
+        STD = 255. * np.array(cfg.MODEL.IMAGE_STD)
+
+        def split_data(source):
+            for item in source:
+                datas = item['data.pyd']
+                for data in datas:
+                    if 'detection.npz' in item:
+                        det_idx = data['extra_info']['detection_npz_idx']
+                        mask = item['detection.npz']['masks'][det_idx]
+                    else:
+                        mask = np.ones_like(item['jpg'][:,:,0], dtype=bool)
+                    yield {
+                        '__key__': item['__key__'],
+                        'jpg': item['jpg'],
+                        'data.pyd': data,
+                        'mask': mask,
+                    }
+
+        def suppress_bad_kps(item, thresh=0.0):
+            if thresh > 0:
+                kp2d = item['data.pyd']['keypoints_2d']
+                kp2d_conf = np.where(kp2d[:, 2] < thresh, 0.0, kp2d[:, 2])
+                item['data.pyd']['keypoints_2d'] = np.concatenate([kp2d[:,:2], kp2d_conf[:,None]], axis=1)
+            return item
+
+        def filter_numkp(item, numkp=4, thresh=0.0):
+            kp_conf = item['data.pyd']['keypoints_2d'][:, 2]
+            return (kp_conf > thresh).sum() > numkp
+
+        def filter_reproj_error(item, thresh=10**4.5):
+            losses = item['data.pyd'].get('extra_info', {}).get('fitting_loss', np.array({})).item()
+            reproj_loss = losses.get('reprojection_loss', None)
+            return reproj_loss is None or reproj_loss < thresh
+
+        def filter_bbox_size(item, thresh=1):
+            bbox_size_min = item['data.pyd']['scale'].min().item() * 200.
+            return bbox_size_min > thresh
+
+        def filter_no_poses(item):
+            return (item['data.pyd']['has_body_pose'] > 0)
+
+        def supress_bad_betas(item, thresh=3):
+            has_betas = item['data.pyd']['has_betas']
+            if thresh > 0 and has_betas:
+                betas_abs = np.abs(item['data.pyd']['betas'])
+                if (betas_abs > thresh).any():
+                    item['data.pyd']['has_betas'] = False
+            return item
+
+        amass_poses_hist100_smooth = load_amass_hist_smooth()
+        def supress_bad_poses(item):
+            has_body_pose = item['data.pyd']['has_body_pose']
+            if has_body_pose:
+                body_pose = item['data.pyd']['body_pose']
+                pose_is_probable = poses_check_probable(torch.from_numpy(body_pose)[None, 3:], amass_poses_hist100_smooth).item()
+                if not pose_is_probable:
+                    item['data.pyd']['has_body_pose'] = False
+            return item
+
+        def poses_betas_simultaneous(item):
+            # We either have both body_pose and betas, or neither
+            has_betas = item['data.pyd']['has_betas']
+            has_body_pose = item['data.pyd']['has_body_pose']
+            item['data.pyd']['has_betas'] = item['data.pyd']['has_body_pose'] = np.array(float((has_body_pose>0) and (has_betas>0)))
+            return item
+
+        def set_betas_for_reg(item):
+            # Always have betas set to true
+            has_betas = item['data.pyd']['has_betas']
+            betas = item['data.pyd']['betas']
+
+            if not (has_betas>0):
+                item['data.pyd']['has_betas'] = np.array(float((True)))
+                item['data.pyd']['betas'] = betas * 0
+            return item
+
+        # Load the dataset
+        if epoch_size is not None:
+            resampled = True
+        corrupt_filter = lambda sample: (sample['__key__'] not in CORRUPT_KEYS)
+        import webdataset as wds
+        dataset = wds.WebDataset(expand_urls(urls),
+                                nodesplitter=wds.split_by_node,
+                                shardshuffle=True,
+                                resampled=resampled,
+                                cache_dir=cache_dir,
+                              ).select(corrupt_filter)
+        if train:
+            dataset = dataset.shuffle(100)
+        dataset = dataset.decode('rgb8').rename(jpg='jpg;jpeg;png')
+
+        # Process the dataset
+        dataset = dataset.compose(split_data)
+
+        # Filter/clean the dataset
+        SUPPRESS_KP_CONF_THRESH = cfg.DATASETS.get('SUPPRESS_KP_CONF_THRESH', 0.0)
+        SUPPRESS_BETAS_THRESH = cfg.DATASETS.get('SUPPRESS_BETAS_THRESH', 0.0)
+        SUPPRESS_BAD_POSES = cfg.DATASETS.get('SUPPRESS_BAD_POSES', False)
+        POSES_BETAS_SIMULTANEOUS = cfg.DATASETS.get('POSES_BETAS_SIMULTANEOUS', False)
+        BETAS_REG = cfg.DATASETS.get('BETAS_REG', False)
+        FILTER_NO_POSES = cfg.DATASETS.get('FILTER_NO_POSES', False)
+        FILTER_NUM_KP = cfg.DATASETS.get('FILTER_NUM_KP', 4)
+        FILTER_NUM_KP_THRESH = cfg.DATASETS.get('FILTER_NUM_KP_THRESH', 0.0)
+        FILTER_REPROJ_THRESH = cfg.DATASETS.get('FILTER_REPROJ_THRESH', 0.0)
+        FILTER_MIN_BBOX_SIZE = cfg.DATASETS.get('FILTER_MIN_BBOX_SIZE', 0.0)
+        if SUPPRESS_KP_CONF_THRESH > 0:
+            dataset = dataset.map(lambda x: suppress_bad_kps(x, thresh=SUPPRESS_KP_CONF_THRESH))
+        if SUPPRESS_BETAS_THRESH > 0:
+            dataset = dataset.map(lambda x: supress_bad_betas(x, thresh=SUPPRESS_BETAS_THRESH))
+        if SUPPRESS_BAD_POSES:
+            dataset = dataset.map(lambda x: supress_bad_poses(x))
+        if POSES_BETAS_SIMULTANEOUS:
+            dataset = dataset.map(lambda x: poses_betas_simultaneous(x))
+        if FILTER_NO_POSES:
+            dataset = dataset.select(lambda x: filter_no_poses(x))
+        if FILTER_NUM_KP > 0:
+            dataset = dataset.select(lambda x: filter_numkp(x, numkp=FILTER_NUM_KP, thresh=FILTER_NUM_KP_THRESH))
+        if FILTER_REPROJ_THRESH > 0:
+            dataset = dataset.select(lambda x: filter_reproj_error(x, thresh=FILTER_REPROJ_THRESH))
+        if FILTER_MIN_BBOX_SIZE > 0:
+            dataset = dataset.select(lambda x: filter_bbox_size(x, thresh=FILTER_MIN_BBOX_SIZE))
+        if BETAS_REG:
+            dataset = dataset.map(lambda x: set_betas_for_reg(x))       # NOTE: Must be at the end
+
+        use_skimage_antialias = cfg.DATASETS.get('USE_SKIMAGE_ANTIALIAS', False)
+        border_mode = {
+            'constant': cv2.BORDER_CONSTANT,
+            'replicate': cv2.BORDER_REPLICATE,
+        }[cfg.DATASETS.get('BORDER_MODE', 'constant')]
+
+        # Process the dataset further
+        dataset = dataset.map(lambda x: ImageDataset.process_webdataset_tar_item(x, train,
+                                                        augm_config=cfg.DATASETS.CONFIG,
+                                                        MEAN=MEAN, STD=STD, IMG_SIZE=IMG_SIZE,
+                                                        BBOX_SHAPE=BBOX_SHAPE,
+                                                        use_skimage_antialias=use_skimage_antialias,
+                                                        border_mode=border_mode,
+                                                        ))
+        if epoch_size is not None:
+            dataset = dataset.with_epoch(epoch_size)
+
+        return dataset
+
+    @staticmethod
+    def process_webdataset_tar_item(item, train, 
+                                    augm_config=None, 
+                                    MEAN=DEFAULT_MEAN, 
+                                    STD=DEFAULT_STD, 
+                                    IMG_SIZE=DEFAULT_IMG_SIZE,
+                                    BBOX_SHAPE=None,
+                                    use_skimage_antialias=False,
+                                    border_mode=cv2.BORDER_CONSTANT,
+                                    ):
+        # Read data from item
+        key = item['__key__']
+        image = item['jpg']
+        data = item['data.pyd']
+        mask = item['mask']
+
+        keypoints_2d = data['keypoints_2d']
+        keypoints_3d = data['keypoints_3d']
+        center = data['center']
+        scale = data['scale']
+        body_pose = data['body_pose']
+        betas = data['betas']
+        has_body_pose = data['has_body_pose']
+        has_betas = data['has_betas']
+        # image_file = data['image_file']
+
+        # Process data
+        orig_keypoints_2d = keypoints_2d.copy()
+        center_x = center[0]
+        center_y = center[1]
+        bbox_size = expand_to_aspect_ratio(scale*200, target_aspect_ratio=BBOX_SHAPE).max()
+        if bbox_size < 1:
+            breakpoint()
+
+
+        smpl_params = {'global_orient': body_pose[:3],
+                    'body_pose': body_pose[3:],
+                    'betas': betas
+                    }
+
+        has_smpl_params = {'global_orient': has_body_pose,
+                        'body_pose': has_body_pose,
+                        'betas': has_betas
+                        }
+
+        smpl_params_is_axis_angle = {'global_orient': True,
+                                    'body_pose': True,
+                                    'betas': False
+                                    }
+
+        augm_config = copy.deepcopy(augm_config)
+        # Crop image and (possibly) perform data augmentation
+        img_rgba = np.concatenate([image, mask.astype(np.uint8)[:,:,None]*255], axis=2)
+        img_patch_rgba, keypoints_2d, keypoints_3d, smpl_params, has_smpl_params, img_size, trans = get_example(img_rgba,
+                                                                                                    center_x, center_y,
+                                                                                                    bbox_size, bbox_size,
+                                                                                                    keypoints_2d, keypoints_3d,
+                                                                                                    smpl_params, has_smpl_params,
+                                                                                                    FLIP_KEYPOINT_PERMUTATION,
+                                                                                                    IMG_SIZE, IMG_SIZE,
+                                                                                                    MEAN, STD, train, augm_config,
+                                                                                                    is_bgr=False, return_trans=True,
+                                                                                                    use_skimage_antialias=use_skimage_antialias,
+                                                                                                    border_mode=border_mode,
+                                                                                                    )
+        img_patch = img_patch_rgba[:3,:,:]
+        mask_patch = (img_patch_rgba[3,:,:] / 255.0).clip(0,1)
+        if (mask_patch < 0.5).all():
+            mask_patch = np.ones_like(mask_patch)
+
+        item = {}
+
+        item['img'] = img_patch
+        item['mask'] = mask_patch
+        # item['img_og'] = image
+        # item['mask_og'] = mask
+        item['keypoints_2d'] = keypoints_2d.astype(np.float32)
+        item['keypoints_3d'] = keypoints_3d.astype(np.float32)
+        item['orig_keypoints_2d'] = orig_keypoints_2d
+        item['box_center'] = center.copy()
+        item['box_size'] = bbox_size
+        item['img_size'] = 1.0 * img_size[::-1].copy()
+        item['smpl_params'] = smpl_params
+        item['has_smpl_params'] = has_smpl_params
+        item['smpl_params_is_axis_angle'] = smpl_params_is_axis_angle
+        item['_scale'] = scale
+        item['_trans'] = trans
+        item['imgname'] = key
+        # item['idx'] = idx
+        return item

humans4d/hmr2/datasets/mocap_dataset.py

@@ -0,0 +1,25 @@
+import numpy as np
+from typing import Dict
+
+class MoCapDataset:
+
+    def __init__(self, dataset_file: str):
+        """
+        Dataset class used for loading a dataset of unpaired SMPL parameter annotations
+        Args:
+            cfg (CfgNode): Model config file.
+            dataset_file (str): Path to npz file containing dataset info.
+        """
+        data = np.load(dataset_file)
+        self.pose = data['body_pose'].astype(np.float32)[:, 3:]
+        self.betas = data['betas'].astype(np.float32)
+        self.length = len(self.pose)
+
+    def __getitem__(self, idx: int) -> Dict:
+        pose = self.pose[idx].copy()
+        betas = self.betas[idx].copy()
+        item = {'body_pose': pose, 'betas': betas}
+        return item
+
+    def __len__(self) -> int:
+        return self.length

humans4d/hmr2/datasets/preprocess/lspet_to_npz.py

@@ -0,0 +1,74 @@
+# Adapted from https://raw.githubusercontent.com/nkolot/SPIN/master/datasets/preprocess/hr_lspet.py
+import os
+import glob
+import numpy as np
+import scipy.io as sio
+# from .read_openpose import read_openpose
+
+def hr_lspet_extract(dataset_path, out_path):
+
+    # training mode
+    png_path = os.path.join(dataset_path, '*.png')
+    imgs = glob.glob(png_path)
+    imgs.sort()
+
+    # structs we use
+    imgnames_, scales_, centers_, parts_, openposes_= [], [], [], [], []
+
+    # scale factor
+    scaleFactor = 1.2
+
+    # annotation files
+    annot_file = os.path.join(dataset_path, 'joints.mat')
+    joints = sio.loadmat(annot_file)['joints']
+
+    # main loop
+    for i, imgname in enumerate(imgs):
+        # image name
+        imgname = imgname.split('/')[-1]
+        # read keypoints
+        part14 = joints[:,:2,i]
+        # scale and center
+        bbox = [min(part14[:,0]), min(part14[:,1]),
+                max(part14[:,0]), max(part14[:,1])]
+        center = [(bbox[2]+bbox[0])/2, (bbox[3]+bbox[1])/2]
+        # scale = scaleFactor*max(bbox[2]-bbox[0], bbox[3]-bbox[1]) # Don't /200
+        scale = scaleFactor*np.array([bbox[2]-bbox[0], bbox[3]-bbox[1]]) # Don't /200
+        # update keypoints
+        part = np.zeros([24,3])
+        part[:14] = np.hstack([part14, np.ones([14,1])])
+
+        # # read openpose detections
+        # json_file = os.path.join(openpose_path, 'hrlspet',
+        #     imgname.replace('.png', '_keypoints.json'))
+        # openpose = read_openpose(json_file, part, 'hrlspet')
+
+        # store the data
+        imgnames_.append(imgname)
+        centers_.append(center)
+        scales_.append(scale)
+        parts_.append(part)
+        # openposes_.append(openpose)
+
+    # Populate extra_keypoints_2d: N,19,3
+    # extra_keypoints_2d[:14] = parts[:14]
+    extra_keypoints_2d = np.zeros((len(parts_), 19, 3))
+    extra_keypoints_2d[:,:14,:] = np.stack(parts_)[:,:14,:3]
+
+    print(f'{extra_keypoints_2d.shape=}')
+
+    # store the data struct
+    if not os.path.isdir(out_path):
+        os.makedirs(out_path)
+    out_file = os.path.join(out_path, 'hr-lspet_train.npz')
+    np.savez(out_file, imgname=imgnames_,
+                       center=centers_,
+                       scale=scales_,
+                       part=parts_,
+                       extra_keypoints_2d=extra_keypoints_2d,
+                    #    openpose=openposes_
+            )
+
+
+if __name__ == '__main__':
+    hr_lspet_extract('/shared/pavlakos/datasets/hr-lspet/', 'hmr2_evaluation_data/')

humans4d/hmr2/datasets/preprocess/posetrack_to_npz.py

@@ -0,0 +1,92 @@
+# Adapted from https://raw.githubusercontent.com/nkolot/SPIN/master/datasets/preprocess/coco.py
+import os
+from os.path import join
+import sys
+import json
+import numpy as np
+from pathlib import Path
+# from .read_openpose import read_openpose
+
+def coco_extract(dataset_path, out_path):
+
+    # # convert joints to global order
+    # joints_idx = [19, 20, 21, 22, 23, 9, 8, 10, 7, 11, 6, 3, 2, 4, 1, 5, 0]
+
+    # bbox expansion factor
+    scaleFactor = 1.2
+
+    # structs we need
+    imgnames_, scales_, centers_, parts_, openposes_ = [], [], [], [], []
+
+    # json annotation file
+    SPLIT='val'
+    json_paths = (Path(dataset_path)/'posetrack_data/annotations'/SPLIT).glob('*.json')
+    for json_path in json_paths:
+        json_data = json.load(open(json_path, 'r'))
+
+        imgs = {}
+        for img in json_data['images']:
+            imgs[img['id']] = img
+
+        for annot in json_data['annotations']:
+            # keypoints processing
+            keypoints = annot['keypoints']
+            keypoints = np.reshape(keypoints, (17,3))
+            keypoints[keypoints[:,2]>0,2] = 1
+            # check if all major body joints are annotated
+            if sum(keypoints[5:,2]>0) < 12:
+                continue
+            # image name
+            image_id = annot['image_id']
+            img_name = str(imgs[image_id]['file_name'])
+            # img_name_full = f'images/{SPLIT}/{json_path.stem}/{img_name}'
+            img_name_full = img_name
+
+            # keypoints
+            part = np.zeros([17,3])
+            # part[joints_idx] = keypoints
+            part = keypoints
+
+            # scale and center
+            bbox = annot['bbox']
+            center = [bbox[0] + bbox[2]/2, bbox[1] + bbox[3]/2]
+            # scale = scaleFactor*max(bbox[2], bbox[3]) # Don't do /200
+            scale = scaleFactor*np.array([bbox[2], bbox[3]]) # Don't /200
+            # # read openpose detections
+            # json_file = os.path.join(openpose_path, 'coco',
+            #     img_name.replace('.jpg', '_keypoints.json'))
+            # openpose = read_openpose(json_file, part, 'coco')
+
+            # store data
+            imgnames_.append(img_name_full)
+            centers_.append(center)
+            scales_.append(scale)
+            parts_.append(part)
+            # openposes_.append(openpose)
+
+
+    # NOTE: Posetrack val doesn't annotate ears (17,18)
+    # But Posetrack does annotate head, neck so that wil have to live in extra_kps.
+    posetrack_to_op_extra = [0, 37, 38, 18, 17, 5, 2, 6, 3, 7, 4, 12, 9, 13, 10, 14, 11] # Will contain 15 keypoints.
+    all_keypoints_2d = np.zeros((len(parts_), 44, 3))
+    all_keypoints_2d[:,posetrack_to_op_extra] = np.stack(parts_)[:,:len(posetrack_to_op_extra),:3]
+    body_keypoints_2d = all_keypoints_2d[:,:25,:]
+    extra_keypoints_2d = all_keypoints_2d[:,25:,:]
+
+    print(f'{extra_keypoints_2d.shape=}')
+
+    # store the data struct
+    if not os.path.isdir(out_path):
+        os.makedirs(out_path)
+    out_file = os.path.join(out_path, f'posetrack_2018_{SPLIT}.npz')
+    np.savez(out_file, imgname=imgnames_,
+                       center=centers_,
+                       scale=scales_,
+                       part=parts_,
+                       body_keypoints_2d=body_keypoints_2d,
+                       extra_keypoints_2d=extra_keypoints_2d,
+                    #    openpose=openposes_
+            )
+
+if __name__ == '__main__':
+    coco_extract('/shared/pavlakos/datasets/posetrack/posetrack2018/', 'hmr2_evaluation_data/')

humans4d/hmr2/datasets/smplh_prob_filter.py

@@ -0,0 +1,154 @@
+import os
+import numpy as np
+import torch
+import torch.nn.functional as F
+
+JOINT_NAMES = [
+    'left_hip', 
+    'right_hip', 
+    'spine1', 
+    'left_knee', 
+    'right_knee', 
+    'spine2', 
+    'left_ankle', 
+    'right_ankle', 
+    'spine3', 
+    'left_foot', 
+    'right_foot', 
+    'neck', 
+    'left_collar', 
+    'right_collar', 
+    'head', 
+    'left_shoulder', 
+    'right_shoulder', 
+    'left_elbow', 
+    'right_elbow', 
+    'left_wrist', 
+    'right_wrist'
+]
+
+# Manually chosen probability density thresholds for each joint
+# Probablities computed using SIGMA=2 gaussian blur on AMASS pose 3D histogram for range (-pi,pi) with 100x100x100 bins
+JOINT_NAME_PROB_THRESHOLDS = {
+    'left_hip': 5e-5,
+    'right_hip': 5e-5,
+    'spine1': 2e-3,
+    'left_knee': 5e-6,
+    'right_knee': 5e-6,
+    'spine2': 0.01,
+    'left_ankle': 5e-6,
+    'right_ankle': 5e-6,
+    'spine3': 0.025,
+    'left_foot': 0,
+    'right_foot': 0,
+    'neck': 2e-4,
+    'left_collar': 4.5e-4 ,
+    'right_collar': 4.5e-4,
+    'head': 5e-4,
+    'left_shoulder': 2e-4,
+    'right_shoulder': 2e-4,
+    'left_elbow': 4e-5,
+    'right_elbow': 4e-5,
+    'left_wrist': 1e-3,
+    'right_wrist': 1e-3,
+}
+
+JOINT_IDX_PROB_THRESHOLDS = torch.tensor([JOINT_NAME_PROB_THRESHOLDS[joint_name] for joint_name in JOINT_NAMES])
+
+###################################################################
+POSE_RANGE_MIN = -np.pi
+POSE_RANGE_MAX = np.pi
+
+# Create 21x100x100x100 histogram of all 21 AMASS body pose joints using `create_pose_hist(amass_poses, nbins=100)`
+AMASS_HIST100_PATH = 'hmr2_training_data/amass_poses_hist100_SMPL+H_G.npy'
+if not os.path.exists(AMASS_HIST100_PATH):
+    AMASS_HIST100_PATH = '/shared/shubham/code/hmr2023/amass_poses_hist100_SMPL+H_G.npy'
+if not os.path.exists(AMASS_HIST100_PATH):
+    AMASS_HIST100_PATH = '/fsx/shubham/code/stable-humans/notebooks/amass_poses_hist100_SMPL+H_G.npy'
+
+
+def create_pose_hist(poses: np.ndarray, nbins: int = 100) -> np.ndarray:
+    N,K,C = poses.shape
+    assert C==3, poses.shape
+    poses_21x3 = normalize_axis_angle(torch.fromnumpy(poses).view(N*K,3)).numpy().reshape(N, K, 3)
+    assert (poses_21x3 > -np.pi).all() and (poses_21x3 < np.pi).all()
+
+    Hs, Es = [], []
+    for i in range(K):
+        H, edges = np.histogramdd(poses_21x3[:, i, :], bins=nbins, range=[(-np.pi, np.pi)]*3)
+        Hs.append(H)
+        Es.append(edges)
+    Hs = np.stack(Hs, axis=0)
+    return Hs
+
+def load_amass_hist_smooth(sigma=2) -> torch.Tensor:
+    amass_poses_hist100 = np.load(AMASS_HIST100_PATH)
+    amass_poses_hist100 = torch.from_numpy(amass_poses_hist100)
+    assert amass_poses_hist100.shape == (21,100,100,100)
+
+    nbins = amass_poses_hist100.shape[1]
+    amass_poses_hist100 = amass_poses_hist100/amass_poses_hist100.sum() / (2*np.pi/nbins)**3
+
+    # Gaussian filter on amass_poses_hist100
+    from scipy.ndimage import gaussian_filter
+    amass_poses_hist100_smooth = gaussian_filter(amass_poses_hist100.numpy(), sigma=sigma, mode='constant')
+    amass_poses_hist100_smooth = torch.from_numpy(amass_poses_hist100_smooth)
+    return amass_poses_hist100_smooth
+
+# Normalize axis angle representation s.t. angle is in [-pi, pi]
+def normalize_axis_angle(poses: torch.Tensor) -> torch.Tensor:
+    # poses: N, 3
+    # print(f'normalize_axis_angle ...')
+    assert poses.shape[1] == 3, poses.shape
+    angle = poses.norm(dim=1)
+    axis = F.normalize(poses, p=2, dim=1, eps=1e-8)
+    
+    angle_fixed = angle.clone()
+    axis_fixed = axis.clone()
+
+    eps = 1e-6
+    ii = 0
+    while True:
+        # print(f'normalize_axis_angle iter {ii}')
+        ii += 1
+        angle_too_big = (angle_fixed > np.pi + eps)
+        if not angle_too_big.any():
+            break
+        
+        angle_fixed[angle_too_big] -= 2 * np.pi
+        angle_too_small = (angle_fixed < -eps)
+        axis_fixed[angle_too_small] *= -1
+        angle_fixed[angle_too_small] *= -1
+    
+    return axis_fixed * angle_fixed[:,None]
+
+def poses_to_joint_probs(poses: torch.Tensor, amass_poses_100_smooth: torch.Tensor) -> torch.Tensor:
+    # poses: Nx69
+    # amass_poses_100_smooth: 21xBINSxBINSxBINS
+    # returns: poses_prob: Nx21
+    N=poses.shape[0]
+    assert poses.shape == (N,69)
+    poses = poses[:,:63].reshape(N*21,3)
+
+    nbins = amass_poses_100_smooth.shape[1]
+    assert amass_poses_100_smooth.shape == (21,nbins,nbins,nbins)
+
+    poses_bin = (poses - POSE_RANGE_MIN) / (POSE_RANGE_MAX - POSE_RANGE_MIN) * (nbins - 1e-6)
+    poses_bin = poses_bin.long().clip(0, nbins-1)
+    joint_id = torch.arange(21, device=poses.device).view(1,21).expand(N,21).reshape(N*21)
+    poses_prob = amass_poses_100_smooth[joint_id, poses_bin[:,0], poses_bin[:,1], poses_bin[:,2]]
+
+    poses_bad = ((poses < POSE_RANGE_MIN) | (poses >= POSE_RANGE_MAX)).any(dim=1)
+    poses_prob[poses_bad] = 0
+
+    return poses_prob.view(N,21)
+
+def poses_check_probable(
+        poses: torch.Tensor, 
+        amass_poses_100_smooth: torch.Tensor, 
+        prob_thresholds: torch.Tensor = JOINT_IDX_PROB_THRESHOLDS
+    ) -> torch.Tensor:
+    N,C=poses.shape
+    poses_norm = normalize_axis_angle(poses.reshape(N*(C//3),3)).reshape(N,C)
+    poses_prob = poses_to_joint_probs(poses_norm, amass_poses_100_smooth)
+    return (poses_prob > prob_thresholds).all(dim=1)

humans4d/hmr2/datasets/vitdet_dataset.py

@@ -0,0 +1,88 @@
+from typing import Dict
+
+import cv2
+import numpy as np
+from skimage.filters import gaussian
+from yacs.config import CfgNode
+import torch
+
+from .utils import (convert_cvimg_to_tensor,
+                    expand_to_aspect_ratio,
+                    generate_image_patch_cv2)
+
+DEFAULT_MEAN = 255. * np.array([0.485, 0.456, 0.406])
+DEFAULT_STD = 255. * np.array([0.229, 0.224, 0.225])
+
+class ViTDetDataset(torch.utils.data.Dataset):
+
+    def __init__(self,
+                 cfg: CfgNode,
+                 img_cv2: np.array,
+                 boxes: np.array,
+                 train: bool = False,
+                 **kwargs):
+        super().__init__()
+        self.cfg = cfg
+        self.img_cv2 = img_cv2
+        # self.boxes = boxes
+
+        assert train == False, "ViTDetDataset is only for inference"
+        self.train = train
+        self.img_size = cfg.MODEL.IMAGE_SIZE
+        self.mean = 255. * np.array(self.cfg.MODEL.IMAGE_MEAN)
+        self.std = 255. * np.array(self.cfg.MODEL.IMAGE_STD)
+
+        # Preprocess annotations
+        boxes = boxes.astype(np.float32)
+        self.center = (boxes[:, 2:4] + boxes[:, 0:2]) / 2.0
+        self.scale = (boxes[:, 2:4] - boxes[:, 0:2]) / 200.0
+        self.personid = np.arange(len(boxes), dtype=np.int32)
+
+    def __len__(self) -> int:
+        return len(self.personid)
+
+    def __getitem__(self, idx: int) -> Dict[str, np.array]:
+
+        center = self.center[idx].copy()
+        center_x = center[0]
+        center_y = center[1]
+
+        scale = self.scale[idx]
+        BBOX_SHAPE = self.cfg.MODEL.get('BBOX_SHAPE', None)
+        bbox_size = expand_to_aspect_ratio(scale*200, target_aspect_ratio=BBOX_SHAPE).max()
+
+        patch_width = patch_height = self.img_size
+
+        # 3. generate image patch
+        # if use_skimage_antialias:
+        cvimg = self.img_cv2.copy()
+        if True:
+            # Blur image to avoid aliasing artifacts
+            downsampling_factor = ((bbox_size*1.0) / patch_width)
+            print(f'{downsampling_factor=}')
+            downsampling_factor = downsampling_factor / 2.0
+            if downsampling_factor > 1.1:
+                cvimg  = gaussian(cvimg, sigma=(downsampling_factor-1)/2, channel_axis=2, preserve_range=True)
+
+
+        img_patch_cv, trans = generate_image_patch_cv2(cvimg,
+                                                    center_x, center_y,
+                                                    bbox_size, bbox_size,
+                                                    patch_width, patch_height,
+                                                    False, 1.0, 0,
+                                                    border_mode=cv2.BORDER_CONSTANT)
+        img_patch_cv = img_patch_cv[:, :, ::-1]
+        img_patch = convert_cvimg_to_tensor(img_patch_cv)
+
+        # apply normalization
+        for n_c in range(min(self.img_cv2.shape[2], 3)):
+            img_patch[n_c, :, :] = (img_patch[n_c, :, :] - self.mean[n_c]) / self.std[n_c]
+
+        item = {
+            'img': img_patch,
+            'personid': int(self.personid[idx]),
+        }
+        item['box_center'] = self.center[idx].copy()
+        item['box_size'] = bbox_size
+        item['img_size'] = 1.0 * np.array([cvimg.shape[1], cvimg.shape[0]])
+        return item

humans4d/hmr2/models/__init__.py

@@ -0,0 +1,94 @@
+import os
+
+from .smpl_wrapper import SMPL
+from .hmr2 import HMR2
+from .discriminator import Discriminator
+
+from ..utils.download import cache_url
+from ..configs import CACHE_DIR_4DHUMANS
+from .... import REPO_PATH
+
+
+def download_models(folder=CACHE_DIR_4DHUMANS):
+    """Download checkpoints and files for running inference.
+    """
+    os.makedirs(folder, exist_ok=True)
+    download_files = {
+        "hmr2_data.tar.gz"      : ["https://people.eecs.berkeley.edu/~jathushan/projects/4dhumans/hmr2_data.tar.gz", folder],
+    }
+
+    for file_name, url in download_files.items():
+        output_path = os.path.join(url[1], file_name)
+        if not os.path.exists(output_path):
+            print("Downloading file: " + file_name)
+            # output = gdown.cached_download(url[0], output_path, fuzzy=True)
+            output = cache_url(url[0], output_path)
+            assert os.path.exists(output_path), f"{output} does not exist"
+
+            # if ends with tar.gz, tar -xzf
+            if file_name.endswith(".tar.gz"):
+                print("Extracting file: " + file_name)
+                os.system("tar -xvf " + output_path + " -C " + url[1])
+
+def check_smpl_exists(smpl_dir):
+    import os
+    candidates = [
+        f'{smpl_dir}/SMPL_NEUTRAL.pkl',
+        f'{smpl_dir}/basicModel_neutral_lbs_10_207_0_v1.0.0.pkl',
+    ]
+    candidates_exist = [os.path.exists(c) for c in candidates]
+    if not any(candidates_exist):
+        raise FileNotFoundError(f"SMPL model not found. Please download it from https://smplify.is.tue.mpg.de/ and place it at {candidates[1]}")
+
+    # Code edxpects SMPL model at CACHE_DIR_4DHUMANS/data/smpl/SMPL_NEUTRAL.pkl. Copy there if needed
+    if (not candidates_exist[0]) and candidates_exist[1]:
+        convert_pkl(candidates[1], candidates[0])
+
+    return True
+
+# Convert SMPL pkl file to be compatible with Python 3
+# Script is from https://rebeccabilbro.github.io/convert-py2-pickles-to-py3/
+def convert_pkl(old_pkl, new_pkl):
+    """
+    Convert a Python 2 pickle to Python 3
+    """
+    import dill
+    import pickle
+
+    # Convert Python 2 "ObjectType" to Python 3 object
+    dill._dill._reverse_typemap["ObjectType"] = object
+
+    # Open the pickle using latin1 encoding
+    with open(old_pkl, "rb") as f:
+        loaded = pickle.load(f, encoding="latin1")
+
+    # Re-save as Python 3 pickle
+    with open(new_pkl, "wb") as outfile:
+        pickle.dump(loaded, outfile)
+
+
+def load_hmr2(checkpoint_path, smpl_dir):
+    from pathlib import Path
+    from ..configs import get_config
+    model_cfg = os.path.join(REPO_PATH, 'configs', 'model_config.yaml')
+    model_cfg = get_config(model_cfg, update_cachedir=True)
+
+    model_cfg.SMPL.MODEL_PATH = smpl_dir
+    model_cfg.SMPL.JOINT_REGRESSION_EXTRA = os.path.join(smpl_dir, 'SMPL_to_J19.pkl')
+    model_cfg.SMPL.JOINT_REGRESSOR_EXTRA = os.path.join(smpl_dir, 'SMPL_to_J19.pkl')
+    model_cfg.SMPL.MEAN_PARAMS = os.path.join(smpl_dir, 'smpl_mean_params.npz')
+
+    model_cfg.freeze()
+
+    # Override some config values, to crop bbox correctly
+    if (model_cfg.MODEL.BACKBONE.TYPE == 'vit') and ('BBOX_SHAPE' not in model_cfg.MODEL):
+        model_cfg.defrost()
+        assert model_cfg.MODEL.IMAGE_SIZE == 256, f"MODEL.IMAGE_SIZE ({model_cfg.MODEL.IMAGE_SIZE}) should be 256 for ViT backbone"
+        model_cfg.MODEL.BBOX_SHAPE = [192,256]
+        model_cfg.freeze()
+
+    # Ensure SMPL model exists
+    check_smpl_exists(smpl_dir)
+
+    model = HMR2.load_from_checkpoint(checkpoint_path, strict=False, cfg=model_cfg)
+    return model, model_cfg

humans4d/hmr2/models/backbones/__init__.py

@@ -0,0 +1,7 @@
+from .vit import vit
+
+def create_backbone(cfg):
+    if cfg.MODEL.BACKBONE.TYPE == 'vit':
+        return vit(cfg)
+    else:
+        raise NotImplementedError('Backbone type is not implemented')

humans4d/hmr2/models/backbones/vit.py

@@ -0,0 +1,348 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+import math
+
+import torch
+from functools import partial
+import torch.nn as nn
+import torch.nn.functional as F
+import torch.utils.checkpoint as checkpoint
+
+from timm.models.layers import drop_path, to_2tuple, trunc_normal_
+
+def vit(cfg):
+    return ViT(
+                img_size=(256, 192),
+                patch_size=16,
+                embed_dim=1280,
+                depth=32,
+                num_heads=16,
+                ratio=1,
+                use_checkpoint=False,
+                mlp_ratio=4,
+                qkv_bias=True,
+                drop_path_rate=0.55,
+            )
+
+def get_abs_pos(abs_pos, h, w, ori_h, ori_w, has_cls_token=True):
+    """
+    Calculate absolute positional embeddings. If needed, resize embeddings and remove cls_token
+        dimension for the original embeddings.
+    Args:
+        abs_pos (Tensor): absolute positional embeddings with (1, num_position, C).
+        has_cls_token (bool): If true, has 1 embedding in abs_pos for cls token.
+        hw (Tuple): size of input image tokens.
+
+    Returns:
+        Absolute positional embeddings after processing with shape (1, H, W, C)
+    """
+    cls_token = None
+    B, L, C = abs_pos.shape
+    if has_cls_token:
+        cls_token = abs_pos[:, 0:1]
+        abs_pos = abs_pos[:, 1:]
+
+    if ori_h != h or ori_w != w:
+        new_abs_pos = F.interpolate(
+            abs_pos.reshape(1, ori_h, ori_w, -1).permute(0, 3, 1, 2),
+            size=(h, w),
+            mode="bicubic",
+            align_corners=False,
+        ).permute(0, 2, 3, 1).reshape(B, -1, C)
+
+    else:
+        new_abs_pos = abs_pos
+    
+    if cls_token is not None:
+        new_abs_pos = torch.cat([cls_token, new_abs_pos], dim=1)
+    return new_abs_pos
+
+class DropPath(nn.Module):
+    """Drop paths (Stochastic Depth) per sample  (when applied in main path of residual blocks).
+    """
+    def __init__(self, drop_prob=None):
+        super(DropPath, self).__init__()
+        self.drop_prob = drop_prob
+
+    def forward(self, x):
+        return drop_path(x, self.drop_prob, self.training)
+    
+    def extra_repr(self):
+        return 'p={}'.format(self.drop_prob)
+
+class Mlp(nn.Module):
+    def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.):
+        super().__init__()
+        out_features = out_features or in_features
+        hidden_features = hidden_features or in_features
+        self.fc1 = nn.Linear(in_features, hidden_features)
+        self.act = act_layer()
+        self.fc2 = nn.Linear(hidden_features, out_features)
+        self.drop = nn.Dropout(drop)
+
+    def forward(self, x):
+        x = self.fc1(x)
+        x = self.act(x)
+        x = self.fc2(x)
+        x = self.drop(x)
+        return x
+
+class Attention(nn.Module):
+    def __init__(
+            self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0.,
+            proj_drop=0., attn_head_dim=None,):
+        super().__init__()
+        self.num_heads = num_heads
+        head_dim = dim // num_heads
+        self.dim = dim
+
+        if attn_head_dim is not None:
+            head_dim = attn_head_dim
+        all_head_dim = head_dim * self.num_heads
+
+        self.scale = qk_scale or head_dim ** -0.5
+
+        self.qkv = nn.Linear(dim, all_head_dim * 3, bias=qkv_bias)
+
+        self.attn_drop = nn.Dropout(attn_drop)
+        self.proj = nn.Linear(all_head_dim, dim)
+        self.proj_drop = nn.Dropout(proj_drop)
+
+    def forward(self, x):
+        B, N, C = x.shape
+        qkv = self.qkv(x)
+        qkv = qkv.reshape(B, N, 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)
+        q, k, v = qkv[0], qkv[1], qkv[2]   # make torchscript happy (cannot use tensor as tuple)
+
+        q = q * self.scale
+        attn = (q @ k.transpose(-2, -1))
+
+        attn = attn.softmax(dim=-1)
+        attn = self.attn_drop(attn)
+
+        x = (attn @ v).transpose(1, 2).reshape(B, N, -1)
+        x = self.proj(x)
+        x = self.proj_drop(x)
+
+        return x
+
+class Block(nn.Module):
+
+    def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, 
+                 drop=0., attn_drop=0., drop_path=0., act_layer=nn.GELU, 
+                 norm_layer=nn.LayerNorm, attn_head_dim=None
+                 ):
+        super().__init__()
+        
+        self.norm1 = norm_layer(dim)
+        self.attn = Attention(
+            dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale,
+            attn_drop=attn_drop, proj_drop=drop, attn_head_dim=attn_head_dim
+            )
+
+        # NOTE: drop path for stochastic depth, we shall see if this is better than dropout here
+        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
+        self.norm2 = norm_layer(dim)
+        mlp_hidden_dim = int(dim * mlp_ratio)
+        self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
+
+    def forward(self, x):
+        x = x + self.drop_path(self.attn(self.norm1(x)))
+        x = x + self.drop_path(self.mlp(self.norm2(x)))
+        return x
+
+
+class PatchEmbed(nn.Module):
+    """ Image to Patch Embedding
+    """
+    def __init__(self, img_size=224, patch_size=16, in_chans=3, embed_dim=768, ratio=1):
+        super().__init__()
+        img_size = to_2tuple(img_size)
+        patch_size = to_2tuple(patch_size)
+        num_patches = (img_size[1] // patch_size[1]) * (img_size[0] // patch_size[0]) * (ratio ** 2)
+        self.patch_shape = (int(img_size[0] // patch_size[0] * ratio), int(img_size[1] // patch_size[1] * ratio))
+        self.origin_patch_shape = (int(img_size[0] // patch_size[0]), int(img_size[1] // patch_size[1]))
+        self.img_size = img_size
+        self.patch_size = patch_size
+        self.num_patches = num_patches
+
+        self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=(patch_size[0] // ratio), padding=4 + 2 * (ratio//2-1))
+
+    def forward(self, x, **kwargs):
+        B, C, H, W = x.shape
+        x = self.proj(x)
+        Hp, Wp = x.shape[2], x.shape[3]
+
+        x = x.flatten(2).transpose(1, 2)
+        return x, (Hp, Wp)
+
+
+class HybridEmbed(nn.Module):
+    """ CNN Feature Map Embedding
+    Extract feature map from CNN, flatten, project to embedding dim.
+    """
+    def __init__(self, backbone, img_size=224, feature_size=None, in_chans=3, embed_dim=768):
+        super().__init__()
+        assert isinstance(backbone, nn.Module)
+        img_size = to_2tuple(img_size)
+        self.img_size = img_size
+        self.backbone = backbone
+        if feature_size is None:
+            with torch.no_grad():
+                training = backbone.training
+                if training:
+                    backbone.eval()
+                o = self.backbone(torch.zeros(1, in_chans, img_size[0], img_size[1]))[-1]
+                feature_size = o.shape[-2:]
+                feature_dim = o.shape[1]
+                backbone.train(training)
+        else:
+            feature_size = to_2tuple(feature_size)
+            feature_dim = self.backbone.feature_info.channels()[-1]
+        self.num_patches = feature_size[0] * feature_size[1]
+        self.proj = nn.Linear(feature_dim, embed_dim)
+
+    def forward(self, x):
+        x = self.backbone(x)[-1]
+        x = x.flatten(2).transpose(1, 2)
+        x = self.proj(x)
+        return x
+
+
+class ViT(nn.Module):
+
+    def __init__(self,
+                 img_size=224, patch_size=16, in_chans=3, num_classes=80, embed_dim=768, depth=12,
+                 num_heads=12, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop_rate=0., attn_drop_rate=0.,
+                 drop_path_rate=0., hybrid_backbone=None, norm_layer=None, use_checkpoint=False, 
+                 frozen_stages=-1, ratio=1, last_norm=True,
+                 patch_padding='pad', freeze_attn=False, freeze_ffn=False,
+                 ):
+        # Protect mutable default arguments
+        super(ViT, self).__init__()
+        norm_layer = norm_layer or partial(nn.LayerNorm, eps=1e-6)
+        self.num_classes = num_classes
+        self.num_features = self.embed_dim = embed_dim  # num_features for consistency with other models
+        self.frozen_stages = frozen_stages
+        self.use_checkpoint = use_checkpoint
+        self.patch_padding = patch_padding
+        self.freeze_attn = freeze_attn
+        self.freeze_ffn = freeze_ffn
+        self.depth = depth
+
+        if hybrid_backbone is not None:
+            self.patch_embed = HybridEmbed(
+                hybrid_backbone, img_size=img_size, in_chans=in_chans, embed_dim=embed_dim)
+        else:
+            self.patch_embed = PatchEmbed(
+                img_size=img_size, patch_size=patch_size, in_chans=in_chans, embed_dim=embed_dim, ratio=ratio)
+        num_patches = self.patch_embed.num_patches
+
+        # since the pretraining model has class token
+        self.pos_embed = nn.Parameter(torch.zeros(1, num_patches + 1, embed_dim))
+
+        dpr = [x.item() for x in torch.linspace(0, drop_path_rate, depth)]  # stochastic depth decay rule
+
+        self.blocks = nn.ModuleList([
+            Block(
+                dim=embed_dim, num_heads=num_heads, mlp_ratio=mlp_ratio, qkv_bias=qkv_bias, qk_scale=qk_scale,
+                drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[i], norm_layer=norm_layer,
+                )
+            for i in range(depth)])
+
+        self.last_norm = norm_layer(embed_dim) if last_norm else nn.Identity()
+
+        if self.pos_embed is not None:
+            trunc_normal_(self.pos_embed, std=.02)
+
+        self._freeze_stages()
+
+    def _freeze_stages(self):
+        """Freeze parameters."""
+        if self.frozen_stages >= 0:
+            self.patch_embed.eval()
+            for param in self.patch_embed.parameters():
+                param.requires_grad = False
+
+        for i in range(1, self.frozen_stages + 1):
+            m = self.blocks[i]
+            m.eval()
+            for param in m.parameters():
+                param.requires_grad = False
+
+        if self.freeze_attn:
+            for i in range(0, self.depth):
+                m = self.blocks[i]
+                m.attn.eval()
+                m.norm1.eval()
+                for param in m.attn.parameters():
+                    param.requires_grad = False
+                for param in m.norm1.parameters():
+                    param.requires_grad = False
+
+        if self.freeze_ffn:
+            self.pos_embed.requires_grad = False
+            self.patch_embed.eval()
+            for param in self.patch_embed.parameters():
+                param.requires_grad = False
+            for i in range(0, self.depth):
+                m = self.blocks[i]
+                m.mlp.eval()
+                m.norm2.eval()
+                for param in m.mlp.parameters():
+                    param.requires_grad = False
+                for param in m.norm2.parameters():
+                    param.requires_grad = False
+
+    def init_weights(self):
+        """Initialize the weights in backbone.
+        Args:
+            pretrained (str, optional): Path to pre-trained weights.
+                Defaults to None.
+        """
+        def _init_weights(m):
+            if isinstance(m, nn.Linear):
+                trunc_normal_(m.weight, std=.02)
+                if isinstance(m, nn.Linear) and m.bias is not None:
+                    nn.init.constant_(m.bias, 0)
+            elif isinstance(m, nn.LayerNorm):
+                nn.init.constant_(m.bias, 0)
+                nn.init.constant_(m.weight, 1.0)
+
+        self.apply(_init_weights)
+
+    def get_num_layers(self):
+        return len(self.blocks)
+
+    @torch.jit.ignore
+    def no_weight_decay(self):
+        return {'pos_embed', 'cls_token'}
+
+    def forward_features(self, x):
+        B, C, H, W = x.shape
+        x, (Hp, Wp) = self.patch_embed(x)
+
+        if self.pos_embed is not None:
+            # fit for multiple GPU training
+            # since the first element for pos embed (sin-cos manner) is zero, it will cause no difference
+            x = x + self.pos_embed[:, 1:] + self.pos_embed[:, :1]
+
+        for blk in self.blocks:
+            if self.use_checkpoint:
+                x = checkpoint.checkpoint(blk, x)
+            else:
+                x = blk(x)
+
+        x = self.last_norm(x)
+
+        xp = x.permute(0, 2, 1).reshape(B, -1, Hp, Wp).contiguous()
+
+        return xp
+
+    def forward(self, x):
+        x = self.forward_features(x)
+        return x
+
+    def train(self, mode=True):
+        """Convert the model into training mode."""
+        super().train(mode)
+        self._freeze_stages()

humans4d/hmr2/models/backbones/vit_vitpose.py

@@ -0,0 +1,17 @@
+# import mmcv
+# import mmpose
+# from mmpose.models import build_posenet
+# from mmcv.runner import load_checkpoint
+# from pathlib import Path
+
+# def vit(cfg):
+#     vitpose_dir = Path(mmpose.__file__).parent.parent
+#     config = f'{vitpose_dir}/configs/body/2d_kpt_sview_rgb_img/topdown_heatmap/coco/ViTPose_huge_coco_256x192.py'
+#     # checkpoint = f'{vitpose_dir}/models/vitpose-h-multi-coco.pth'
+
+#     config = mmcv.Config.fromfile(config)
+#     config.model.pretrained = None
+#     model = build_posenet(config.model)
+#     # load_checkpoint(model, checkpoint, map_location='cpu')
+
+#     return model.backbone

humans4d/hmr2/models/components/pose_transformer.py

@@ -0,0 +1,358 @@
+from inspect import isfunction
+from typing import Callable, Optional
+
+import torch
+from einops import rearrange
+from einops.layers.torch import Rearrange
+from torch import nn
+
+from .t_cond_mlp import (
+    AdaptiveLayerNorm1D,
+    FrequencyEmbedder,
+    normalization_layer,
+)
+# from .vit import Attention, FeedForward
+
+
+def exists(val):
+    return val is not None
+
+
+def default(val, d):
+    if exists(val):
+        return val
+    return d() if isfunction(d) else d
+
+
+class PreNorm(nn.Module):
+    def __init__(self, dim: int, fn: Callable, norm: str = "layer", norm_cond_dim: int = -1):
+        super().__init__()
+        self.norm = normalization_layer(norm, dim, norm_cond_dim)
+        self.fn = fn
+
+    def forward(self, x: torch.Tensor, *args, **kwargs):
+        if isinstance(self.norm, AdaptiveLayerNorm1D):
+            return self.fn(self.norm(x, *args), **kwargs)
+        else:
+            return self.fn(self.norm(x), **kwargs)
+
+
+class FeedForward(nn.Module):
+    def __init__(self, dim, hidden_dim, dropout=0.0):
+        super().__init__()
+        self.net = nn.Sequential(
+            nn.Linear(dim, hidden_dim),
+            nn.GELU(),
+            nn.Dropout(dropout),
+            nn.Linear(hidden_dim, dim),
+            nn.Dropout(dropout),
+        )
+
+    def forward(self, x):
+        return self.net(x)
+
+
+class Attention(nn.Module):
+    def __init__(self, dim, heads=8, dim_head=64, dropout=0.0):
+        super().__init__()
+        inner_dim = dim_head * heads
+        project_out = not (heads == 1 and dim_head == dim)
+
+        self.heads = heads
+        self.scale = dim_head**-0.5
+
+        self.attend = nn.Softmax(dim=-1)
+        self.dropout = nn.Dropout(dropout)
+
+        self.to_qkv = nn.Linear(dim, inner_dim * 3, bias=False)
+
+        self.to_out = (
+            nn.Sequential(nn.Linear(inner_dim, dim), nn.Dropout(dropout))
+            if project_out
+            else nn.Identity()
+        )
+
+    def forward(self, x):
+        qkv = self.to_qkv(x).chunk(3, dim=-1)
+        q, k, v = map(lambda t: rearrange(t, "b n (h d) -> b h n d", h=self.heads), qkv)
+
+        dots = torch.matmul(q, k.transpose(-1, -2)) * self.scale
+
+        attn = self.attend(dots)
+        attn = self.dropout(attn)
+
+        out = torch.matmul(attn, v)
+        out = rearrange(out, "b h n d -> b n (h d)")
+        return self.to_out(out)
+
+
+class CrossAttention(nn.Module):
+    def __init__(self, dim, context_dim=None, heads=8, dim_head=64, dropout=0.0):
+        super().__init__()
+        inner_dim = dim_head * heads
+        project_out = not (heads == 1 and dim_head == dim)
+
+        self.heads = heads
+        self.scale = dim_head**-0.5
+
+        self.attend = nn.Softmax(dim=-1)
+        self.dropout = nn.Dropout(dropout)
+
+        context_dim = default(context_dim, dim)
+        self.to_kv = nn.Linear(context_dim, inner_dim * 2, bias=False)
+        self.to_q = nn.Linear(dim, inner_dim, bias=False)
+
+        self.to_out = (
+            nn.Sequential(nn.Linear(inner_dim, dim), nn.Dropout(dropout))
+            if project_out
+            else nn.Identity()
+        )
+
+    def forward(self, x, context=None):
+        context = default(context, x)
+        k, v = self.to_kv(context).chunk(2, dim=-1)
+        q = self.to_q(x)
+        q, k, v = map(lambda t: rearrange(t, "b n (h d) -> b h n d", h=self.heads), [q, k, v])
+
+        dots = torch.matmul(q, k.transpose(-1, -2)) * self.scale
+
+        attn = self.attend(dots)
+        attn = self.dropout(attn)
+
+        out = torch.matmul(attn, v)
+        out = rearrange(out, "b h n d -> b n (h d)")
+        return self.to_out(out)
+
+
+class Transformer(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        depth: int,
+        heads: int,
+        dim_head: int,
+        mlp_dim: int,
+        dropout: float = 0.0,
+        norm: str = "layer",
+        norm_cond_dim: int = -1,
+    ):
+        super().__init__()
+        self.layers = nn.ModuleList([])
+        for _ in range(depth):
+            sa = Attention(dim, heads=heads, dim_head=dim_head, dropout=dropout)
+            ff = FeedForward(dim, mlp_dim, dropout=dropout)
+            self.layers.append(
+                nn.ModuleList(
+                    [
+                        PreNorm(dim, sa, norm=norm, norm_cond_dim=norm_cond_dim),
+                        PreNorm(dim, ff, norm=norm, norm_cond_dim=norm_cond_dim),
+                    ]
+                )
+            )
+
+    def forward(self, x: torch.Tensor, *args):
+        for attn, ff in self.layers:
+            x = attn(x, *args) + x
+            x = ff(x, *args) + x
+        return x
+
+
+class TransformerCrossAttn(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        depth: int,
+        heads: int,
+        dim_head: int,
+        mlp_dim: int,
+        dropout: float = 0.0,
+        norm: str = "layer",
+        norm_cond_dim: int = -1,
+        context_dim: Optional[int] = None,
+    ):
+        super().__init__()
+        self.layers = nn.ModuleList([])
+        for _ in range(depth):
+            sa = Attention(dim, heads=heads, dim_head=dim_head, dropout=dropout)
+            ca = CrossAttention(
+                dim, context_dim=context_dim, heads=heads, dim_head=dim_head, dropout=dropout
+            )
+            ff = FeedForward(dim, mlp_dim, dropout=dropout)
+            self.layers.append(
+                nn.ModuleList(
+                    [
+                        PreNorm(dim, sa, norm=norm, norm_cond_dim=norm_cond_dim),
+                        PreNorm(dim, ca, norm=norm, norm_cond_dim=norm_cond_dim),
+                        PreNorm(dim, ff, norm=norm, norm_cond_dim=norm_cond_dim),
+                    ]
+                )
+            )
+
+    def forward(self, x: torch.Tensor, *args, context=None, context_list=None):
+        if context_list is None:
+            context_list = [context] * len(self.layers)
+        if len(context_list) != len(self.layers):
+            raise ValueError(f"len(context_list) != len(self.layers) ({len(context_list)} != {len(self.layers)})")
+
+        for i, (self_attn, cross_attn, ff) in enumerate(self.layers):
+            x = self_attn(x, *args) + x
+            x = cross_attn(x, *args, context=context_list[i]) + x
+            x = ff(x, *args) + x
+        return x
+
+
+class DropTokenDropout(nn.Module):
+    def __init__(self, p: float = 0.1):
+        super().__init__()
+        if p < 0 or p > 1:
+            raise ValueError(
+                "dropout probability has to be between 0 and 1, " "but got {}".format(p)
+            )
+        self.p = p
+
+    def forward(self, x: torch.Tensor):
+        # x: (batch_size, seq_len, dim)
+        if self.training and self.p > 0:
+            zero_mask = torch.full_like(x[0, :, 0], self.p).bernoulli().bool()
+            # TODO: permutation idx for each batch using torch.argsort
+            if zero_mask.any():
+                x = x[:, ~zero_mask, :]
+        return x
+
+
+class ZeroTokenDropout(nn.Module):
+    def __init__(self, p: float = 0.1):
+        super().__init__()
+        if p < 0 or p > 1:
+            raise ValueError(
+                "dropout probability has to be between 0 and 1, " "but got {}".format(p)
+            )
+        self.p = p
+
+    def forward(self, x: torch.Tensor):
+        # x: (batch_size, seq_len, dim)
+        if self.training and self.p > 0:
+            zero_mask = torch.full_like(x[:, :, 0], self.p).bernoulli().bool()
+            # Zero-out the masked tokens
+            x[zero_mask, :] = 0
+        return x
+
+
+class TransformerEncoder(nn.Module):
+    def __init__(
+        self,
+        num_tokens: int,
+        token_dim: int,
+        dim: int,
+        depth: int,
+        heads: int,
+        mlp_dim: int,
+        dim_head: int = 64,
+        dropout: float = 0.0,
+        emb_dropout: float = 0.0,
+        emb_dropout_type: str = "drop",
+        emb_dropout_loc: str = "token",
+        norm: str = "layer",
+        norm_cond_dim: int = -1,
+        token_pe_numfreq: int = -1,
+    ):
+        super().__init__()
+        if token_pe_numfreq > 0:
+            token_dim_new = token_dim * (2 * token_pe_numfreq + 1)
+            self.to_token_embedding = nn.Sequential(
+                Rearrange("b n d -> (b n) d", n=num_tokens, d=token_dim),
+                FrequencyEmbedder(token_pe_numfreq, token_pe_numfreq - 1),
+                Rearrange("(b n) d -> b n d", n=num_tokens, d=token_dim_new),
+                nn.Linear(token_dim_new, dim),
+            )
+        else:
+            self.to_token_embedding = nn.Linear(token_dim, dim)
+        self.pos_embedding = nn.Parameter(torch.randn(1, num_tokens, dim))
+        if emb_dropout_type == "drop":
+            self.dropout = DropTokenDropout(emb_dropout)
+        elif emb_dropout_type == "zero":
+            self.dropout = ZeroTokenDropout(emb_dropout)
+        else:
+            raise ValueError(f"Unknown emb_dropout_type: {emb_dropout_type}")
+        self.emb_dropout_loc = emb_dropout_loc
+
+        self.transformer = Transformer(
+            dim, depth, heads, dim_head, mlp_dim, dropout, norm=norm, norm_cond_dim=norm_cond_dim
+        )
+
+    def forward(self, inp: torch.Tensor, *args, **kwargs):
+        x = inp
+
+        if self.emb_dropout_loc == "input":
+            x = self.dropout(x)
+        x = self.to_token_embedding(x)
+
+        if self.emb_dropout_loc == "token":
+            x = self.dropout(x)
+        b, n, _ = x.shape
+        x += self.pos_embedding[:, :n]
+
+        if self.emb_dropout_loc == "token_afterpos":
+            x = self.dropout(x)
+        x = self.transformer(x, *args)
+        return x
+
+
+class TransformerDecoder(nn.Module):
+    def __init__(
+        self,
+        num_tokens: int,
+        token_dim: int,
+        dim: int,
+        depth: int,
+        heads: int,
+        mlp_dim: int,
+        dim_head: int = 64,
+        dropout: float = 0.0,
+        emb_dropout: float = 0.0,
+        emb_dropout_type: str = 'drop',
+        norm: str = "layer",
+        norm_cond_dim: int = -1,
+        context_dim: Optional[int] = None,
+        skip_token_embedding: bool = False,
+    ):
+        super().__init__()
+        if not skip_token_embedding:
+            self.to_token_embedding = nn.Linear(token_dim, dim)
+        else:
+            self.to_token_embedding = nn.Identity()
+            if token_dim != dim:
+                raise ValueError(
+                    f"token_dim ({token_dim}) != dim ({dim}) when skip_token_embedding is True"
+                )
+
+        self.pos_embedding = nn.Parameter(torch.randn(1, num_tokens, dim))
+        if emb_dropout_type == "drop":
+            self.dropout = DropTokenDropout(emb_dropout)
+        elif emb_dropout_type == "zero":
+            self.dropout = ZeroTokenDropout(emb_dropout)
+        elif emb_dropout_type == "normal":
+            self.dropout = nn.Dropout(emb_dropout)
+
+        self.transformer = TransformerCrossAttn(
+            dim,
+            depth,
+            heads,
+            dim_head,
+            mlp_dim,
+            dropout,
+            norm=norm,
+            norm_cond_dim=norm_cond_dim,
+            context_dim=context_dim,
+        )
+
+    def forward(self, inp: torch.Tensor, *args, context=None, context_list=None):
+        x = self.to_token_embedding(inp)
+        b, n, _ = x.shape
+
+        x = self.dropout(x)
+        x += self.pos_embedding[:, :n]
+
+        x = self.transformer(x, *args, context=context, context_list=context_list)
+        return x
+

humans4d/hmr2/models/components/t_cond_mlp.py

@@ -0,0 +1,199 @@
+import copy
+from typing import List, Optional
+
+import torch
+
+
+class AdaptiveLayerNorm1D(torch.nn.Module):
+    def __init__(self, data_dim: int, norm_cond_dim: int):
+        super().__init__()
+        if data_dim <= 0:
+            raise ValueError(f"data_dim must be positive, but got {data_dim}")
+        if norm_cond_dim <= 0:
+            raise ValueError(f"norm_cond_dim must be positive, but got {norm_cond_dim}")
+        self.norm = torch.nn.LayerNorm(
+            data_dim
+        )  # TODO: Check if elementwise_affine=True is correct
+        self.linear = torch.nn.Linear(norm_cond_dim, 2 * data_dim)
+        torch.nn.init.zeros_(self.linear.weight)
+        torch.nn.init.zeros_(self.linear.bias)
+
+    def forward(self, x: torch.Tensor, t: torch.Tensor) -> torch.Tensor:
+        # x: (batch, ..., data_dim)
+        # t: (batch, norm_cond_dim)
+        # return: (batch, data_dim)
+        x = self.norm(x)
+        alpha, beta = self.linear(t).chunk(2, dim=-1)
+
+        # Add singleton dimensions to alpha and beta
+        if x.dim() > 2:
+            alpha = alpha.view(alpha.shape[0], *([1] * (x.dim() - 2)), alpha.shape[1])
+            beta = beta.view(beta.shape[0], *([1] * (x.dim() - 2)), beta.shape[1])
+
+        return x * (1 + alpha) + beta
+
+
+class SequentialCond(torch.nn.Sequential):
+    def forward(self, input, *args, **kwargs):
+        for module in self:
+            if isinstance(module, (AdaptiveLayerNorm1D, SequentialCond, ResidualMLPBlock)):
+                # print(f'Passing on args to {module}', [a.shape for a in args])
+                input = module(input, *args, **kwargs)
+            else:
+                # print(f'Skipping passing args to {module}', [a.shape for a in args])
+                input = module(input)
+        return input
+
+
+def normalization_layer(norm: Optional[str], dim: int, norm_cond_dim: int = -1):
+    if norm == "batch":
+        return torch.nn.BatchNorm1d(dim)
+    elif norm == "layer":
+        return torch.nn.LayerNorm(dim)
+    elif norm == "ada":
+        assert norm_cond_dim > 0, f"norm_cond_dim must be positive, got {norm_cond_dim}"
+        return AdaptiveLayerNorm1D(dim, norm_cond_dim)
+    elif norm is None:
+        return torch.nn.Identity()
+    else:
+        raise ValueError(f"Unknown norm: {norm}")
+
+
+def linear_norm_activ_dropout(
+    input_dim: int,
+    output_dim: int,
+    activation: torch.nn.Module = torch.nn.ReLU(),
+    bias: bool = True,
+    norm: Optional[str] = "layer",  # Options: ada/batch/layer
+    dropout: float = 0.0,
+    norm_cond_dim: int = -1,
+) -> SequentialCond:
+    layers = []
+    layers.append(torch.nn.Linear(input_dim, output_dim, bias=bias))
+    if norm is not None:
+        layers.append(normalization_layer(norm, output_dim, norm_cond_dim))
+    layers.append(copy.deepcopy(activation))
+    if dropout > 0.0:
+        layers.append(torch.nn.Dropout(dropout))
+    return SequentialCond(*layers)
+
+
+def create_simple_mlp(
+    input_dim: int,
+    hidden_dims: List[int],
+    output_dim: int,
+    activation: torch.nn.Module = torch.nn.ReLU(),
+    bias: bool = True,
+    norm: Optional[str] = "layer",  # Options: ada/batch/layer
+    dropout: float = 0.0,
+    norm_cond_dim: int = -1,
+) -> SequentialCond:
+    layers = []
+    prev_dim = input_dim
+    for hidden_dim in hidden_dims:
+        layers.extend(
+            linear_norm_activ_dropout(
+                prev_dim, hidden_dim, activation, bias, norm, dropout, norm_cond_dim
+            )
+        )
+        prev_dim = hidden_dim
+    layers.append(torch.nn.Linear(prev_dim, output_dim, bias=bias))
+    return SequentialCond(*layers)
+
+
+class ResidualMLPBlock(torch.nn.Module):
+    def __init__(
+        self,
+        input_dim: int,
+        hidden_dim: int,
+        num_hidden_layers: int,
+        output_dim: int,
+        activation: torch.nn.Module = torch.nn.ReLU(),
+        bias: bool = True,
+        norm: Optional[str] = "layer",  # Options: ada/batch/layer
+        dropout: float = 0.0,
+        norm_cond_dim: int = -1,
+    ):
+        super().__init__()
+        if not (input_dim == output_dim == hidden_dim):
+            raise NotImplementedError(
+                f"input_dim {input_dim} != output_dim {output_dim} is not implemented"
+            )
+
+        layers = []
+        prev_dim = input_dim
+        for i in range(num_hidden_layers):
+            layers.append(
+                linear_norm_activ_dropout(
+                    prev_dim, hidden_dim, activation, bias, norm, dropout, norm_cond_dim
+                )
+            )
+            prev_dim = hidden_dim
+        self.model = SequentialCond(*layers)
+        self.skip = torch.nn.Identity()
+
+    def forward(self, x: torch.Tensor, *args, **kwargs) -> torch.Tensor:
+        return x + self.model(x, *args, **kwargs)
+
+
+class ResidualMLP(torch.nn.Module):
+    def __init__(
+        self,
+        input_dim: int,
+        hidden_dim: int,
+        num_hidden_layers: int,
+        output_dim: int,
+        activation: torch.nn.Module = torch.nn.ReLU(),
+        bias: bool = True,
+        norm: Optional[str] = "layer",  # Options: ada/batch/layer
+        dropout: float = 0.0,
+        num_blocks: int = 1,
+        norm_cond_dim: int = -1,
+    ):
+        super().__init__()
+        self.input_dim = input_dim
+        self.model = SequentialCond(
+            linear_norm_activ_dropout(
+                input_dim, hidden_dim, activation, bias, norm, dropout, norm_cond_dim
+            ),
+            *[
+                ResidualMLPBlock(
+                    hidden_dim,
+                    hidden_dim,
+                    num_hidden_layers,
+                    hidden_dim,
+                    activation,
+                    bias,
+                    norm,
+                    dropout,
+                    norm_cond_dim,
+                )
+                for _ in range(num_blocks)
+            ],
+            torch.nn.Linear(hidden_dim, output_dim, bias=bias),
+        )
+
+    def forward(self, x: torch.Tensor, *args, **kwargs) -> torch.Tensor:
+        return self.model(x, *args, **kwargs)
+
+
+class FrequencyEmbedder(torch.nn.Module):
+    def __init__(self, num_frequencies, max_freq_log2):
+        super().__init__()
+        frequencies = 2 ** torch.linspace(0, max_freq_log2, steps=num_frequencies)
+        self.register_buffer("frequencies", frequencies)
+
+    def forward(self, x):
+        # x should be of size (N,) or (N, D)
+        N = x.size(0)
+        if x.dim() == 1:  # (N,)
+            x = x.unsqueeze(1)  # (N, D) where D=1
+        x_unsqueezed = x.unsqueeze(-1)  # (N, D, 1)
+        scaled = self.frequencies.view(1, 1, -1) * x_unsqueezed  # (N, D, num_frequencies)
+        s = torch.sin(scaled)
+        c = torch.cos(scaled)
+        embedded = torch.cat([s, c, x_unsqueezed], dim=-1).view(
+            N, -1
+        )  # (N, D * 2 * num_frequencies + D)
+        return embedded
+

humans4d/hmr2/models/discriminator.py

@@ -0,0 +1,99 @@
+import torch
+import torch.nn as nn
+
+class Discriminator(nn.Module):
+
+    def __init__(self):
+        """
+        Pose + Shape discriminator proposed in HMR
+        """
+        super(Discriminator, self).__init__()
+
+        self.num_joints = 23
+        # poses_alone
+        self.D_conv1 = nn.Conv2d(9, 32, kernel_size=1)
+        nn.init.xavier_uniform_(self.D_conv1.weight)
+        nn.init.zeros_(self.D_conv1.bias)
+        self.relu = nn.ReLU(inplace=True)
+        self.D_conv2 = nn.Conv2d(32, 32, kernel_size=1)
+        nn.init.xavier_uniform_(self.D_conv2.weight)
+        nn.init.zeros_(self.D_conv2.bias)
+        pose_out = []
+        for i in range(self.num_joints):
+            pose_out_temp = nn.Linear(32, 1)
+            nn.init.xavier_uniform_(pose_out_temp.weight)
+            nn.init.zeros_(pose_out_temp.bias)
+            pose_out.append(pose_out_temp)
+        self.pose_out = nn.ModuleList(pose_out)
+
+        # betas
+        self.betas_fc1 = nn.Linear(10, 10)
+        nn.init.xavier_uniform_(self.betas_fc1.weight)
+        nn.init.zeros_(self.betas_fc1.bias)
+        self.betas_fc2 = nn.Linear(10, 5)
+        nn.init.xavier_uniform_(self.betas_fc2.weight)
+        nn.init.zeros_(self.betas_fc2.bias)
+        self.betas_out = nn.Linear(5, 1)
+        nn.init.xavier_uniform_(self.betas_out.weight)
+        nn.init.zeros_(self.betas_out.bias)
+
+        # poses_joint
+        self.D_alljoints_fc1 = nn.Linear(32*self.num_joints, 1024)
+        nn.init.xavier_uniform_(self.D_alljoints_fc1.weight)
+        nn.init.zeros_(self.D_alljoints_fc1.bias)
+        self.D_alljoints_fc2 = nn.Linear(1024, 1024)
+        nn.init.xavier_uniform_(self.D_alljoints_fc2.weight)
+        nn.init.zeros_(self.D_alljoints_fc2.bias)
+        self.D_alljoints_out = nn.Linear(1024, 1)
+        nn.init.xavier_uniform_(self.D_alljoints_out.weight)
+        nn.init.zeros_(self.D_alljoints_out.bias)
+
+
+    def forward(self, poses: torch.Tensor, betas: torch.Tensor) -> torch.Tensor:
+        """
+        Forward pass of the discriminator.
+        Args:
+            poses (torch.Tensor): Tensor of shape (B, 23, 3, 3) containing a batch of SMPL body poses (excluding the global orientation).
+            betas (torch.Tensor): Tensor of shape (B, 10) containign a batch of SMPL beta coefficients.
+        Returns:
+            torch.Tensor: Discriminator output with shape (B, 25)
+        """
+        #import ipdb; ipdb.set_trace()
+        #bn = poses.shape[0]
+        # poses B x 207
+        #poses = poses.reshape(bn, -1)
+        # poses B x num_joints x 1 x 9
+        poses = poses.reshape(-1, self.num_joints, 1, 9)
+        bn = poses.shape[0]
+        # poses B x 9 x num_joints x 1
+        poses = poses.permute(0, 3, 1, 2).contiguous()
+
+        # poses_alone
+        poses = self.D_conv1(poses)
+        poses = self.relu(poses)
+        poses = self.D_conv2(poses)
+        poses = self.relu(poses)
+
+        poses_out = []
+        for i in range(self.num_joints):
+            poses_out_ = self.pose_out[i](poses[:, :, i, 0])
+            poses_out.append(poses_out_)
+        poses_out = torch.cat(poses_out, dim=1)
+
+        # betas
+        betas = self.betas_fc1(betas)
+        betas = self.relu(betas)
+        betas = self.betas_fc2(betas)
+        betas = self.relu(betas)
+        betas_out = self.betas_out(betas)
+
+        # poses_joint
+        poses = poses.reshape(bn,-1)
+        poses_all = self.D_alljoints_fc1(poses)
+        poses_all = self.relu(poses_all)
+        poses_all = self.D_alljoints_fc2(poses_all)
+        poses_all = self.relu(poses_all)
+        poses_all_out = self.D_alljoints_out(poses_all)
+
+        disc_out = torch.cat((poses_out, betas_out, poses_all_out), 1)
+        return disc_out

humans4d/hmr2/models/heads/__init__.py

@@ -0,0 +1 @@
+from .smpl_head import build_smpl_head

humans4d/hmr2/models/heads/smpl_head.py

@@ -0,0 +1,111 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import numpy as np
+import einops
+
+from ...utils.geometry import rot6d_to_rotmat, aa_to_rotmat
+from ..components.pose_transformer import TransformerDecoder
+
+def build_smpl_head(cfg):
+    smpl_head_type = cfg.MODEL.SMPL_HEAD.get('TYPE', 'hmr')
+    if  smpl_head_type == 'transformer_decoder':
+        return SMPLTransformerDecoderHead(cfg)
+    else:
+        raise ValueError('Unknown SMPL head type: {}'.format(smpl_head_type))
+
+class SMPLTransformerDecoderHead(nn.Module):
+    """ Cross-attention based SMPL Transformer decoder
+    """
+
+    def __init__(self, cfg):
+        super().__init__()
+        self.cfg = cfg
+        self.joint_rep_type = cfg.MODEL.SMPL_HEAD.get('JOINT_REP', '6d')
+        self.joint_rep_dim = {'6d': 6, 'aa': 3}[self.joint_rep_type]
+        npose = self.joint_rep_dim * (cfg.SMPL.NUM_BODY_JOINTS + 1)
+        self.npose = npose
+        self.input_is_mean_shape = cfg.MODEL.SMPL_HEAD.get('TRANSFORMER_INPUT', 'zero') == 'mean_shape'
+        transformer_args = dict(
+            num_tokens=1,
+            token_dim=(npose + 10 + 3) if self.input_is_mean_shape else 1,
+            dim=1024,
+        )
+        transformer_args = (transformer_args | dict(cfg.MODEL.SMPL_HEAD.TRANSFORMER_DECODER))
+        self.transformer = TransformerDecoder(
+            **transformer_args
+        )
+        dim=transformer_args['dim']
+        self.decpose = nn.Linear(dim, npose)
+        self.decshape = nn.Linear(dim, 10)
+        self.deccam = nn.Linear(dim, 3)
+
+        if cfg.MODEL.SMPL_HEAD.get('INIT_DECODER_XAVIER', False):
+            # True by default in MLP. False by default in Transformer
+            nn.init.xavier_uniform_(self.decpose.weight, gain=0.01)
+            nn.init.xavier_uniform_(self.decshape.weight, gain=0.01)
+            nn.init.xavier_uniform_(self.deccam.weight, gain=0.01)
+
+        mean_params = np.load(cfg.SMPL.MEAN_PARAMS)
+        init_body_pose = torch.from_numpy(mean_params['pose'].astype(np.float32)).unsqueeze(0)
+        init_betas = torch.from_numpy(mean_params['shape'].astype('float32')).unsqueeze(0)
+        init_cam = torch.from_numpy(mean_params['cam'].astype(np.float32)).unsqueeze(0)
+        self.register_buffer('init_body_pose', init_body_pose)
+        self.register_buffer('init_betas', init_betas)
+        self.register_buffer('init_cam', init_cam)
+
+    def forward(self, x, **kwargs):
+
+        batch_size = x.shape[0]
+        # vit pretrained backbone is channel-first. Change to token-first
+        x = einops.rearrange(x, 'b c h w -> b (h w) c')
+
+        init_body_pose = self.init_body_pose.expand(batch_size, -1)
+        init_betas = self.init_betas.expand(batch_size, -1)
+        init_cam = self.init_cam.expand(batch_size, -1)
+
+        # TODO: Convert init_body_pose to aa rep if needed
+        if self.joint_rep_type == 'aa':
+            raise NotImplementedError
+
+        pred_body_pose = init_body_pose
+        pred_betas = init_betas
+        pred_cam = init_cam
+        pred_body_pose_list = []
+        pred_betas_list = []
+        pred_cam_list = []
+        for i in range(self.cfg.MODEL.SMPL_HEAD.get('IEF_ITERS', 1)):
+            # Input token to transformer is zero token
+            if self.input_is_mean_shape:
+                token = torch.cat([pred_body_pose, pred_betas, pred_cam], dim=1)[:,None,:]
+            else:
+                token = torch.zeros(batch_size, 1, 1).to(x.device)
+
+            # Pass through transformer
+            token_out = self.transformer(token, context=x)
+            token_out = token_out.squeeze(1) # (B, C)
+
+            # Readout from token_out
+            pred_body_pose = self.decpose(token_out) + pred_body_pose
+            pred_betas = self.decshape(token_out) + pred_betas
+            pred_cam = self.deccam(token_out) + pred_cam
+            pred_body_pose_list.append(pred_body_pose)
+            pred_betas_list.append(pred_betas)
+            pred_cam_list.append(pred_cam)
+
+        # Convert self.joint_rep_type -> rotmat
+        joint_conversion_fn = {
+            '6d': rot6d_to_rotmat,
+            'aa': lambda x: aa_to_rotmat(x.view(-1, 3).contiguous())
+        }[self.joint_rep_type]
+
+        pred_smpl_params_list = {}
+        pred_smpl_params_list['body_pose'] = torch.cat([joint_conversion_fn(pbp).view(batch_size, -1, 3, 3)[:, 1:, :, :] for pbp in pred_body_pose_list], dim=0)
+        pred_smpl_params_list['betas'] = torch.cat(pred_betas_list, dim=0)
+        pred_smpl_params_list['cam'] = torch.cat(pred_cam_list, dim=0)
+        pred_body_pose = joint_conversion_fn(pred_body_pose).view(batch_size, self.cfg.SMPL.NUM_BODY_JOINTS+1, 3, 3)
+
+        pred_smpl_params = {'global_orient': pred_body_pose[:, [0]],
+                            'body_pose': pred_body_pose[:, 1:],
+                            'betas': pred_betas}
+        return pred_smpl_params, pred_cam, pred_smpl_params_list

humans4d/hmr2/models/hmr2.py

@@ -0,0 +1,355 @@
+import torch
+import pytorch_lightning as pl
+from typing import Any, Dict, Mapping, Tuple
+
+from yacs.config import CfgNode
+
+from ..utils import SkeletonRenderer, MeshRenderer
+from ..utils.geometry import aa_to_rotmat, perspective_projection
+from ..utils.pylogger import get_pylogger
+from .backbones import create_backbone
+from .heads import build_smpl_head
+from .discriminator import Discriminator
+from .losses import Keypoint3DLoss, Keypoint2DLoss, ParameterLoss
+from . import SMPL
+
+log = get_pylogger(__name__)
+
+class HMR2(pl.LightningModule):
+
+    def __init__(self, cfg: CfgNode, init_renderer: bool = True):
+        """
+        Setup HMR2 model
+        Args:
+            cfg (CfgNode): Config file as a yacs CfgNode
+        """
+        super().__init__()
+
+        # Save hyperparameters
+        self.save_hyperparameters(logger=False, ignore=['init_renderer'])
+
+        self.cfg = cfg
+        # Create backbone feature extractor
+        self.backbone = create_backbone(cfg)
+        if cfg.MODEL.BACKBONE.get('PRETRAINED_WEIGHTS', None):
+            log.info(f'Loading backbone weights from {cfg.MODEL.BACKBONE.PRETRAINED_WEIGHTS}')
+            self.backbone.load_state_dict(torch.load(cfg.MODEL.BACKBONE.PRETRAINED_WEIGHTS, map_location='cpu')['state_dict'])
+
+        # Create SMPL head
+        self.smpl_head = build_smpl_head(cfg)
+
+        # Create discriminator
+        if self.cfg.LOSS_WEIGHTS.ADVERSARIAL > 0:
+            self.discriminator = Discriminator()
+
+        # Define loss functions
+        self.keypoint_3d_loss = Keypoint3DLoss(loss_type='l1')
+        self.keypoint_2d_loss = Keypoint2DLoss(loss_type='l1')
+        self.smpl_parameter_loss = ParameterLoss()
+
+        # Instantiate SMPL model
+        smpl_cfg = {k.lower(): v for k,v in dict(cfg.SMPL).items()}
+        self.smpl = SMPL(**smpl_cfg)
+
+        # Buffer that shows whetheer we need to initialize ActNorm layers
+        self.register_buffer('initialized', torch.tensor(False))
+        # Setup renderer for visualization
+        if init_renderer:
+            self.renderer = SkeletonRenderer(self.cfg)
+            self.mesh_renderer = MeshRenderer(self.cfg, faces=self.smpl.faces)
+        else:
+            self.renderer = None
+            self.mesh_renderer = None
+
+        # Disable automatic optimization since we use adversarial training
+        self.automatic_optimization = False
+
+    def get_parameters(self):
+        all_params = list(self.smpl_head.parameters())
+        all_params += list(self.backbone.parameters())
+        return all_params
+
+    def configure_optimizers(self) -> Tuple[torch.optim.Optimizer, torch.optim.Optimizer]:
+        """
+        Setup model and distriminator Optimizers
+        Returns:
+            Tuple[torch.optim.Optimizer, torch.optim.Optimizer]: Model and discriminator optimizers
+        """
+        param_groups = [{'params': filter(lambda p: p.requires_grad, self.get_parameters()), 'lr': self.cfg.TRAIN.LR}]
+
+        optimizer = torch.optim.AdamW(params=param_groups,
+                                        # lr=self.cfg.TRAIN.LR,
+                                        weight_decay=self.cfg.TRAIN.WEIGHT_DECAY)
+        optimizer_disc = torch.optim.AdamW(params=self.discriminator.parameters(),
+                                            lr=self.cfg.TRAIN.LR,
+                                            weight_decay=self.cfg.TRAIN.WEIGHT_DECAY)
+
+        return optimizer, optimizer_disc
+
+    def forward_step(self, batch: Dict, train: bool = False) -> Dict:
+        """
+        Run a forward step of the network
+        Args:
+            batch (Dict): Dictionary containing batch data
+            train (bool): Flag indicating whether it is training or validation mode
+        Returns:
+            Dict: Dictionary containing the regression output
+        """
+
+        # Use RGB image as input
+        x = batch['img']
+        batch_size = x.shape[0]
+
+        # Compute conditioning features using the backbone
+        # if using ViT backbone, we need to use a different aspect ratio
+        conditioning_feats = self.backbone(x[:,:,:,32:-32])
+
+        pred_smpl_params, pred_cam, _ = self.smpl_head(conditioning_feats)
+
+        # Store useful regression outputs to the output dict
+        output = {}
+        output['pred_cam'] = pred_cam
+        output['pred_smpl_params'] = {k: v.clone() for k,v in pred_smpl_params.items()}
+
+        # Compute camera translation
+        device = pred_smpl_params['body_pose'].device
+        dtype = pred_smpl_params['body_pose'].dtype
+        focal_length = self.cfg.EXTRA.FOCAL_LENGTH * torch.ones(batch_size, 2, device=device, dtype=dtype)
+        pred_cam_t = torch.stack([pred_cam[:, 1],
+                                  pred_cam[:, 2],
+                                  2*focal_length[:, 0]/(self.cfg.MODEL.IMAGE_SIZE * pred_cam[:, 0] +1e-9)],dim=-1)
+        output['pred_cam_t'] = pred_cam_t
+        output['focal_length'] = focal_length
+
+        # Compute model vertices, joints and the projected joints
+        pred_smpl_params['global_orient'] = pred_smpl_params['global_orient'].reshape(batch_size, -1, 3, 3)
+        pred_smpl_params['body_pose'] = pred_smpl_params['body_pose'].reshape(batch_size, -1, 3, 3)
+        pred_smpl_params['betas'] = pred_smpl_params['betas'].reshape(batch_size, -1)
+        smpl_output = self.smpl(**{k: v.float() for k,v in pred_smpl_params.items()}, pose2rot=False)
+        pred_keypoints_3d = smpl_output.joints
+        pred_vertices = smpl_output.vertices
+        output['pred_keypoints_3d'] = pred_keypoints_3d.reshape(batch_size, -1, 3)
+        output['pred_vertices'] = pred_vertices.reshape(batch_size, -1, 3)
+        pred_cam_t = pred_cam_t.reshape(-1, 3)
+        focal_length = focal_length.reshape(-1, 2)
+        pred_keypoints_2d = perspective_projection(pred_keypoints_3d,
+                                                   translation=pred_cam_t,
+                                                   focal_length=focal_length / self.cfg.MODEL.IMAGE_SIZE)
+
+        output['pred_keypoints_2d'] = pred_keypoints_2d.reshape(batch_size, -1, 2)
+        return output
+
+    def compute_loss(self, batch: Dict, output: Dict, train: bool = True) -> torch.Tensor:
+        """
+        Compute losses given the input batch and the regression output
+        Args:
+            batch (Dict): Dictionary containing batch data
+            output (Dict): Dictionary containing the regression output
+            train (bool): Flag indicating whether it is training or validation mode
+        Returns:
+            torch.Tensor : Total loss for current batch
+        """
+
+        pred_smpl_params = output['pred_smpl_params']
+        pred_keypoints_2d = output['pred_keypoints_2d']
+        pred_keypoints_3d = output['pred_keypoints_3d']
+
+
+        batch_size = pred_smpl_params['body_pose'].shape[0]
+        device = pred_smpl_params['body_pose'].device
+        dtype = pred_smpl_params['body_pose'].dtype
+
+        # Get annotations
+        gt_keypoints_2d = batch['keypoints_2d']
+        gt_keypoints_3d = batch['keypoints_3d']
+        gt_smpl_params = batch['smpl_params']
+        has_smpl_params = batch['has_smpl_params']
+        is_axis_angle = batch['smpl_params_is_axis_angle']
+
+        # Compute 3D keypoint loss
+        loss_keypoints_2d = self.keypoint_2d_loss(pred_keypoints_2d, gt_keypoints_2d)
+        loss_keypoints_3d = self.keypoint_3d_loss(pred_keypoints_3d, gt_keypoints_3d, pelvis_id=25+14)
+
+        # Compute loss on SMPL parameters
+        loss_smpl_params = {}
+        for k, pred in pred_smpl_params.items():
+            gt = gt_smpl_params[k].view(batch_size, -1)
+            if is_axis_angle[k].all():
+                gt = aa_to_rotmat(gt.reshape(-1, 3)).view(batch_size, -1, 3, 3)
+            has_gt = has_smpl_params[k]
+            loss_smpl_params[k] = self.smpl_parameter_loss(pred.reshape(batch_size, -1), gt.reshape(batch_size, -1), has_gt)
+
+        loss = self.cfg.LOSS_WEIGHTS['KEYPOINTS_3D'] * loss_keypoints_3d+\
+               self.cfg.LOSS_WEIGHTS['KEYPOINTS_2D'] * loss_keypoints_2d+\
+               sum([loss_smpl_params[k] * self.cfg.LOSS_WEIGHTS[k.upper()] for k in loss_smpl_params])
+
+        losses = dict(loss=loss.detach(),
+                      loss_keypoints_2d=loss_keypoints_2d.detach(),
+                      loss_keypoints_3d=loss_keypoints_3d.detach())
+
+        for k, v in loss_smpl_params.items():
+            losses['loss_' + k] = v.detach()
+
+        output['losses'] = losses
+
+        return loss
+
+    # Tensoroboard logging should run from first rank only
+    @pl.utilities.rank_zero.rank_zero_only
+    def tensorboard_logging(self, batch: Dict, output: Dict, step_count: int, train: bool = True, write_to_summary_writer: bool = True) -> None:
+        """
+        Log results to Tensorboard
+        Args:
+            batch (Dict): Dictionary containing batch data
+            output (Dict): Dictionary containing the regression output
+            step_count (int): Global training step count
+            train (bool): Flag indicating whether it is training or validation mode
+        """
+
+        mode = 'train' if train else 'val'
+        batch_size = batch['keypoints_2d'].shape[0]
+        images = batch['img']
+        images = images * torch.tensor([0.229, 0.224, 0.225], device=images.device).reshape(1,3,1,1)
+        images = images + torch.tensor([0.485, 0.456, 0.406], device=images.device).reshape(1,3,1,1)
+        #images = 255*images.permute(0, 2, 3, 1).cpu().numpy()
+
+        pred_keypoints_3d = output['pred_keypoints_3d'].detach().reshape(batch_size, -1, 3)
+        pred_vertices = output['pred_vertices'].detach().reshape(batch_size, -1, 3)
+        focal_length = output['focal_length'].detach().reshape(batch_size, 2)
+        gt_keypoints_3d = batch['keypoints_3d']
+        gt_keypoints_2d = batch['keypoints_2d']
+        losses = output['losses']
+        pred_cam_t = output['pred_cam_t'].detach().reshape(batch_size, 3)
+        pred_keypoints_2d = output['pred_keypoints_2d'].detach().reshape(batch_size, -1, 2)
+
+        if write_to_summary_writer:
+            summary_writer = self.logger.experiment
+            for loss_name, val in losses.items():
+                summary_writer.add_scalar(mode +'/' + loss_name, val.detach().item(), step_count)
+        num_images = min(batch_size, self.cfg.EXTRA.NUM_LOG_IMAGES)
+
+        gt_keypoints_3d = batch['keypoints_3d']
+        pred_keypoints_3d = output['pred_keypoints_3d'].detach().reshape(batch_size, -1, 3)
+
+        # We render the skeletons instead of the full mesh because rendering a lot of meshes will make the training slow.
+        #predictions = self.renderer(pred_keypoints_3d[:num_images],
+        #                            gt_keypoints_3d[:num_images],
+        #                            2 * gt_keypoints_2d[:num_images],
+        #                            images=images[:num_images],
+        #                            camera_translation=pred_cam_t[:num_images])
+        predictions = self.mesh_renderer.visualize_tensorboard(pred_vertices[:num_images].cpu().numpy(),
+                                                               pred_cam_t[:num_images].cpu().numpy(),
+                                                               images[:num_images].cpu().numpy(),
+                                                               pred_keypoints_2d[:num_images].cpu().numpy(),
+                                                               gt_keypoints_2d[:num_images].cpu().numpy(),
+                                                               focal_length=focal_length[:num_images].cpu().numpy())
+        if write_to_summary_writer:
+            summary_writer.add_image('%s/predictions' % mode, predictions, step_count)
+
+        return predictions
+
+    def forward(self, batch: Dict) -> Dict:
+        """
+        Run a forward step of the network in val mode
+        Args:
+            batch (Dict): Dictionary containing batch data
+        Returns:
+            Dict: Dictionary containing the regression output
+        """
+        return self.forward_step(batch, train=False)
+
+    def training_step_discriminator(self, batch: Dict,
+                                    body_pose: torch.Tensor,
+                                    betas: torch.Tensor,
+                                    optimizer: torch.optim.Optimizer) -> torch.Tensor:
+        """
+        Run a discriminator training step
+        Args:
+            batch (Dict): Dictionary containing mocap batch data
+            body_pose (torch.Tensor): Regressed body pose from current step
+            betas (torch.Tensor): Regressed betas from current step
+            optimizer (torch.optim.Optimizer): Discriminator optimizer
+        Returns:
+            torch.Tensor: Discriminator loss
+        """
+        batch_size = body_pose.shape[0]
+        gt_body_pose = batch['body_pose']
+        gt_betas = batch['betas']
+        gt_rotmat = aa_to_rotmat(gt_body_pose.view(-1,3)).view(batch_size, -1, 3, 3)
+        disc_fake_out = self.discriminator(body_pose.detach(), betas.detach())
+        loss_fake = ((disc_fake_out - 0.0) ** 2).sum() / batch_size
+        disc_real_out = self.discriminator(gt_rotmat, gt_betas)
+        loss_real = ((disc_real_out - 1.0) ** 2).sum() / batch_size
+        loss_disc = loss_fake + loss_real
+        loss = self.cfg.LOSS_WEIGHTS.ADVERSARIAL * loss_disc
+        optimizer.zero_grad()
+        self.manual_backward(loss)
+        optimizer.step()
+        return loss_disc.detach()
+
+    def training_step(self, joint_batch: Dict, batch_idx: int) -> Dict:
+        """
+        Run a full training step
+        Args:
+            joint_batch (Dict): Dictionary containing image and mocap batch data
+            batch_idx (int): Unused.
+            batch_idx (torch.Tensor): Unused.
+        Returns:
+            Dict: Dictionary containing regression output.
+        """
+        batch = joint_batch['img']
+        mocap_batch = joint_batch['mocap']
+        optimizer = self.optimizers(use_pl_optimizer=True)
+        if self.cfg.LOSS_WEIGHTS.ADVERSARIAL > 0:
+            optimizer, optimizer_disc = optimizer
+
+        batch_size = batch['img'].shape[0]
+        output = self.forward_step(batch, train=True)
+        pred_smpl_params = output['pred_smpl_params']
+        if self.cfg.get('UPDATE_GT_SPIN', False):
+            self.update_batch_gt_spin(batch, output)
+        loss = self.compute_loss(batch, output, train=True)
+        if self.cfg.LOSS_WEIGHTS.ADVERSARIAL > 0:
+            disc_out = self.discriminator(pred_smpl_params['body_pose'].reshape(batch_size, -1), pred_smpl_params['betas'].reshape(batch_size, -1))
+            loss_adv = ((disc_out - 1.0) ** 2).sum() / batch_size
+            loss = loss + self.cfg.LOSS_WEIGHTS.ADVERSARIAL * loss_adv
+
+        # Error if Nan
+        if torch.isnan(loss):
+            raise ValueError('Loss is NaN')
+
+        optimizer.zero_grad()
+        self.manual_backward(loss)
+        # Clip gradient
+        if self.cfg.TRAIN.get('GRAD_CLIP_VAL', 0) > 0:
+            gn = torch.nn.utils.clip_grad_norm_(self.get_parameters(), self.cfg.TRAIN.GRAD_CLIP_VAL, error_if_nonfinite=True)
+            self.log('train/grad_norm', gn, on_step=True, on_epoch=True, prog_bar=True, logger=True)
+        optimizer.step()
+        if self.cfg.LOSS_WEIGHTS.ADVERSARIAL > 0:
+            loss_disc = self.training_step_discriminator(mocap_batch, pred_smpl_params['body_pose'].reshape(batch_size, -1), pred_smpl_params['betas'].reshape(batch_size, -1), optimizer_disc)
+            output['losses']['loss_gen'] = loss_adv
+            output['losses']['loss_disc'] = loss_disc
+
+        if self.global_step > 0 and self.global_step % self.cfg.GENERAL.LOG_STEPS == 0:
+            self.tensorboard_logging(batch, output, self.global_step, train=True)
+
+        self.log('train/loss', output['losses']['loss'], on_step=True, on_epoch=True, prog_bar=True, logger=False)
+
+        return output
+
+    def validation_step(self, batch: Dict, batch_idx: int, dataloader_idx=0) -> Dict:
+        """
+        Run a validation step and log to Tensorboard
+        Args:
+            batch (Dict): Dictionary containing batch data
+            batch_idx (int): Unused.
+        Returns:
+            Dict: Dictionary containing regression output.
+        """
+        # batch_size = batch['img'].shape[0]
+        output = self.forward_step(batch, train=False)
+        loss = self.compute_loss(batch, output, train=False)
+        output['loss'] = loss
+        self.tensorboard_logging(batch, output, self.global_step, train=False)
+
+        return output

humans4d/hmr2/models/losses.py

@@ -0,0 +1,92 @@
+import torch
+import torch.nn as nn
+
+class Keypoint2DLoss(nn.Module):
+
+    def __init__(self, loss_type: str = 'l1'):
+        """
+        2D keypoint loss module.
+        Args:
+            loss_type (str): Choose between l1 and l2 losses.
+        """
+        super(Keypoint2DLoss, self).__init__()
+        if loss_type == 'l1':
+            self.loss_fn = nn.L1Loss(reduction='none')
+        elif loss_type == 'l2':
+            self.loss_fn = nn.MSELoss(reduction='none')
+        else:
+            raise NotImplementedError('Unsupported loss function')
+
+    def forward(self, pred_keypoints_2d: torch.Tensor, gt_keypoints_2d: torch.Tensor) -> torch.Tensor:
+        """
+        Compute 2D reprojection loss on the keypoints.
+        Args:
+            pred_keypoints_2d (torch.Tensor): Tensor of shape [B, S, N, 2] containing projected 2D keypoints (B: batch_size, S: num_samples, N: num_keypoints)
+            gt_keypoints_2d (torch.Tensor): Tensor of shape [B, S, N, 3] containing the ground truth 2D keypoints and confidence.
+        Returns:
+            torch.Tensor: 2D keypoint loss.
+        """
+        conf = gt_keypoints_2d[:, :, -1].unsqueeze(-1).clone()
+        batch_size = conf.shape[0]
+        loss = (conf * self.loss_fn(pred_keypoints_2d, gt_keypoints_2d[:, :, :-1])).sum(dim=(1,2))
+        return loss.sum()
+
+
+class Keypoint3DLoss(nn.Module):
+
+    def __init__(self, loss_type: str = 'l1'):
+        """
+        3D keypoint loss module.
+        Args:
+            loss_type (str): Choose between l1 and l2 losses.
+        """
+        super(Keypoint3DLoss, self).__init__()
+        if loss_type == 'l1':
+            self.loss_fn = nn.L1Loss(reduction='none')
+        elif loss_type == 'l2':
+            self.loss_fn = nn.MSELoss(reduction='none')
+        else:
+            raise NotImplementedError('Unsupported loss function')
+
+    def forward(self, pred_keypoints_3d: torch.Tensor, gt_keypoints_3d: torch.Tensor, pelvis_id: int = 39):
+        """
+        Compute 3D keypoint loss.
+        Args:
+            pred_keypoints_3d (torch.Tensor): Tensor of shape [B, S, N, 3] containing the predicted 3D keypoints (B: batch_size, S: num_samples, N: num_keypoints)
+            gt_keypoints_3d (torch.Tensor): Tensor of shape [B, S, N, 4] containing the ground truth 3D keypoints and confidence.
+        Returns:
+            torch.Tensor: 3D keypoint loss.
+        """
+        batch_size = pred_keypoints_3d.shape[0]
+        gt_keypoints_3d = gt_keypoints_3d.clone()
+        pred_keypoints_3d = pred_keypoints_3d - pred_keypoints_3d[:, pelvis_id, :].unsqueeze(dim=1)
+        gt_keypoints_3d[:, :, :-1] = gt_keypoints_3d[:, :, :-1] - gt_keypoints_3d[:, pelvis_id, :-1].unsqueeze(dim=1)
+        conf = gt_keypoints_3d[:, :, -1].unsqueeze(-1).clone()
+        gt_keypoints_3d = gt_keypoints_3d[:, :, :-1]
+        loss = (conf * self.loss_fn(pred_keypoints_3d, gt_keypoints_3d)).sum(dim=(1,2))
+        return loss.sum()
+
+class ParameterLoss(nn.Module):
+
+    def __init__(self):
+        """
+        SMPL parameter loss module.
+        """
+        super(ParameterLoss, self).__init__()
+        self.loss_fn = nn.MSELoss(reduction='none')
+
+    def forward(self, pred_param: torch.Tensor, gt_param: torch.Tensor, has_param: torch.Tensor):
+        """
+        Compute SMPL parameter loss.
+        Args:
+            pred_param (torch.Tensor): Tensor of shape [B, S, ...] containing the predicted parameters (body pose / global orientation / betas)
+            gt_param (torch.Tensor): Tensor of shape [B, S, ...] containing the ground truth SMPL parameters.
+        Returns:
+            torch.Tensor: L2 parameter loss loss.
+        """
+        batch_size = pred_param.shape[0]
+        num_dims = len(pred_param.shape)
+        mask_dimension = [batch_size] + [1] * (num_dims-1)
+        has_param = has_param.type(pred_param.type()).view(*mask_dimension)
+        loss_param = (has_param * self.loss_fn(pred_param, gt_param))
+        return loss_param.sum()

humans4d/hmr2/models/smpl_wrapper.py

@@ -0,0 +1,41 @@
+import torch
+import numpy as np
+import pickle
+from typing import Optional
+import smplx
+from smplx.lbs import vertices2joints
+from smplx.utils import SMPLOutput
+
+
+class SMPL(smplx.SMPLLayer):
+    def __init__(self, *args, joint_regressor_extra: Optional[str] = None, update_hips: bool = False, **kwargs):
+        """
+        Extension of the official SMPL implementation to support more joints.
+        Args:
+            Same as SMPLLayer.
+            joint_regressor_extra (str): Path to extra joint regressor.
+        """
+        super(SMPL, self).__init__(*args, **kwargs)
+        smpl_to_openpose = [24, 12, 17, 19, 21, 16, 18, 20, 0, 2, 5, 8, 1, 4,
+                            7, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34]
+            
+        if joint_regressor_extra is not None:
+            self.register_buffer('joint_regressor_extra', torch.tensor(pickle.load(open(joint_regressor_extra, 'rb'), encoding='latin1'), dtype=torch.float32))
+        self.register_buffer('joint_map', torch.tensor(smpl_to_openpose, dtype=torch.long))
+        self.update_hips = update_hips
+
+    def forward(self, *args, **kwargs) -> SMPLOutput:
+        """
+        Run forward pass. Same as SMPL and also append an extra set of joints if joint_regressor_extra is specified.
+        """
+        smpl_output = super(SMPL, self).forward(*args, **kwargs)
+        joints = smpl_output.joints[:, self.joint_map, :]
+        if self.update_hips:
+            joints[:,[9,12]] = joints[:,[9,12]] + \
+                0.25*(joints[:,[9,12]]-joints[:,[12,9]]) + \
+                0.5*(joints[:,[8]] - 0.5*(joints[:,[9,12]] + joints[:,[12,9]]))
+        if hasattr(self, 'joint_regressor_extra'):
+            extra_joints = vertices2joints(self.joint_regressor_extra, smpl_output.vertices)
+            joints = torch.cat([joints, extra_joints], dim=1)
+        smpl_output.joints = joints
+        return smpl_output

humans4d/hmr2/utils/__init__.py

@@ -0,0 +1,25 @@
+import torch
+from typing import Any
+
+from .renderer import Renderer
+from .mesh_renderer import MeshRenderer
+from .skeleton_renderer import SkeletonRenderer
+from .pose_utils import eval_pose, Evaluator
+
+def recursive_to(x: Any, target: torch.device):
+    """
+    Recursively transfer a batch of data to the target device
+    Args:
+        x (Any): Batch of data.
+        target (torch.device): Target device.
+    Returns:
+        Batch of data where all tensors are transfered to the target device.
+    """
+    if isinstance(x, dict):
+        return {k: recursive_to(v, target) for k, v in x.items()}
+    elif isinstance(x, torch.Tensor):
+        return x.to(target)
+    elif isinstance(x, list):
+        return [recursive_to(i, target) for i in x]
+    else:
+        return x

humans4d/hmr2/utils/download.py

@@ -0,0 +1,66 @@
+import os
+import re
+import sys
+from urllib import request as urlrequest
+
+
+def _progress_bar(count, total):
+    """Report download progress. Credit:
+    https://stackoverflow.com/questions/3173320/text-progress-bar-in-the-console/27871113
+    """
+    bar_len = 60
+    filled_len = int(round(bar_len * count / float(total)))
+    percents = round(100.0 * count / float(total), 1)
+    bar = "=" * filled_len + "-" * (bar_len - filled_len)
+    sys.stdout.write(
+        "  [{}] {}% of {:.1f}MB file  \r".format(bar, percents, total / 1024 / 1024)
+    )
+    sys.stdout.flush()
+    if count >= total:
+        sys.stdout.write("\n")
+
+
+def download_url(url, dst_file_path, chunk_size=8192, progress_hook=_progress_bar):
+    """Download url and write it to dst_file_path. Credit:
+    https://stackoverflow.com/questions/2028517/python-urllib2-progress-hook
+    """
+    # url = url + "?dl=1" if "dropbox" in url else url
+    req = urlrequest.Request(url)
+    response = urlrequest.urlopen(req)
+    total_size = response.info().get("Content-Length")
+    if total_size is None:
+        raise ValueError("Cannot determine size of download from {}".format(url))
+    total_size = int(total_size.strip())
+    bytes_so_far = 0
+
+    with open(dst_file_path, "wb") as f:
+        while 1:
+            chunk = response.read(chunk_size)
+            bytes_so_far += len(chunk)
+            if not chunk:
+                break
+
+            if progress_hook:
+                progress_hook(bytes_so_far, total_size)
+
+            f.write(chunk)
+    return bytes_so_far
+
+
+def cache_url(url_or_file, cache_file_path, download=True):
+    """Download the file specified by the URL to the cache_dir and return the path to
+    the cached file. If the argument is not a URL, simply return it as is.
+    """
+    is_url = re.match(r"^(?:http)s?://", url_or_file, re.IGNORECASE) is not None
+    if not is_url:
+        return url_or_file
+    url = url_or_file
+    if os.path.exists(cache_file_path):
+        return cache_file_path
+    cache_file_dir = os.path.dirname(cache_file_path)
+    if not os.path.exists(cache_file_dir):
+        os.makedirs(cache_file_dir)
+    if download:
+        print("Downloading remote file {} to {}".format(url, cache_file_path))
+        download_url(url, cache_file_path)
+    return cache_file_path

humans4d/hmr2/utils/geometry.py

@@ -0,0 +1,102 @@
+from typing import Optional
+import torch
+from torch.nn import functional as F
+
+def aa_to_rotmat(theta: torch.Tensor):
+    """
+    Convert axis-angle representation to rotation matrix.
+    Works by first converting it to a quaternion.
+    Args:
+        theta (torch.Tensor): Tensor of shape (B, 3) containing axis-angle representations.
+    Returns:
+        torch.Tensor: Corresponding rotation matrices with shape (B, 3, 3).
+    """
+    norm = torch.norm(theta + 1e-8, p = 2, dim = 1)
+    angle = torch.unsqueeze(norm, -1)
+    normalized = torch.div(theta, angle)
+    angle = angle * 0.5
+    v_cos = torch.cos(angle)
+    v_sin = torch.sin(angle)
+    quat = torch.cat([v_cos, v_sin * normalized], dim = 1)
+    return quat_to_rotmat(quat)
+
+def quat_to_rotmat(quat: torch.Tensor) -> torch.Tensor:
+    """
+    Convert quaternion representation to rotation matrix.
+    Args:
+        quat (torch.Tensor) of shape (B, 4); 4 <===> (w, x, y, z).
+    Returns:
+        torch.Tensor: Corresponding rotation matrices with shape (B, 3, 3).
+    """
+    norm_quat = quat
+    norm_quat = norm_quat/norm_quat.norm(p=2, dim=1, keepdim=True)
+    w, x, y, z = norm_quat[:,0], norm_quat[:,1], norm_quat[:,2], norm_quat[:,3]
+
+    B = quat.size(0)
+
+    w2, x2, y2, z2 = w.pow(2), x.pow(2), y.pow(2), z.pow(2)
+    wx, wy, wz = w*x, w*y, w*z
+    xy, xz, yz = x*y, x*z, y*z
+
+    rotMat = torch.stack([w2 + x2 - y2 - z2, 2*xy - 2*wz, 2*wy + 2*xz,
+                          2*wz + 2*xy, w2 - x2 + y2 - z2, 2*yz - 2*wx,
+                          2*xz - 2*wy, 2*wx + 2*yz, w2 - x2 - y2 + z2], dim=1).view(B, 3, 3)
+    return rotMat
+
+
+def rot6d_to_rotmat(x: torch.Tensor) -> torch.Tensor:
+    """
+    Convert 6D rotation representation to 3x3 rotation matrix.
+    Based on Zhou et al., "On the Continuity of Rotation Representations in Neural Networks", CVPR 2019
+    Args:
+        x (torch.Tensor): (B,6) Batch of 6-D rotation representations.
+    Returns:
+        torch.Tensor: Batch of corresponding rotation matrices with shape (B,3,3).
+    """
+    x = x.reshape(-1,2,3).permute(0, 2, 1).contiguous()
+    a1 = x[:, :, 0]
+    a2 = x[:, :, 1]
+    b1 = F.normalize(a1)
+    b2 = F.normalize(a2 - torch.einsum('bi,bi->b', b1, a2).unsqueeze(-1) * b1)
+    b3 = torch.cross(b1, b2)
+    return torch.stack((b1, b2, b3), dim=-1)
+
+def perspective_projection(points: torch.Tensor,
+                           translation: torch.Tensor,
+                           focal_length: torch.Tensor,
+                           camera_center: Optional[torch.Tensor] = None,
+                           rotation: Optional[torch.Tensor] = None) -> torch.Tensor:
+    """
+    Computes the perspective projection of a set of 3D points.
+    Args:
+        points (torch.Tensor): Tensor of shape (B, N, 3) containing the input 3D points.
+        translation (torch.Tensor): Tensor of shape (B, 3) containing the 3D camera translation.
+        focal_length (torch.Tensor): Tensor of shape (B, 2) containing the focal length in pixels.
+        camera_center (torch.Tensor): Tensor of shape (B, 2) containing the camera center in pixels.
+        rotation (torch.Tensor): Tensor of shape (B, 3, 3) containing the camera rotation.
+    Returns:
+        torch.Tensor: Tensor of shape (B, N, 2) containing the projection of the input points.
+    """
+    batch_size = points.shape[0]
+    if rotation is None:
+        rotation = torch.eye(3, device=points.device, dtype=points.dtype).unsqueeze(0).expand(batch_size, -1, -1)
+    if camera_center is None:
+        camera_center = torch.zeros(batch_size, 2, device=points.device, dtype=points.dtype)
+    # Populate intrinsic camera matrix K.
+    K = torch.zeros([batch_size, 3, 3], device=points.device, dtype=points.dtype)
+    K[:,0,0] = focal_length[:,0]
+    K[:,1,1] = focal_length[:,1]
+    K[:,2,2] = 1.
+    K[:,:-1, -1] = camera_center
+
+    # Transform points
+    points = torch.einsum('bij,bkj->bki', rotation, points)
+    points = points + translation.unsqueeze(1)
+
+    # Apply perspective distortion
+    projected_points = points / points[:,:,-1].unsqueeze(-1)
+
+    # Apply camera intrinsics
+    projected_points = torch.einsum('bij,bkj->bki', K, projected_points)
+
+    return projected_points[:, :, :-1]

humans4d/hmr2/utils/mesh_renderer.py

@@ -0,0 +1,149 @@
+import os
+if 'PYOPENGL_PLATFORM' not in os.environ:
+    os.environ['PYOPENGL_PLATFORM'] = 'egl'
+import torch
+from torchvision.utils import make_grid
+import numpy as np
+import pyrender
+import trimesh
+import cv2
+import torch.nn.functional as F
+
+from .render_openpose import render_openpose
+
+def create_raymond_lights():
+    import pyrender
+    thetas = np.pi * np.array([1.0 / 6.0, 1.0 / 6.0, 1.0 / 6.0])
+    phis = np.pi * np.array([0.0, 2.0 / 3.0, 4.0 / 3.0])
+
+    nodes = []
+
+    for phi, theta in zip(phis, thetas):
+        xp = np.sin(theta) * np.cos(phi)
+        yp = np.sin(theta) * np.sin(phi)
+        zp = np.cos(theta)
+
+        z = np.array([xp, yp, zp])
+        z = z / np.linalg.norm(z)
+        x = np.array([-z[1], z[0], 0.0])
+        if np.linalg.norm(x) == 0:
+            x = np.array([1.0, 0.0, 0.0])
+        x = x / np.linalg.norm(x)
+        y = np.cross(z, x)
+
+        matrix = np.eye(4)
+        matrix[:3,:3] = np.c_[x,y,z]
+        nodes.append(pyrender.Node(
+            light=pyrender.DirectionalLight(color=np.ones(3), intensity=1.0),
+            matrix=matrix
+        ))
+
+    return nodes
+
+class MeshRenderer:
+
+    def __init__(self, cfg, faces=None):
+        self.cfg = cfg
+        self.focal_length = cfg.EXTRA.FOCAL_LENGTH
+        self.img_res = cfg.MODEL.IMAGE_SIZE
+        self.renderer = pyrender.OffscreenRenderer(viewport_width=self.img_res,
+                                       viewport_height=self.img_res,
+                                       point_size=1.0)
+        
+        self.camera_center = [self.img_res // 2, self.img_res // 2]
+        self.faces = faces
+
+    def visualize(self, vertices, camera_translation, images, focal_length=None, nrow=3, padding=2):
+        images_np = np.transpose(images, (0,2,3,1))
+        rend_imgs = []
+        for i in range(vertices.shape[0]):
+            fl = self.focal_length
+            rend_img = torch.from_numpy(np.transpose(self.__call__(vertices[i], camera_translation[i], images_np[i], focal_length=fl, side_view=False), (2,0,1))).float()
+            rend_img_side = torch.from_numpy(np.transpose(self.__call__(vertices[i], camera_translation[i], images_np[i], focal_length=fl, side_view=True), (2,0,1))).float()
+            rend_imgs.append(torch.from_numpy(images[i]))
+            rend_imgs.append(rend_img)
+            rend_imgs.append(rend_img_side)
+        rend_imgs = make_grid(rend_imgs, nrow=nrow, padding=padding)
+        return rend_imgs
+
+    def visualize_tensorboard(self, vertices, camera_translation, images, pred_keypoints, gt_keypoints, focal_length=None, nrow=5, padding=2):
+        images_np = np.transpose(images, (0,2,3,1))
+        rend_imgs = []
+        pred_keypoints = np.concatenate((pred_keypoints, np.ones_like(pred_keypoints)[:, :, [0]]), axis=-1)
+        pred_keypoints = self.img_res * (pred_keypoints + 0.5)
+        gt_keypoints[:, :, :-1] = self.img_res * (gt_keypoints[:, :, :-1] + 0.5)
+        keypoint_matches = [(1, 12), (2, 8), (3, 7), (4, 6), (5, 9), (6, 10), (7, 11), (8, 14), (9, 2), (10, 1), (11, 0), (12, 3), (13, 4), (14, 5)]
+        for i in range(vertices.shape[0]):
+            fl = self.focal_length
+            rend_img = torch.from_numpy(np.transpose(self.__call__(vertices[i], camera_translation[i], images_np[i], focal_length=fl, side_view=False), (2,0,1))).float()
+            rend_img_side = torch.from_numpy(np.transpose(self.__call__(vertices[i], camera_translation[i], images_np[i], focal_length=fl, side_view=True), (2,0,1))).float()
+            body_keypoints = pred_keypoints[i, :25]
+            extra_keypoints = pred_keypoints[i, -19:]
+            for pair in keypoint_matches:
+                body_keypoints[pair[0], :] = extra_keypoints[pair[1], :]
+            pred_keypoints_img = render_openpose(255 * images_np[i].copy(), body_keypoints) / 255
+            body_keypoints = gt_keypoints[i, :25]
+            extra_keypoints = gt_keypoints[i, -19:]
+            for pair in keypoint_matches:
+                if extra_keypoints[pair[1], -1] > 0 and body_keypoints[pair[0], -1] == 0:
+                    body_keypoints[pair[0], :] = extra_keypoints[pair[1], :]
+            gt_keypoints_img = render_openpose(255*images_np[i].copy(), body_keypoints) / 255
+            rend_imgs.append(torch.from_numpy(images[i]))
+            rend_imgs.append(rend_img)
+            rend_imgs.append(rend_img_side)
+            rend_imgs.append(torch.from_numpy(pred_keypoints_img).permute(2,0,1))
+            rend_imgs.append(torch.from_numpy(gt_keypoints_img).permute(2,0,1))
+        rend_imgs = make_grid(rend_imgs, nrow=nrow, padding=padding)
+        return rend_imgs
+
+    def __call__(self, vertices, camera_translation, image, focal_length=5000, text=None, resize=None, side_view=False, baseColorFactor=(1.0, 1.0, 0.9, 1.0), rot_angle=90):
+        renderer = pyrender.OffscreenRenderer(viewport_width=image.shape[1],
+                                              viewport_height=image.shape[0],
+                                              point_size=1.0)
+        material = pyrender.MetallicRoughnessMaterial(
+            metallicFactor=0.0,
+            alphaMode='OPAQUE',
+            baseColorFactor=baseColorFactor)
+
+        camera_translation[0] *= -1.
+
+        mesh = trimesh.Trimesh(vertices.copy(), self.faces.copy())
+        if side_view:
+            rot = trimesh.transformations.rotation_matrix(
+                np.radians(rot_angle), [0, 1, 0])
+            mesh.apply_transform(rot)
+        rot = trimesh.transformations.rotation_matrix(
+            np.radians(180), [1, 0, 0])
+        mesh.apply_transform(rot)
+        mesh = pyrender.Mesh.from_trimesh(mesh, material=material)
+
+        scene = pyrender.Scene(bg_color=[0.0, 0.0, 0.0, 0.0],
+                               ambient_light=(0.3, 0.3, 0.3))
+        scene.add(mesh, 'mesh')
+
+        camera_pose = np.eye(4)
+        camera_pose[:3, 3] = camera_translation
+        camera_center = [image.shape[1] / 2., image.shape[0] / 2.]
+        camera = pyrender.IntrinsicsCamera(fx=focal_length, fy=focal_length,
+                                           cx=camera_center[0], cy=camera_center[1])
+        scene.add(camera, pose=camera_pose)
+
+
+        light_nodes = create_raymond_lights()
+        for node in light_nodes:
+            scene.add_node(node)
+
+        color, rend_depth = renderer.render(scene, flags=pyrender.RenderFlags.RGBA)
+        color = color.astype(np.float32) / 255.0
+        valid_mask = (color[:, :, -1] > 0)[:, :, np.newaxis]
+        if not side_view:
+            output_img = (color[:, :, :3] * valid_mask +
+                      (1 - valid_mask) * image)
+        else:
+            output_img = color[:, :, :3]
+        if resize is not None:
+            output_img = cv2.resize(output_img, resize)
+
+        output_img = output_img.astype(np.float32)
+        renderer.delete()
+        return output_img

humans4d/hmr2/utils/misc.py

@@ -0,0 +1,203 @@
+import time
+import warnings
+from importlib.util import find_spec
+from pathlib import Path
+from typing import Callable, List
+
+import hydra
+from omegaconf import DictConfig, OmegaConf
+from pytorch_lightning import Callback
+from pytorch_lightning.loggers import Logger
+from pytorch_lightning.utilities import rank_zero_only
+
+from . import pylogger, rich_utils
+
+log = pylogger.get_pylogger(__name__)
+
+
+def task_wrapper(task_func: Callable) -> Callable:
+    """Optional decorator that wraps the task function in extra utilities.
+
+    Makes multirun more resistant to failure.
+
+    Utilities:
+    - Calling the `utils.extras()` before the task is started
+    - Calling the `utils.close_loggers()` after the task is finished
+    - Logging the exception if occurs
+    - Logging the task total execution time
+    - Logging the output dir
+    """
+
+    def wrap(cfg: DictConfig):
+
+        # apply extra utilities
+        extras(cfg)
+
+        # execute the task
+        try:
+            start_time = time.time()
+            ret = task_func(cfg=cfg)
+        except Exception as ex:
+            log.exception("")  # save exception to `.log` file
+            raise ex
+        finally:
+            path = Path(cfg.paths.output_dir, "exec_time.log")
+            content = f"'{cfg.task_name}' execution time: {time.time() - start_time} (s)"
+            save_file(path, content)  # save task execution time (even if exception occurs)
+            close_loggers()  # close loggers (even if exception occurs so multirun won't fail)
+
+        log.info(f"Output dir: {cfg.paths.output_dir}")
+
+        return ret
+
+    return wrap
+
+
+def extras(cfg: DictConfig) -> None:
+    """Applies optional utilities before the task is started.
+
+    Utilities:
+    - Ignoring python warnings
+    - Setting tags from command line
+    - Rich config printing
+    """
+
+    # return if no `extras` config
+    if not cfg.get("extras"):
+        log.warning("Extras config not found! <cfg.extras=null>")
+        return
+
+    # disable python warnings
+    if cfg.extras.get("ignore_warnings"):
+        log.info("Disabling python warnings! <cfg.extras.ignore_warnings=True>")
+        warnings.filterwarnings("ignore")
+
+    # prompt user to input tags from command line if none are provided in the config
+    if cfg.extras.get("enforce_tags"):
+        log.info("Enforcing tags! <cfg.extras.enforce_tags=True>")
+        rich_utils.enforce_tags(cfg, save_to_file=True)
+
+    # pretty print config tree using Rich library
+    if cfg.extras.get("print_config"):
+        log.info("Printing config tree with Rich! <cfg.extras.print_config=True>")
+        rich_utils.print_config_tree(cfg, resolve=True, save_to_file=True)
+
+
+@rank_zero_only
+def save_file(path: str, content: str) -> None:
+    """Save file in rank zero mode (only on one process in multi-GPU setup)."""
+    with open(path, "w+") as file:
+        file.write(content)
+
+
+def instantiate_callbacks(callbacks_cfg: DictConfig) -> List[Callback]:
+    """Instantiates callbacks from config."""
+    callbacks: List[Callback] = []
+
+    if not callbacks_cfg:
+        log.warning("Callbacks config is empty.")
+        return callbacks
+
+    if not isinstance(callbacks_cfg, DictConfig):
+        raise TypeError("Callbacks config must be a DictConfig!")
+
+    for _, cb_conf in callbacks_cfg.items():
+        if isinstance(cb_conf, DictConfig) and "_target_" in cb_conf:
+            log.info(f"Instantiating callback <{cb_conf._target_}>")
+            callbacks.append(hydra.utils.instantiate(cb_conf))
+
+    return callbacks
+
+
+def instantiate_loggers(logger_cfg: DictConfig) -> List[Logger]:
+    """Instantiates loggers from config."""
+    logger: List[Logger] = []
+
+    if not logger_cfg:
+        log.warning("Logger config is empty.")
+        return logger
+
+    if not isinstance(logger_cfg, DictConfig):
+        raise TypeError("Logger config must be a DictConfig!")
+
+    for _, lg_conf in logger_cfg.items():
+        if isinstance(lg_conf, DictConfig) and "_target_" in lg_conf:
+            log.info(f"Instantiating logger <{lg_conf._target_}>")
+            logger.append(hydra.utils.instantiate(lg_conf))
+
+    return logger
+
+
+@rank_zero_only
+def log_hyperparameters(object_dict: dict) -> None:
+    """Controls which config parts are saved by lightning loggers.
+
+    Additionally saves:
+    - Number of model parameters
+    """
+
+    hparams = {}
+
+    cfg = object_dict["cfg"]
+    model = object_dict["model"]
+    trainer = object_dict["trainer"]
+
+    if not trainer.logger:
+        log.warning("Logger not found! Skipping hyperparameter logging...")
+        return
+
+    # save number of model parameters
+    hparams["model/params/total"] = sum(p.numel() for p in model.parameters())
+    hparams["model/params/trainable"] = sum(
+        p.numel() for p in model.parameters() if p.requires_grad
+    )
+    hparams["model/params/non_trainable"] = sum(
+        p.numel() for p in model.parameters() if not p.requires_grad
+    )
+
+    for k in cfg.keys():
+        hparams[k] = cfg.get(k)
+
+    # Resolve all interpolations
+    def _resolve(_cfg):
+        if isinstance(_cfg, DictConfig):
+            _cfg = OmegaConf.to_container(_cfg, resolve=True)
+        return _cfg
+
+    hparams = {k: _resolve(v) for k, v in hparams.items()}
+
+    # send hparams to all loggers
+    trainer.logger.log_hyperparams(hparams)
+
+
+def get_metric_value(metric_dict: dict, metric_name: str) -> float:
+    """Safely retrieves value of the metric logged in LightningModule."""
+
+    if not metric_name:
+        log.info("Metric name is None! Skipping metric value retrieval...")
+        return None
+
+    if metric_name not in metric_dict:
+        raise Exception(
+            f"Metric value not found! <metric_name={metric_name}>\n"
+            "Make sure metric name logged in LightningModule is correct!\n"
+            "Make sure `optimized_metric` name in `hparams_search` config is correct!"
+        )
+
+    metric_value = metric_dict[metric_name].item()
+    log.info(f"Retrieved metric value! <{metric_name}={metric_value}>")
+
+    return metric_value
+
+
+def close_loggers() -> None:
+    """Makes sure all loggers closed properly (prevents logging failure during multirun)."""
+
+    log.info("Closing loggers...")
+
+    if find_spec("wandb"):  # if wandb is installed
+        import wandb
+
+        if wandb.run:
+            log.info("Closing wandb!")
+            wandb.finish()

humans4d/hmr2/utils/pck_accuracy.py

@@ -0,0 +1,94 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+
+import numpy as np
+
+def _calc_distances(preds, targets, mask, normalize):
+    """Calculate the normalized distances between preds and target.
+
+    Note:
+        batch_size: N
+        num_keypoints: K
+        dimension of keypoints: D (normally, D=2 or D=3)
+
+    Args:
+        preds (np.ndarray[N, K, D]): Predicted keypoint location.
+        targets (np.ndarray[N, K, D]): Groundtruth keypoint location.
+        mask (np.ndarray[N, K]): Visibility of the target. False for invisible
+            joints, and True for visible. Invisible joints will be ignored for
+            accuracy calculation.
+        normalize (np.ndarray[N, D]): Typical value is heatmap_size
+
+    Returns:
+        np.ndarray[K, N]: The normalized distances. \
+            If target keypoints are missing, the distance is -1.
+    """
+    N, K, _ = preds.shape
+    # set mask=0 when normalize==0
+    _mask = mask.copy()
+    _mask[np.where((normalize == 0).sum(1))[0], :] = False
+    distances = np.full((N, K), -1, dtype=np.float32)
+    # handle invalid values
+    normalize[np.where(normalize <= 0)] = 1e6
+    distances[_mask] = np.linalg.norm(
+        ((preds - targets) / normalize[:, None, :])[_mask], axis=-1)
+    return distances.T
+
+
+def _distance_acc(distances, thr=0.5):
+    """Return the percentage below the distance threshold, while ignoring
+    distances values with -1.
+
+    Note:
+        batch_size: N
+    Args:
+        distances (np.ndarray[N, ]): The normalized distances.
+        thr (float): Threshold of the distances.
+
+    Returns:
+        float: Percentage of distances below the threshold. \
+            If all target keypoints are missing, return -1.
+    """
+    distance_valid = distances != -1
+    num_distance_valid = distance_valid.sum()
+    if num_distance_valid > 0:
+        return (distances[distance_valid] < thr).sum() / num_distance_valid
+    return -1
+
+
+def keypoint_pck_accuracy(pred, gt, mask, thr, normalize):
+    """Calculate the pose accuracy of PCK for each individual keypoint and the
+    averaged accuracy across all keypoints for coordinates.
+
+    Note:
+        PCK metric measures accuracy of the localization of the body joints.
+        The distances between predicted positions and the ground-truth ones
+        are typically normalized by the bounding box size.
+        The threshold (thr) of the normalized distance is commonly set
+        as 0.05, 0.1 or 0.2 etc.
+
+        - batch_size: N
+        - num_keypoints: K
+
+    Args:
+        pred (np.ndarray[N, K, 2]): Predicted keypoint location.
+        gt (np.ndarray[N, K, 2]): Groundtruth keypoint location.
+        mask (np.ndarray[N, K]): Visibility of the target. False for invisible
+            joints, and True for visible. Invisible joints will be ignored for
+            accuracy calculation.
+        thr (float): Threshold of PCK calculation.
+        normalize (np.ndarray[N, 2]): Normalization factor for H&W.
+
+    Returns:
+        tuple: A tuple containing keypoint accuracy.
+
+        - acc (np.ndarray[K]): Accuracy of each keypoint.
+        - avg_acc (float): Averaged accuracy across all keypoints.
+        - cnt (int): Number of valid keypoints.
+    """
+    distances = _calc_distances(pred, gt, mask, normalize)
+
+    acc = np.array([_distance_acc(d, thr) for d in distances])
+    valid_acc = acc[acc >= 0]
+    cnt = len(valid_acc)
+    avg_acc = valid_acc.mean() if cnt > 0 else 0
+    return acc, avg_acc, cnt

humans4d/hmr2/utils/pose_utils.py

@@ -0,0 +1,310 @@
+"""
+Code adapted from: https://github.com/akanazawa/hmr/blob/master/src/benchmark/eval_util.py
+"""
+
+import torch
+import numpy as np
+from typing import Optional, Dict, List, Tuple
+
+def compute_similarity_transform(S1: torch.Tensor, S2: torch.Tensor) -> torch.Tensor:
+    """
+    Computes a similarity transform (sR, t) in a batched way that takes
+    a set of 3D points S1 (B, N, 3) closest to a set of 3D points S2 (B, N, 3),
+    where R is a 3x3 rotation matrix, t 3x1 translation, s scale.
+    i.e. solves the orthogonal Procrutes problem.
+    Args:
+        S1 (torch.Tensor): First set of points of shape (B, N, 3).
+        S2 (torch.Tensor): Second set of points of shape (B, N, 3).
+    Returns:
+        (torch.Tensor): The first set of points after applying the similarity transformation.
+    """
+
+    batch_size = S1.shape[0]
+    S1 = S1.permute(0, 2, 1)
+    S2 = S2.permute(0, 2, 1)
+    # 1. Remove mean.
+    mu1 = S1.mean(dim=2, keepdim=True)
+    mu2 = S2.mean(dim=2, keepdim=True)
+    X1 = S1 - mu1
+    X2 = S2 - mu2
+
+    # 2. Compute variance of X1 used for scale.
+    var1 = (X1**2).sum(dim=(1,2))
+
+    # 3. The outer product of X1 and X2.
+    K = torch.matmul(X1, X2.permute(0, 2, 1))
+
+    # 4. Solution that Maximizes trace(R'K) is R=U*V', where U, V are singular vectors of K.
+    U, s, V = torch.svd(K)
+    Vh = V.permute(0, 2, 1)
+
+    # Construct Z that fixes the orientation of R to get det(R)=1.
+    Z = torch.eye(U.shape[1], device=U.device).unsqueeze(0).repeat(batch_size, 1, 1)
+    Z[:, -1, -1] *= torch.sign(torch.linalg.det(torch.matmul(U, Vh)))
+
+    # Construct R.
+    R = torch.matmul(torch.matmul(V, Z), U.permute(0, 2, 1))
+
+    # 5. Recover scale.
+    trace = torch.matmul(R, K).diagonal(offset=0, dim1=-1, dim2=-2).sum(dim=-1)
+    scale = (trace / var1).unsqueeze(dim=-1).unsqueeze(dim=-1)
+
+    # 6. Recover translation.
+    t = mu2 - scale*torch.matmul(R, mu1)
+
+    # 7. Error:
+    S1_hat = scale*torch.matmul(R, S1) + t
+
+    return S1_hat.permute(0, 2, 1)
+
+def reconstruction_error(S1, S2) -> np.array:
+    """
+    Computes the mean Euclidean distance of 2 set of points S1, S2 after performing Procrustes alignment.
+    Args:
+        S1 (torch.Tensor): First set of points of shape (B, N, 3).
+        S2 (torch.Tensor): Second set of points of shape (B, N, 3).
+    Returns:
+        (np.array): Reconstruction error.
+    """
+    S1_hat = compute_similarity_transform(S1, S2)
+    re = torch.sqrt( ((S1_hat - S2)** 2).sum(dim=-1)).mean(dim=-1)
+    return re
+
+def eval_pose(pred_joints, gt_joints) -> Tuple[np.array, np.array]:
+    """
+    Compute joint errors in mm before and after Procrustes alignment.
+    Args:
+        pred_joints (torch.Tensor): Predicted 3D joints of shape (B, N, 3).
+        gt_joints (torch.Tensor): Ground truth 3D joints of shape (B, N, 3).
+    Returns:
+        Tuple[np.array, np.array]: Joint errors in mm before and after alignment.
+    """
+    # Absolute error (MPJPE)
+    mpjpe = torch.sqrt(((pred_joints - gt_joints) ** 2).sum(dim=-1)).mean(dim=-1).cpu().numpy()
+
+    # Reconstruction_error
+    r_error = reconstruction_error(pred_joints, gt_joints).cpu().numpy()
+    return 1000 * mpjpe, 1000 * r_error
+
+class Evaluator:
+
+    def __init__(self,
+                 dataset_length: int,
+                 keypoint_list: List,
+                 pelvis_ind: int,
+                 metrics: List = ['mode_mpjpe', 'mode_re', 'min_mpjpe', 'min_re'],
+                 pck_thresholds: Optional[List] = None):
+        """
+        Class used for evaluating trained models on different 3D pose datasets.
+        Args:
+            dataset_length (int): Total dataset length.
+            keypoint_list [List]: List of keypoints used for evaluation.
+            pelvis_ind (int): Index of pelvis keypoint; used for aligning the predictions and ground truth.
+            metrics [List]: List of evaluation metrics to record.
+        """
+        self.dataset_length = dataset_length
+        self.keypoint_list = keypoint_list
+        self.pelvis_ind = pelvis_ind
+        self.metrics = metrics
+        for metric in self.metrics:
+            setattr(self, metric, np.zeros((dataset_length,)))
+        self.counter = 0
+        if pck_thresholds is None:
+            self.pck_evaluator = None
+        else:
+            self.pck_evaluator = EvaluatorPCK(pck_thresholds)
+
+    def log(self):
+        """
+        Print current evaluation metrics
+        """
+        if self.counter == 0:
+            print('Evaluation has not started')
+            return
+        print(f'{self.counter} / {self.dataset_length} samples')
+        if self.pck_evaluator is not None:
+            self.pck_evaluator.log()
+        for metric in self.metrics:
+            if metric in ['mode_mpjpe', 'mode_re', 'min_mpjpe', 'min_re']:
+                unit = 'mm'
+            else:
+                unit = ''
+            print(f'{metric}: {getattr(self, metric)[:self.counter].mean()} {unit}')
+        print('***')
+
+    def get_metrics_dict(self) -> Dict:
+        """
+        Returns:
+            Dict: Dictionary of evaluation metrics.
+        """
+        d1 = {metric: getattr(self, metric)[:self.counter].mean() for metric in self.metrics}
+        if self.pck_evaluator is not None:
+            d2 = self.pck_evaluator.get_metrics_dict()
+            d1.update(d2)
+        return d1
+
+    def __call__(self, output: Dict, batch: Dict, opt_output: Optional[Dict] = None):
+        """
+        Evaluate current batch.
+        Args:
+            output (Dict): Regression output.
+            batch (Dict): Dictionary containing images and their corresponding annotations.
+            opt_output (Dict): Optimization output.
+        """
+        if self.pck_evaluator is not None:
+            self.pck_evaluator(output, batch, opt_output)
+
+        pred_keypoints_3d = output['pred_keypoints_3d'].detach()
+        pred_keypoints_3d = pred_keypoints_3d[:,None,:,:]
+        batch_size = pred_keypoints_3d.shape[0]
+        num_samples = pred_keypoints_3d.shape[1]
+        gt_keypoints_3d = batch['keypoints_3d'][:, :, :-1].unsqueeze(1).repeat(1, num_samples, 1, 1)
+
+        # Align predictions and ground truth such that the pelvis location is at the origin
+        pred_keypoints_3d -= pred_keypoints_3d[:, :, [self.pelvis_ind]]
+        gt_keypoints_3d -= gt_keypoints_3d[:, :, [self.pelvis_ind]]
+
+        # Compute joint errors
+        mpjpe, re = eval_pose(pred_keypoints_3d.reshape(batch_size * num_samples, -1, 3)[:, self.keypoint_list], gt_keypoints_3d.reshape(batch_size * num_samples, -1 ,3)[:, self.keypoint_list])
+        mpjpe = mpjpe.reshape(batch_size, num_samples)
+        re = re.reshape(batch_size, num_samples)
+
+        # Compute 2d keypoint errors
+        pred_keypoints_2d = output['pred_keypoints_2d'].detach()
+        pred_keypoints_2d = pred_keypoints_2d[:,None,:,:]
+        gt_keypoints_2d = batch['keypoints_2d'][:,None,:,:].repeat(1, num_samples, 1, 1)
+        conf = gt_keypoints_2d[:, :, :, -1].clone()
+        kp_err = torch.nn.functional.mse_loss(
+                        pred_keypoints_2d,
+                        gt_keypoints_2d[:, :, :, :-1],
+                        reduction='none'
+                    ).sum(dim=3)
+        kp_l2_loss = (conf * kp_err).mean(dim=2)
+        kp_l2_loss = kp_l2_loss.detach().cpu().numpy()
+
+        # Compute joint errors after optimization, if available.
+        if opt_output is not None:
+            opt_keypoints_3d = opt_output['model_joints']
+            opt_keypoints_3d -= opt_keypoints_3d[:, [self.pelvis_ind]]
+            opt_mpjpe, opt_re = eval_pose(opt_keypoints_3d[:, self.keypoint_list], gt_keypoints_3d[:, 0, self.keypoint_list])
+
+        # The 0-th sample always corresponds to the mode
+        if hasattr(self, 'mode_mpjpe'):
+            mode_mpjpe = mpjpe[:, 0]
+            self.mode_mpjpe[self.counter:self.counter+batch_size] = mode_mpjpe
+        if hasattr(self, 'mode_re'):
+            mode_re = re[:, 0]
+            self.mode_re[self.counter:self.counter+batch_size] = mode_re
+        if hasattr(self, 'mode_kpl2'):
+            mode_kpl2 = kp_l2_loss[:, 0]
+            self.mode_kpl2[self.counter:self.counter+batch_size] = mode_kpl2
+        if hasattr(self, 'min_mpjpe'):
+            min_mpjpe = mpjpe.min(axis=-1)
+            self.min_mpjpe[self.counter:self.counter+batch_size] = min_mpjpe
+        if hasattr(self, 'min_re'):
+            min_re = re.min(axis=-1)
+            self.min_re[self.counter:self.counter+batch_size] = min_re
+        if hasattr(self, 'min_kpl2'):
+            min_kpl2 = kp_l2_loss.min(axis=-1)
+            self.min_kpl2[self.counter:self.counter+batch_size] = min_kpl2
+        if hasattr(self, 'opt_mpjpe'):
+            self.opt_mpjpe[self.counter:self.counter+batch_size] = opt_mpjpe
+        if hasattr(self, 'opt_re'):
+            self.opt_re[self.counter:self.counter+batch_size] = opt_re
+
+        self.counter += batch_size
+
+        if hasattr(self, 'mode_mpjpe') and hasattr(self, 'mode_re'):
+            return {
+                'mode_mpjpe': mode_mpjpe,
+                'mode_re': mode_re,
+            }
+        else:
+            return {}
+
+
+class EvaluatorPCK:
+
+    def __init__(self, thresholds: List = [0.05, 0.1, 0.2, 0.3, 0.4, 0.5],):
+        """
+        Class used for evaluating trained models on different 3D pose datasets.
+        Args:
+            thresholds [List]: List of PCK thresholds to evaluate.
+            metrics [List]: List of evaluation metrics to record.
+        """
+        self.thresholds = thresholds
+        self.pred_kp_2d = []
+        self.gt_kp_2d = []
+        self.gt_conf_2d = []
+        self.counter = 0
+
+    def log(self):
+        """
+        Print current evaluation metrics
+        """
+        if self.counter == 0:
+            print('Evaluation has not started')
+            return
+        print(f'{self.counter} samples')
+        metrics_dict = self.get_metrics_dict()
+        for metric in metrics_dict:
+            print(f'{metric}: {metrics_dict[metric]}')
+        print('***')
+
+    def get_metrics_dict(self) -> Dict:
+        """
+        Returns:
+            Dict: Dictionary of evaluation metrics.
+        """
+        pcks = self.compute_pcks()
+        metrics = {}
+        for thr, (acc,avg_acc,cnt) in zip(self.thresholds, pcks):
+            metrics.update({f'kp{i}_pck_{thr}': float(a) for i, a in enumerate(acc) if a>=0})
+            metrics.update({f'kpAvg_pck_{thr}': float(avg_acc)})
+        return metrics
+
+    def compute_pcks(self):
+        pred_kp_2d = np.concatenate(self.pred_kp_2d, axis=0)
+        gt_kp_2d = np.concatenate(self.gt_kp_2d, axis=0)
+        gt_conf_2d = np.concatenate(self.gt_conf_2d, axis=0)
+        assert pred_kp_2d.shape == gt_kp_2d.shape
+        assert pred_kp_2d[..., 0].shape == gt_conf_2d.shape
+        assert pred_kp_2d.shape[1] == 1 # num_samples
+
+        from .pck_accuracy import keypoint_pck_accuracy
+        pcks = [
+            keypoint_pck_accuracy(
+                pred_kp_2d[:, 0, :, :],
+                gt_kp_2d[:, 0, :, :],
+                gt_conf_2d[:, 0, :]>0.5,
+                thr=thr,
+                normalize = np.ones((len(pred_kp_2d),2))   # Already in [-0.5,0.5] range. No need to normalize
+            )
+            for thr in self.thresholds
+        ]
+        return pcks
+
+    def __call__(self, output: Dict, batch: Dict, opt_output: Optional[Dict] = None):
+        """
+        Evaluate current batch.
+        Args:
+            output (Dict): Regression output.
+            batch (Dict): Dictionary containing images and their corresponding annotations.
+            opt_output (Dict): Optimization output.
+        """
+        pred_keypoints_2d = output['pred_keypoints_2d'].detach()
+        num_samples = 1
+        batch_size = pred_keypoints_2d.shape[0]
+
+        pred_keypoints_2d = pred_keypoints_2d[:,None,:,:]
+        gt_keypoints_2d = batch['keypoints_2d'][:,None,:,:].repeat(1, num_samples, 1, 1)
+
+        gt_bbox_expand_factor = (batch['box_size']/(batch['_scale']*200).max(dim=-1).values)
+        gt_bbox_expand_factor = gt_bbox_expand_factor[:,None,None,None].repeat(1, num_samples, 1, 1)
+        gt_bbox_expand_factor = gt_bbox_expand_factor.detach().cpu().numpy()
+
+        self.pred_kp_2d.append(pred_keypoints_2d[:, :, :, :2].detach().cpu().numpy() * gt_bbox_expand_factor)
+        self.gt_conf_2d.append(gt_keypoints_2d[:, :, :, -1].detach().cpu().numpy())
+        self.gt_kp_2d.append(gt_keypoints_2d[:, :, :, :2].detach().cpu().numpy() * gt_bbox_expand_factor)
+
+        self.counter += batch_size

humans4d/hmr2/utils/pylogger.py

@@ -0,0 +1,17 @@
+import logging
+
+from pytorch_lightning.utilities import rank_zero_only
+
+
+def get_pylogger(name=__name__) -> logging.Logger:
+    """Initializes multi-GPU-friendly python command line logger."""
+
+    logger = logging.getLogger(name)
+
+    # this ensures all logging levels get marked with the rank zero decorator
+    # otherwise logs would get multiplied for each GPU process in multi-GPU setup
+    logging_levels = ("debug", "info", "warning", "error", "exception", "fatal", "critical")
+    for level in logging_levels:
+        setattr(logger, level, rank_zero_only(getattr(logger, level)))
+
+    return logger

humans4d/hmr2/utils/render_openpose.py

@@ -0,0 +1,149 @@
+"""
+Render OpenPose keypoints.
+Code was ported to Python from the official C++ implementation https://github.com/CMU-Perceptual-Computing-Lab/openpose/blob/master/src/openpose/utilities/keypoint.cpp
+"""
+import cv2
+import math
+import numpy as np
+from typing import List, Tuple
+
+def get_keypoints_rectangle(keypoints: np.array, threshold: float) -> Tuple[float, float, float]:
+    """
+    Compute rectangle enclosing keypoints above the threshold.
+    Args:
+        keypoints (np.array): Keypoint array of shape (N, 3).
+        threshold (float): Confidence visualization threshold.
+    Returns:
+        Tuple[float, float, float]: Rectangle width, height and area.
+    """
+    valid_ind = keypoints[:, -1] > threshold
+    if valid_ind.sum() > 0:
+        valid_keypoints = keypoints[valid_ind][:, :-1]
+        max_x = valid_keypoints[:,0].max()
+        max_y = valid_keypoints[:,1].max()
+        min_x = valid_keypoints[:,0].min()
+        min_y = valid_keypoints[:,1].min()
+        width = max_x - min_x
+        height = max_y - min_y
+        area = width * height
+        return width, height, area
+    else:
+        return 0,0,0
+
+def render_keypoints(img: np.array,
+                     keypoints: np.array,
+                     pairs: List,
+                     colors: List,
+                     thickness_circle_ratio: float,
+                     thickness_line_ratio_wrt_circle: float,
+                     pose_scales: List,
+                     threshold: float = 0.1) -> np.array:
+    """
+    Render keypoints on input image.
+    Args:
+        img (np.array): Input image of shape (H, W, 3) with pixel values in the [0,255] range.
+        keypoints (np.array): Keypoint array of shape (N, 3).
+        pairs (List): List of keypoint pairs per limb.
+        colors: (List): List of colors per keypoint.
+        thickness_circle_ratio (float): Circle thickness ratio.
+        thickness_line_ratio_wrt_circle (float): Line thickness ratio wrt the circle.
+        pose_scales (List): List of pose scales.
+        threshold (float): Only visualize keypoints with confidence above the threshold.
+    Returns:
+        (np.array): Image of shape (H, W, 3) with keypoints drawn on top of the original image. 
+    """
+    img_orig = img.copy()
+    width, height = img.shape[1], img.shape[2]
+    area = width * height
+
+    lineType = 8
+    shift = 0
+    numberColors = len(colors)
+    thresholdRectangle = 0.1
+
+    person_width, person_height, person_area = get_keypoints_rectangle(keypoints, thresholdRectangle)
+    if person_area > 0:
+        ratioAreas = min(1, max(person_width / width, person_height / height))
+        thicknessRatio = np.maximum(np.round(math.sqrt(area) * thickness_circle_ratio * ratioAreas), 2)
+        thicknessCircle = np.maximum(1, thicknessRatio if ratioAreas > 0.05 else -np.ones_like(thicknessRatio))
+        thicknessLine = np.maximum(1, np.round(thicknessRatio * thickness_line_ratio_wrt_circle))
+        radius = thicknessRatio / 2
+
+        img = np.ascontiguousarray(img.copy())
+        for i, pair in enumerate(pairs):
+            index1, index2 = pair
+            if keypoints[index1, -1] > threshold and keypoints[index2, -1] > threshold:
+                thicknessLineScaled = int(round(min(thicknessLine[index1], thicknessLine[index2]) * pose_scales[0]))
+                colorIndex = index2
+                color = colors[colorIndex % numberColors]
+                keypoint1 = keypoints[index1, :-1].astype(np.int)
+                keypoint2 = keypoints[index2, :-1].astype(np.int)
+                cv2.line(img, tuple(keypoint1.tolist()), tuple(keypoint2.tolist()), tuple(color.tolist()), thicknessLineScaled, lineType, shift)
+        for part in range(len(keypoints)):
+            faceIndex = part
+            if keypoints[faceIndex, -1] > threshold:
+                radiusScaled = int(round(radius[faceIndex] * pose_scales[0]))
+                thicknessCircleScaled = int(round(thicknessCircle[faceIndex] * pose_scales[0]))
+                colorIndex = part
+                color = colors[colorIndex % numberColors]
+                center = keypoints[faceIndex, :-1].astype(np.int)
+                cv2.circle(img, tuple(center.tolist()), radiusScaled, tuple(color.tolist()), thicknessCircleScaled, lineType, shift)
+    return img
+
+def render_body_keypoints(img: np.array,
+                          body_keypoints: np.array) -> np.array:
+    """
+    Render OpenPose body keypoints on input image.
+    Args:
+        img (np.array): Input image of shape (H, W, 3) with pixel values in the [0,255] range.
+        body_keypoints (np.array): Keypoint array of shape (N, 3); 3 <====> (x, y, confidence).
+    Returns:
+        (np.array): Image of shape (H, W, 3) with keypoints drawn on top of the original image. 
+    """
+
+    thickness_circle_ratio = 1./75. * np.ones(body_keypoints.shape[0])
+    thickness_line_ratio_wrt_circle = 0.75
+    pairs = []
+    pairs = [1,8,1,2,1,5,2,3,3,4,5,6,6,7,8,9,9,10,10,11,8,12,12,13,13,14,1,0,0,15,15,17,0,16,16,18,14,19,19,20,14,21,11,22,22,23,11,24]
+    pairs = np.array(pairs).reshape(-1,2)
+    colors = [255.,     0.,     85.,
+              255.,     0.,     0.,
+              255.,    85.,     0.,
+              255.,   170.,     0.,
+              255.,   255.,     0.,
+              170.,   255.,     0.,
+               85.,   255.,     0.,
+                0.,   255.,     0.,
+              255.,     0.,     0.,
+                0.,   255.,    85.,
+                0.,   255.,   170.,
+                0.,   255.,   255.,
+                0.,   170.,   255.,
+                0.,    85.,   255.,
+                0.,     0.,   255.,
+              255.,     0.,   170.,
+              170.,     0.,   255.,
+              255.,     0.,   255.,
+               85.,     0.,   255.,
+                0.,     0.,   255.,
+                0.,     0.,   255.,
+                0.,     0.,   255.,
+                0.,   255.,   255.,
+                0.,   255.,   255.,
+                0.,   255.,   255.]
+    colors = np.array(colors).reshape(-1,3)
+    pose_scales = [1]
+    return render_keypoints(img, body_keypoints, pairs, colors, thickness_circle_ratio, thickness_line_ratio_wrt_circle, pose_scales, 0.1)
+
+def render_openpose(img: np.array,
+                    body_keypoints: np.array) -> np.array:
+    """
+    Render keypoints in the OpenPose format on input image.
+    Args:
+        img (np.array): Input image of shape (H, W, 3) with pixel values in the [0,255] range.
+        body_keypoints (np.array): Keypoint array of shape (N, 3); 3 <====> (x, y, confidence).
+    Returns:
+        (np.array): Image of shape (H, W, 3) with keypoints drawn on top of the original image. 
+    """
+    img = render_body_keypoints(img, body_keypoints)
+    return img

humans4d/hmr2/utils/renderer.py

@@ -0,0 +1,400 @@
+import os
+if 'PYOPENGL_PLATFORM' not in os.environ:
+    os.environ['PYOPENGL_PLATFORM'] = 'egl'
+import torch
+import numpy as np
+import pyrender
+import trimesh
+import cv2
+from yacs.config import CfgNode
+from typing import List, Optional
+
+def cam_crop_to_full(cam_bbox, box_center, box_size, img_size, focal_length=5000.):
+    # Convert cam_bbox to full image
+    img_w, img_h = img_size[:, 0], img_size[:, 1]
+    cx, cy, b = box_center[:, 0], box_center[:, 1], box_size
+    w_2, h_2 = img_w / 2., img_h / 2.
+    bs = b * cam_bbox[:, 0] + 1e-9
+    tz = 2 * focal_length / bs
+    tx = (2 * (cx - w_2) / bs) + cam_bbox[:, 1]
+    ty = (2 * (cy - h_2) / bs) + cam_bbox[:, 2]
+    full_cam = torch.stack([tx, ty, tz], dim=-1)
+    return full_cam
+
+def get_light_poses(n_lights=5, elevation=np.pi / 3, dist=12):
+    # get lights in a circle around origin at elevation
+    thetas = elevation * np.ones(n_lights)
+    phis = 2 * np.pi * np.arange(n_lights) / n_lights
+    poses = []
+    trans = make_translation(torch.tensor([0, 0, dist]))
+    for phi, theta in zip(phis, thetas):
+        rot = make_rotation(rx=-theta, ry=phi, order="xyz")
+        poses.append((rot @ trans).numpy())
+    return poses
+
+def make_translation(t):
+    return make_4x4_pose(torch.eye(3), t)
+
+def make_rotation(rx=0, ry=0, rz=0, order="xyz"):
+    Rx = rotx(rx)
+    Ry = roty(ry)
+    Rz = rotz(rz)
+    if order == "xyz":
+        R = Rz @ Ry @ Rx
+    elif order == "xzy":
+        R = Ry @ Rz @ Rx
+    elif order == "yxz":
+        R = Rz @ Rx @ Ry
+    elif order == "yzx":
+        R = Rx @ Rz @ Ry
+    elif order == "zyx":
+        R = Rx @ Ry @ Rz
+    elif order == "zxy":
+        R = Ry @ Rx @ Rz
+    return make_4x4_pose(R, torch.zeros(3))
+
+def make_4x4_pose(R, t):
+    """
+    :param R (*, 3, 3)
+    :param t (*, 3)
+    return (*, 4, 4)
+    """
+    dims = R.shape[:-2]
+    pose_3x4 = torch.cat([R, t.view(*dims, 3, 1)], dim=-1)
+    bottom = (
+        torch.tensor([0, 0, 0, 1], device=R.device)
+        .reshape(*(1,) * len(dims), 1, 4)
+        .expand(*dims, 1, 4)
+    )
+    return torch.cat([pose_3x4, bottom], dim=-2)
+
+
+def rotx(theta):
+    return torch.tensor(
+        [
+            [1, 0, 0],
+            [0, np.cos(theta), -np.sin(theta)],
+            [0, np.sin(theta), np.cos(theta)],
+        ],
+        dtype=torch.float32,
+    )
+
+
+def roty(theta):
+    return torch.tensor(
+        [
+            [np.cos(theta), 0, np.sin(theta)],
+            [0, 1, 0],
+            [-np.sin(theta), 0, np.cos(theta)],
+        ],
+        dtype=torch.float32,
+    )
+
+
+def rotz(theta):
+    return torch.tensor(
+        [
+            [np.cos(theta), -np.sin(theta), 0],
+            [np.sin(theta), np.cos(theta), 0],
+            [0, 0, 1],
+        ],
+        dtype=torch.float32,
+    )
+    
+
+def create_raymond_lights() -> List[pyrender.Node]:
+    """
+    Return raymond light nodes for the scene.
+    """
+    thetas = np.pi * np.array([1.0 / 6.0, 1.0 / 6.0, 1.0 / 6.0])
+    phis = np.pi * np.array([0.0, 2.0 / 3.0, 4.0 / 3.0])
+
+    nodes = []
+
+    for phi, theta in zip(phis, thetas):
+        xp = np.sin(theta) * np.cos(phi)
+        yp = np.sin(theta) * np.sin(phi)
+        zp = np.cos(theta)
+
+        z = np.array([xp, yp, zp])
+        z = z / np.linalg.norm(z)
+        x = np.array([-z[1], z[0], 0.0])
+        if np.linalg.norm(x) == 0:
+            x = np.array([1.0, 0.0, 0.0])
+        x = x / np.linalg.norm(x)
+        y = np.cross(z, x)
+
+        matrix = np.eye(4)
+        matrix[:3,:3] = np.c_[x,y,z]
+        nodes.append(pyrender.Node(
+            light=pyrender.DirectionalLight(color=np.ones(3), intensity=1.0),
+            matrix=matrix
+        ))
+
+    return nodes
+
+class Renderer:
+
+    def __init__(self, cfg: CfgNode, faces: np.array):
+        """
+        Wrapper around the pyrender renderer to render SMPL meshes.
+        Args:
+            cfg (CfgNode): Model config file.
+            faces (np.array): Array of shape (F, 3) containing the mesh faces.
+        """
+        self.cfg = cfg
+        self.focal_length = cfg.EXTRA.FOCAL_LENGTH
+        self.img_res = cfg.MODEL.IMAGE_SIZE
+
+        self.camera_center = [self.img_res // 2, self.img_res // 2]
+        self.faces = faces
+
+    def __call__(self,
+                vertices: np.array,
+                camera_translation: np.array,
+                image: torch.Tensor,
+                full_frame: bool = False,
+                imgname: Optional[str] = None,
+                side_view=False, top_view=False,
+                rot_angle=90,
+                mesh_base_color=(1.0, 1.0, 0.9),
+                scene_bg_color=(0,0,0),
+                return_rgba=False,
+                ) -> np.array:
+        """
+        Render meshes on input image
+        Args:
+            vertices (np.array): Array of shape (V, 3) containing the mesh vertices.
+            camera_translation (np.array): Array of shape (3,) with the camera translation.
+            image (torch.Tensor): Tensor of shape (3, H, W) containing the image crop with normalized pixel values.
+            full_frame (bool): If True, then render on the full image.
+            imgname (Optional[str]): Contains the original image filenamee. Used only if full_frame == True.
+        """
+        
+        if full_frame:
+            image = cv2.imread(imgname).astype(np.float32)[:, :, ::-1] / 255.
+        else:
+            image = image.clone() * torch.tensor(self.cfg.MODEL.IMAGE_STD, device=image.device).reshape(3,1,1)
+            image = image + torch.tensor(self.cfg.MODEL.IMAGE_MEAN, device=image.device).reshape(3,1,1)
+            image = image.permute(1, 2, 0).cpu().numpy()
+
+        renderer = pyrender.OffscreenRenderer(viewport_width=image.shape[1],
+                                              viewport_height=image.shape[0],
+                                              point_size=1.0)
+        material = pyrender.MetallicRoughnessMaterial(
+            metallicFactor=0.0,
+            alphaMode='OPAQUE',
+            baseColorFactor=(*mesh_base_color, 1.0))
+
+        camera_translation[0] *= -1.
+
+        mesh = trimesh.Trimesh(vertices.copy(), self.faces.copy())
+        if side_view:
+            rot = trimesh.transformations.rotation_matrix(
+                np.radians(rot_angle), [0, 1, 0])
+            mesh.apply_transform(rot)
+        elif top_view:
+            rot = trimesh.transformations.rotation_matrix(
+                np.radians(rot_angle), [1, 0, 0])
+            mesh.apply_transform(rot)
+        rot = trimesh.transformations.rotation_matrix(
+            np.radians(180), [1, 0, 0])
+        mesh.apply_transform(rot)
+        mesh = pyrender.Mesh.from_trimesh(mesh, material=material)
+
+        scene = pyrender.Scene(bg_color=[*scene_bg_color, 0.0],
+                               ambient_light=(0.3, 0.3, 0.3))
+        scene.add(mesh, 'mesh')
+
+        camera_pose = np.eye(4)
+        camera_pose[:3, 3] = camera_translation
+        camera_center = [image.shape[1] / 2., image.shape[0] / 2.]
+        camera = pyrender.IntrinsicsCamera(fx=self.focal_length, fy=self.focal_length,
+                                           cx=camera_center[0], cy=camera_center[1], zfar=1e12)
+        scene.add(camera, pose=camera_pose)
+
+
+        light_nodes = create_raymond_lights()
+        for node in light_nodes:
+            scene.add_node(node)
+
+        color, rend_depth = renderer.render(scene, flags=pyrender.RenderFlags.RGBA)
+        color = color.astype(np.float32) / 255.0
+        renderer.delete()
+
+        if return_rgba:
+            return color
+
+        valid_mask = (color[:, :, -1])[:, :, np.newaxis]
+        if not side_view and not top_view:
+            output_img = (color[:, :, :3] * valid_mask + (1 - valid_mask) * image)
+        else:
+            output_img = color[:, :, :3]
+
+        output_img = output_img.astype(np.float32)
+        return output_img
+
+    def vertices_to_trimesh(self, vertices, camera_translation, mesh_base_color=(1.0, 1.0, 0.9), 
+                            rot_axis=[1,0,0], rot_angle=0,):
+        # material = pyrender.MetallicRoughnessMaterial(
+        #     metallicFactor=0.0,
+        #     alphaMode='OPAQUE',
+        #     baseColorFactor=(*mesh_base_color, 1.0))
+        vertex_colors = np.array([(*mesh_base_color, 1.0)] * vertices.shape[0])
+        print(vertices.shape, camera_translation.shape)
+        mesh = trimesh.Trimesh(vertices.copy() + camera_translation, self.faces.copy(), vertex_colors=vertex_colors)
+        # mesh = trimesh.Trimesh(vertices.copy(), self.faces.copy())
+        
+        rot = trimesh.transformations.rotation_matrix(
+                np.radians(rot_angle), rot_axis)
+        mesh.apply_transform(rot)
+
+        rot = trimesh.transformations.rotation_matrix(
+            np.radians(180), [1, 0, 0])
+        mesh.apply_transform(rot)
+        return mesh
+
+    def render_rgba(
+            self,
+            vertices: np.array,
+            cam_t = None,
+            rot=None,
+            rot_axis=[1,0,0],
+            rot_angle=0,
+            camera_z=3,
+            # camera_translation: np.array,
+            mesh_base_color=(1.0, 1.0, 0.9),
+            scene_bg_color=(0,0,0),
+            render_res=[256, 256],
+        ):
+
+        renderer = pyrender.OffscreenRenderer(viewport_width=render_res[0],
+                                              viewport_height=render_res[1],
+                                              point_size=1.0)
+        # material = pyrender.MetallicRoughnessMaterial(
+        #     metallicFactor=0.0,
+        #     alphaMode='OPAQUE',
+        #     baseColorFactor=(*mesh_base_color, 1.0))
+
+        if cam_t is not None:
+            camera_translation = cam_t.copy()
+            # camera_translation[0] *= -1.
+        else:
+            camera_translation = np.array([0, 0, camera_z * self.focal_length/render_res[1]])
+
+        mesh = self.vertices_to_trimesh(vertices, camera_translation, mesh_base_color, rot_axis, rot_angle)
+        mesh = pyrender.Mesh.from_trimesh(mesh)
+        # mesh = pyrender.Mesh.from_trimesh(mesh, material=material)
+
+        scene = pyrender.Scene(bg_color=[*scene_bg_color, 0.0],
+                               ambient_light=(0.3, 0.3, 0.3))
+        scene.add(mesh, 'mesh')
+
+        camera_pose = np.eye(4)
+        # camera_pose[:3, 3] = camera_translation
+        camera_center = [render_res[0] / 2., render_res[1] / 2.]
+        camera = pyrender.IntrinsicsCamera(fx=self.focal_length, fy=self.focal_length,
+                                           cx=camera_center[0], cy=camera_center[1], zfar=1e12)
+
+        # Create camera node and add it to pyRender scene
+        camera_node = pyrender.Node(camera=camera, matrix=camera_pose)
+        scene.add_node(camera_node)
+        self.add_point_lighting(scene, camera_node)
+        self.add_lighting(scene, camera_node)
+
+        light_nodes = create_raymond_lights()
+        for node in light_nodes:
+            scene.add_node(node)
+
+        color, rend_depth = renderer.render(scene, flags=pyrender.RenderFlags.RGBA)
+        color = color.astype(np.float32) / 255.0
+        renderer.delete()
+
+        return color
+
+    def render_rgba_multiple(
+            self,
+            vertices: List[np.array],
+            cam_t: List[np.array],
+            rot_axis=[1,0,0],
+            rot_angle=0,
+            mesh_base_color=(1.0, 1.0, 0.9),
+            scene_bg_color=(0,0,0),
+            render_res=[256, 256],
+            focal_length=None,
+        ):
+
+        renderer = pyrender.OffscreenRenderer(viewport_width=render_res[0],
+                                              viewport_height=render_res[1],
+                                              point_size=1.0)
+        # material = pyrender.MetallicRoughnessMaterial(
+        #     metallicFactor=0.0,
+        #     alphaMode='OPAQUE',
+        #     baseColorFactor=(*mesh_base_color, 1.0))
+
+        mesh_list = [pyrender.Mesh.from_trimesh(self.vertices_to_trimesh(vvv, ttt.copy(), mesh_base_color, rot_axis, rot_angle)) for vvv,ttt in zip(vertices, cam_t)]
+
+        scene = pyrender.Scene(bg_color=[*scene_bg_color, 0.0],
+                               ambient_light=(0.3, 0.3, 0.3))
+        for i,mesh in enumerate(mesh_list):
+            scene.add(mesh, f'mesh_{i}')
+
+        camera_pose = np.eye(4)
+        # camera_pose[:3, 3] = camera_translation
+        camera_center = [render_res[0] / 2., render_res[1] / 2.]
+        focal_length = focal_length if focal_length is not None else self.focal_length
+        camera = pyrender.IntrinsicsCamera(fx=focal_length, fy=focal_length,
+                                           cx=camera_center[0], cy=camera_center[1], zfar=1e12)
+
+        # Create camera node and add it to pyRender scene
+        camera_node = pyrender.Node(camera=camera, matrix=camera_pose)
+        scene.add_node(camera_node)
+        self.add_point_lighting(scene, camera_node)
+        self.add_lighting(scene, camera_node)
+
+        light_nodes = create_raymond_lights()
+        for node in light_nodes:
+            scene.add_node(node)
+
+        color, rend_depth = renderer.render(scene, flags=pyrender.RenderFlags.RGBA)
+        color = color.astype(np.float32) / 255.0
+        renderer.delete()
+
+        return color
+
+    def add_lighting(self, scene, cam_node, color=np.ones(3), intensity=1.0):
+        # from phalp.visualize.py_renderer import get_light_poses
+        light_poses = get_light_poses()
+        light_poses.append(np.eye(4))
+        cam_pose = scene.get_pose(cam_node)
+        for i, pose in enumerate(light_poses):
+            matrix = cam_pose @ pose
+            node = pyrender.Node(
+                name=f"light-{i:02d}",
+                light=pyrender.DirectionalLight(color=color, intensity=intensity),
+                matrix=matrix,
+            )
+            if scene.has_node(node):
+                continue
+            scene.add_node(node)
+
+    def add_point_lighting(self, scene, cam_node, color=np.ones(3), intensity=1.0):
+        # from phalp.visualize.py_renderer import get_light_poses
+        light_poses = get_light_poses(dist=0.5)
+        light_poses.append(np.eye(4))
+        cam_pose = scene.get_pose(cam_node)
+        for i, pose in enumerate(light_poses):
+            matrix = cam_pose @ pose
+            # node = pyrender.Node(
+            #     name=f"light-{i:02d}",
+            #     light=pyrender.DirectionalLight(color=color, intensity=intensity),
+            #     matrix=matrix,
+            # )
+            node = pyrender.Node(
+                name=f"plight-{i:02d}",
+                light=pyrender.PointLight(color=color, intensity=intensity),
+                matrix=matrix,
+            )
+            if scene.has_node(node):
+                continue
+            scene.add_node(node)

humans4d/hmr2/utils/rich_utils.py

@@ -0,0 +1,105 @@
+from pathlib import Path
+from typing import Sequence
+
+import rich
+import rich.syntax
+import rich.tree
+from hydra.core.hydra_config import HydraConfig
+from omegaconf import DictConfig, OmegaConf, open_dict
+from pytorch_lightning.utilities import rank_zero_only
+from rich.prompt import Prompt
+
+from . import pylogger
+
+log = pylogger.get_pylogger(__name__)
+
+
+@rank_zero_only
+def print_config_tree(
+    cfg: DictConfig,
+    print_order: Sequence[str] = (
+        "datamodule",
+        "model",
+        "callbacks",
+        "logger",
+        "trainer",
+        "paths",
+        "extras",
+    ),
+    resolve: bool = False,
+    save_to_file: bool = False,
+) -> None:
+    """Prints content of DictConfig using Rich library and its tree structure.
+
+    Args:
+        cfg (DictConfig): Configuration composed by Hydra.
+        print_order (Sequence[str], optional): Determines in what order config components are printed.
+        resolve (bool, optional): Whether to resolve reference fields of DictConfig.
+        save_to_file (bool, optional): Whether to export config to the hydra output folder.
+    """
+
+    style = "dim"
+    tree = rich.tree.Tree("CONFIG", style=style, guide_style=style)
+
+    queue = []
+
+    # add fields from `print_order` to queue
+    for field in print_order:
+        queue.append(field) if field in cfg else log.warning(
+            f"Field '{field}' not found in config. Skipping '{field}' config printing..."
+        )
+
+    # add all the other fields to queue (not specified in `print_order`)
+    for field in cfg:
+        if field not in queue:
+            queue.append(field)
+
+    # generate config tree from queue
+    for field in queue:
+        branch = tree.add(field, style=style, guide_style=style)
+
+        config_group = cfg[field]
+        if isinstance(config_group, DictConfig):
+            branch_content = OmegaConf.to_yaml(config_group, resolve=resolve)
+        else:
+            branch_content = str(config_group)
+
+        branch.add(rich.syntax.Syntax(branch_content, "yaml"))
+
+    # print config tree
+    rich.print(tree)
+
+    # save config tree to file
+    if save_to_file:
+        with open(Path(cfg.paths.output_dir, "config_tree.log"), "w") as file:
+            rich.print(tree, file=file)
+
+
+@rank_zero_only
+def enforce_tags(cfg: DictConfig, save_to_file: bool = False) -> None:
+    """Prompts user to input tags from command line if no tags are provided in config."""
+
+    if not cfg.get("tags"):
+        if "id" in HydraConfig().cfg.hydra.job:
+            raise ValueError("Specify tags before launching a multirun!")
+
+        log.warning("No tags provided in config. Prompting user to input tags...")
+        tags = Prompt.ask("Enter a list of comma separated tags", default="dev")
+        tags = [t.strip() for t in tags.split(",") if t != ""]
+
+        with open_dict(cfg):
+            cfg.tags = tags
+
+        log.info(f"Tags: {cfg.tags}")
+
+    if save_to_file:
+        with open(Path(cfg.paths.output_dir, "tags.log"), "w") as file:
+            rich.print(cfg.tags, file=file)
+
+
+if __name__ == "__main__":
+    from hydra import compose, initialize
+
+    with initialize(version_base="1.2", config_path="../../configs"):
+        cfg = compose(config_name="train.yaml", return_hydra_config=False, overrides=[])
+        print_config_tree(cfg, resolve=False, save_to_file=False)

humans4d/hmr2/utils/skeleton_renderer.py

@@ -0,0 +1,122 @@
+import torch
+import numpy as np
+import trimesh
+from typing import Optional
+from yacs.config import CfgNode
+
+from .geometry import perspective_projection
+from .render_openpose import render_openpose
+
+class SkeletonRenderer:
+
+    def __init__(self, cfg: CfgNode):
+        """
+        Object used to render 3D keypoints. Faster for use during training.
+        Args:
+            cfg (CfgNode): Model config file.
+        """
+        self.cfg = cfg
+
+    def __call__(self,
+                 pred_keypoints_3d: torch.Tensor,
+                 gt_keypoints_3d: torch.Tensor,
+                 gt_keypoints_2d: torch.Tensor,
+                 images: Optional[np.array] = None,
+                 camera_translation: Optional[torch.Tensor] = None) -> np.array:
+        """
+        Render batch of 3D keypoints.
+        Args:
+            pred_keypoints_3d (torch.Tensor): Tensor of shape (B, S, N, 3) containing a batch of predicted 3D keypoints, with S samples per image.
+            gt_keypoints_3d (torch.Tensor): Tensor of shape (B, N, 4) containing corresponding ground truth 3D keypoints; last value is the confidence.
+            gt_keypoints_2d (torch.Tensor): Tensor of shape (B, N, 3) containing corresponding ground truth 2D keypoints.
+            images (torch.Tensor): Tensor of shape (B, H, W, 3) containing images with values in the [0,255] range.
+            camera_translation (torch.Tensor): Tensor of shape (B, 3) containing the camera translation.
+        Returns:
+            np.array : Image with the following layout. Each row contains the a) input image,
+                                                                              b) image with gt 2D keypoints,
+                                                                              c) image with projected gt 3D keypoints,
+                                                                              d_1, ... , d_S) image with projected predicted 3D keypoints,
+                                                                              e) gt 3D keypoints rendered from a side view,
+                                                                              f_1, ... , f_S) predicted 3D keypoints frorm a side view
+        """
+        batch_size = pred_keypoints_3d.shape[0]
+#        num_samples = pred_keypoints_3d.shape[1]
+        pred_keypoints_3d = pred_keypoints_3d.clone().cpu().float()
+        gt_keypoints_3d = gt_keypoints_3d.clone().cpu().float()
+        gt_keypoints_3d[:, :, :-1] = gt_keypoints_3d[:, :, :-1] - gt_keypoints_3d[:, [25+14], :-1] + pred_keypoints_3d[:, [25+14]]
+        gt_keypoints_2d = gt_keypoints_2d.clone().cpu().float().numpy()
+        gt_keypoints_2d[:, :, :-1] = self.cfg.MODEL.IMAGE_SIZE * (gt_keypoints_2d[:, :, :-1] + 1.0) / 2.0
+
+        openpose_indices = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14]
+        gt_indices = [12, 8, 7, 6, 9, 10, 11, 14, 2, 1, 0, 3, 4, 5]
+        gt_indices = [25 + i for i in gt_indices]
+        keypoints_to_render = torch.ones(batch_size, gt_keypoints_3d.shape[1], 1)
+        rotation = torch.eye(3).unsqueeze(0)
+        if camera_translation is None:
+            camera_translation = torch.tensor([0.0, 0.0, 2 * self.cfg.EXTRA.FOCAL_LENGTH / (0.8 * self.cfg.MODEL.IMAGE_SIZE)]).unsqueeze(0).repeat(batch_size, 1)
+        else:
+            camera_translation = camera_translation.cpu()
+
+        if images is None:
+            images = np.zeros((batch_size, self.cfg.MODEL.IMAGE_SIZE, self.cfg.MODEL.IMAGE_SIZE, 3))
+        focal_length = torch.tensor([self.cfg.EXTRA.FOCAL_LENGTH, self.cfg.EXTRA.FOCAL_LENGTH]).reshape(1, 2)
+        camera_center = torch.tensor([self.cfg.MODEL.IMAGE_SIZE, self.cfg.MODEL.IMAGE_SIZE], dtype=torch.float).reshape(1, 2) / 2.
+        gt_keypoints_3d_proj = perspective_projection(gt_keypoints_3d[:, :, :-1], rotation=rotation.repeat(batch_size, 1, 1), translation=camera_translation[:, :], focal_length=focal_length.repeat(batch_size, 1), camera_center=camera_center.repeat(batch_size, 1))
+        pred_keypoints_3d_proj = perspective_projection(pred_keypoints_3d.reshape(batch_size, -1, 3), rotation=rotation.repeat(batch_size, 1, 1), translation=camera_translation.reshape(batch_size, -1), focal_length=focal_length.repeat(batch_size, 1), camera_center=camera_center.repeat(batch_size, 1)).reshape(batch_size, -1, 2)
+        gt_keypoints_3d_proj = torch.cat([gt_keypoints_3d_proj, gt_keypoints_3d[:, :, [-1]]], dim=-1).cpu().numpy()
+        pred_keypoints_3d_proj = torch.cat([pred_keypoints_3d_proj, keypoints_to_render.reshape(batch_size, -1, 1)], dim=-1).cpu().numpy()
+        rows = []
+        # Rotate keypoints to visualize side view
+        R = torch.tensor(trimesh.transformations.rotation_matrix(np.radians(90), [0, 1, 0])[:3, :3]).float()
+        gt_keypoints_3d_side = gt_keypoints_3d.clone()
+        gt_keypoints_3d_side[:, :, :-1] = torch.einsum('bni,ij->bnj', gt_keypoints_3d_side[:, :, :-1], R)
+        pred_keypoints_3d_side = pred_keypoints_3d.clone()
+        pred_keypoints_3d_side = torch.einsum('bni,ij->bnj', pred_keypoints_3d_side, R)
+        gt_keypoints_3d_proj_side = perspective_projection(gt_keypoints_3d_side[:, :, :-1], rotation=rotation.repeat(batch_size, 1, 1), translation=camera_translation[:, :], focal_length=focal_length.repeat(batch_size, 1), camera_center=camera_center.repeat(batch_size, 1))
+        pred_keypoints_3d_proj_side = perspective_projection(pred_keypoints_3d_side.reshape(batch_size, -1, 3), rotation=rotation.repeat(batch_size, 1, 1), translation=camera_translation.reshape(batch_size, -1), focal_length=focal_length.repeat(batch_size, 1), camera_center=camera_center.repeat(batch_size, 1)).reshape(batch_size, -1, 2)
+        gt_keypoints_3d_proj_side = torch.cat([gt_keypoints_3d_proj_side, gt_keypoints_3d_side[:, :, [-1]]], dim=-1).cpu().numpy()
+        pred_keypoints_3d_proj_side = torch.cat([pred_keypoints_3d_proj_side, keypoints_to_render.reshape(batch_size, -1, 1)], dim=-1).cpu().numpy()
+        for i in range(batch_size):
+            img = images[i]
+            side_img = np.zeros((self.cfg.MODEL.IMAGE_SIZE, self.cfg.MODEL.IMAGE_SIZE, 3))
+            # gt 2D keypoints
+            body_keypoints_2d = gt_keypoints_2d[i, :25].copy()
+            for op, gt in zip(openpose_indices, gt_indices):
+                if gt_keypoints_2d[i, gt, -1] > body_keypoints_2d[op, -1]:
+                    body_keypoints_2d[op] = gt_keypoints_2d[i, gt]
+            gt_keypoints_img = render_openpose(img, body_keypoints_2d) / 255.
+            # gt 3D keypoints
+            body_keypoints_3d_proj = gt_keypoints_3d_proj[i, :25].copy()
+            for op, gt in zip(openpose_indices, gt_indices):
+                if gt_keypoints_3d_proj[i, gt, -1] > body_keypoints_3d_proj[op, -1]:
+                    body_keypoints_3d_proj[op] = gt_keypoints_3d_proj[i, gt]
+            gt_keypoints_3d_proj_img = render_openpose(img, body_keypoints_3d_proj) / 255.
+            # gt 3D keypoints from the side
+            body_keypoints_3d_proj = gt_keypoints_3d_proj_side[i, :25].copy()
+            for op, gt in zip(openpose_indices, gt_indices):
+                if gt_keypoints_3d_proj_side[i, gt, -1] > body_keypoints_3d_proj[op, -1]:
+                    body_keypoints_3d_proj[op] = gt_keypoints_3d_proj_side[i, gt]
+            gt_keypoints_3d_proj_img_side = render_openpose(side_img, body_keypoints_3d_proj) / 255.
+            # pred 3D keypoints
+            pred_keypoints_3d_proj_imgs = []
+            body_keypoints_3d_proj = pred_keypoints_3d_proj[i, :25].copy()
+            for op, gt in zip(openpose_indices, gt_indices):
+                if pred_keypoints_3d_proj[i, gt, -1] >= body_keypoints_3d_proj[op, -1]:
+                    body_keypoints_3d_proj[op] = pred_keypoints_3d_proj[i, gt]
+            pred_keypoints_3d_proj_imgs.append(render_openpose(img, body_keypoints_3d_proj) / 255.)
+            pred_keypoints_3d_proj_img = np.concatenate(pred_keypoints_3d_proj_imgs, axis=1)
+            # gt 3D keypoints from the side
+            pred_keypoints_3d_proj_imgs_side = []
+            body_keypoints_3d_proj = pred_keypoints_3d_proj_side[i, :25].copy()
+            for op, gt in zip(openpose_indices, gt_indices):
+                if pred_keypoints_3d_proj_side[i, gt, -1] >= body_keypoints_3d_proj[op, -1]:
+                    body_keypoints_3d_proj[op] = pred_keypoints_3d_proj_side[i, gt]
+            pred_keypoints_3d_proj_imgs_side.append(render_openpose(side_img, body_keypoints_3d_proj) / 255.)
+            pred_keypoints_3d_proj_img_side = np.concatenate(pred_keypoints_3d_proj_imgs_side, axis=1)
+            rows.append(np.concatenate((gt_keypoints_img, gt_keypoints_3d_proj_img, pred_keypoints_3d_proj_img, gt_keypoints_3d_proj_img_side, pred_keypoints_3d_proj_img_side), axis=1))
+        # Concatenate images
+        img = np.concatenate(rows, axis=0)
+        img[:, ::self.cfg.MODEL.IMAGE_SIZE, :] = 1.0
+        img[::self.cfg.MODEL.IMAGE_SIZE, :, :] = 1.0
+        img[:, (1+1+1)*self.cfg.MODEL.IMAGE_SIZE, :] = 0.5
+        return img

humans4d/hmr2/utils/texture_utils.py

@@ -0,0 +1,85 @@
+import numpy as np
+import torch
+from torch.nn import functional as F
+# from psbody.mesh.visibility import visibility_compute
+
+def uv_to_xyz_and_normals(verts, f, fmap, bmap, ftov):
+    vn = estimate_vertex_normals(verts, f, ftov)
+    pixels_to_set = torch.nonzero(fmap+1)
+    x_to_set = pixels_to_set[:,0]
+    y_to_set = pixels_to_set[:,1]
+    b_coords = bmap[x_to_set, y_to_set, :]
+    f_coords = fmap[x_to_set, y_to_set]
+    v_ids = f[f_coords]
+    points = (b_coords[:,0,None]*verts[:,v_ids[:,0]]
+             + b_coords[:,1,None]*verts[:,v_ids[:,1]]
+             + b_coords[:,2,None]*verts[:,v_ids[:,2]])
+    normals = (b_coords[:,0,None]*vn[:,v_ids[:,0]]
+             + b_coords[:,1,None]*vn[:,v_ids[:,1]]
+             + b_coords[:,2,None]*vn[:,v_ids[:,2]])
+    return points, normals, vn, f_coords
+
+def estimate_vertex_normals(v, f, ftov):
+    face_normals = TriNormalsScaled(v, f)
+    non_scaled_normals = torch.einsum('ij,bjk->bik', ftov, face_normals)
+    norms = torch.sum(non_scaled_normals ** 2.0, 2) ** 0.5
+    norms[norms == 0] = 1.0
+    return torch.div(non_scaled_normals, norms[:,:,None])
+
+def TriNormalsScaled(v, f):
+    return torch.cross(_edges_for(v, f, 1, 0), _edges_for(v, f, 2, 0))
+
+def _edges_for(v, f, cplus, cminus):
+    return v[:,f[:,cplus]] - v[:,f[:,cminus]]
+
+def psbody_get_face_visibility(v, n, f, cams, normal_threshold=0.5):
+    bn, nverts, _ = v.shape
+    nfaces, _ = f.shape
+    vis_f = np.zeros([bn, nfaces], dtype='float32')
+    for i in range(bn):
+        vis, n_dot_cam = visibility_compute(v=v[i], n=n[i], f=f, cams=cams)
+        vis_v = (vis == 1) & (n_dot_cam > normal_threshold)
+        vis_f[i] = np.all(vis_v[0,f],1)
+    return vis_f
+
+def compute_uvsampler(vt, ft, tex_size=6):
+    """
+    For this mesh, pre-computes the UV coordinates for
+    F x T x T points.
+    Returns F x T x T x 2
+    """
+    uv = obj2nmr_uvmap(ft, vt, tex_size=tex_size)
+    uv = uv.reshape(-1, tex_size, tex_size, 2)
+    return uv
+
+def obj2nmr_uvmap(ft, vt, tex_size=6):
+    """
+    Converts obj uv_map to NMR uv_map (F x T x T x 2),
+    where tex_size (T) is the sample rate on each face.
+    """
+    # This is F x 3 x 2
+    uv_map_for_verts = vt[ft]
+
+    # obj's y coordinate is [1-0], but image is [0-1]
+    uv_map_for_verts[:, :, 1] = 1 - uv_map_for_verts[:, :, 1]
+
+    # range [0, 1] -> [-1, 1]
+    uv_map_for_verts = (2 * uv_map_for_verts) - 1
+
+    alpha = np.arange(tex_size, dtype=float) / (tex_size - 1)
+    beta = np.arange(tex_size, dtype=float) / (tex_size - 1)
+    import itertools
+    # Barycentric coordinate values
+    coords = np.stack([p for p in itertools.product(*[alpha, beta])])
+
+    # Compute alpha, beta (this is the same order as NMR)
+    v2 = uv_map_for_verts[:, 2]
+    v0v2 = uv_map_for_verts[:, 0] - uv_map_for_verts[:, 2]
+    v1v2 = uv_map_for_verts[:, 1] - uv_map_for_verts[:, 2]
+    # Interpolate the vertex uv values: F x 2 x T*2
+    uv_map = np.dstack([v0v2, v1v2]).dot(coords.T) + v2.reshape(-1, 2, 1)
+
+    # F x T*2 x 2  -> F x T x T x 2
+    uv_map = np.transpose(uv_map, (0, 2, 1)).reshape(-1, tex_size, tex_size, 2)
+
+    return uv_map

humans4d/hmr2/utils/utils_detectron2.py

@@ -0,0 +1,93 @@
+import detectron2.data.transforms as T
+import torch
+from detectron2.checkpoint import DetectionCheckpointer
+from detectron2.config import CfgNode, instantiate
+from detectron2.data import MetadataCatalog
+from omegaconf import OmegaConf
+
+
+class DefaultPredictor_Lazy:
+    """Create a simple end-to-end predictor with the given config that runs on single device for a
+    single input image.
+
+    Compared to using the model directly, this class does the following additions:
+
+    1. Load checkpoint from the weights specified in config (cfg.MODEL.WEIGHTS).
+    2. Always take BGR image as the input and apply format conversion internally.
+    3. Apply resizing defined by the config (`cfg.INPUT.{MIN,MAX}_SIZE_TEST`).
+    4. Take one input image and produce a single output, instead of a batch.
+
+    This is meant for simple demo purposes, so it does the above steps automatically.
+    This is not meant for benchmarks or running complicated inference logic.
+    If you'd like to do anything more complicated, please refer to its source code as
+    examples to build and use the model manually.
+
+    Attributes:
+        metadata (Metadata): the metadata of the underlying dataset, obtained from
+            test dataset name in the config.
+
+
+    Examples:
+    ::
+        pred = DefaultPredictor(cfg)
+        inputs = cv2.imread("input.jpg")
+        outputs = pred(inputs)
+    """
+
+    def __init__(self, cfg):
+        """
+        Args:
+            cfg: a yacs CfgNode or a omegaconf dict object.
+        """
+        if isinstance(cfg, CfgNode):
+            self.cfg = cfg.clone()  # cfg can be modified by model
+            self.model = build_model(self.cfg)  # noqa: F821
+            if len(cfg.DATASETS.TEST):
+                test_dataset = cfg.DATASETS.TEST[0]
+
+            checkpointer = DetectionCheckpointer(self.model)
+            checkpointer.load(cfg.MODEL.WEIGHTS)
+
+            self.aug = T.ResizeShortestEdge(
+                [cfg.INPUT.MIN_SIZE_TEST, cfg.INPUT.MIN_SIZE_TEST], cfg.INPUT.MAX_SIZE_TEST
+            )
+
+            self.input_format = cfg.INPUT.FORMAT
+        else:  # new LazyConfig
+            self.cfg = cfg
+            self.model = instantiate(cfg.model)
+            test_dataset = OmegaConf.select(cfg, "dataloader.test.dataset.names", default=None)
+            if isinstance(test_dataset, (list, tuple)):
+                test_dataset = test_dataset[0]
+
+            checkpointer = DetectionCheckpointer(self.model)
+            checkpointer.load(OmegaConf.select(cfg, "train.init_checkpoint", default=""))
+
+            mapper = instantiate(cfg.dataloader.test.mapper)
+            self.aug = mapper.augmentations
+            self.input_format = mapper.image_format
+
+        self.model.eval().cuda()
+        if test_dataset:
+            self.metadata = MetadataCatalog.get(test_dataset)
+        assert self.input_format in ["RGB", "BGR"], self.input_format
+
+    def __call__(self, original_image):
+        """
+        Args:
+            original_image (np.ndarray): an image of shape (H, W, C) (in BGR order).
+
+        Returns:
+            predictions (dict):
+                the output of the model for one image only.
+                See :doc:`/tutorials/models` for details about the format.
+        """
+        with torch.no_grad():
+            if self.input_format == "RGB":
+                original_image = original_image[:, :, ::-1]
+            height, width = original_image.shape[:2]
+            image = self.aug(T.AugInput(original_image)).apply_image(original_image)
+            image = torch.as_tensor(image.astype("float32").transpose(2, 0, 1))
+            inputs = {"image": image, "height": height, "width": width}
+            predictions = self.model([inputs])[0]
+            return predictions

humans4d/setup.py

@@ -0,0 +1,31 @@
+from setuptools import setup, find_packages
+
+print('Found packages:', find_packages())
+setup(
+    description='HMR2 as a package',
+    name='hmr2',
+    packages=find_packages(),
+    install_requires=[
+        'gdown',
+        'numpy',
+        'torch',
+        'torchvision',
+        'pytorch-lightning',
+        'smplx==0.1.28',
+        'pyrender',
+        'opencv-python',
+        'yacs',
+        'scikit-image',
+        'einops',
+        'timm',
+        'webdataset',
+        'dill',
+        'pandas',
+        'chumpy @ git+https://github.com/mattloper/chumpy',
+    ],
+    extras_require={
+        'all': [
+            'detectron2 @ git+https://github.com/facebookresearch/detectron2',
+        ],
+    },
+)

humans4d/track.py

@@ -0,0 +1,185 @@
+import warnings
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Optional, Tuple
+
+import os
+import hydra
+import torch
+import numpy as np
+from hydra.core.config_store import ConfigStore
+from omegaconf import DictConfig
+
+from phalp.configs.base import FullConfig
+from phalp.models.hmar.hmr import HMR2018Predictor
+from phalp.trackers.PHALP import PHALP
+from phalp.utils import get_pylogger
+from phalp.configs.base import CACHE_DIR
+
+from hmr2.datasets.utils import expand_bbox_to_aspect_ratio
+
+warnings.filterwarnings('ignore')
+
+log = get_pylogger(__name__)
+
+class HMR2Predictor(HMR2018Predictor):
+    def __init__(self, cfg) -> None:
+        super().__init__(cfg)
+        # Setup our new model
+        from hmr2.models import download_models, load_hmr2
+
+        # Download and load checkpoints
+        download_models()
+        model, _ = load_hmr2()
+
+        self.model = model
+        self.model.eval()
+
+    def forward(self, x):
+        hmar_out = self.hmar_old(x)
+        batch = {
+            'img': x[:,:3,:,:],
+            'mask': (x[:,3,:,:]).clip(0,1),
+        }
+        model_out = self.model(batch)
+        out = hmar_out | {
+            'pose_smpl': model_out['pred_smpl_params'],
+            'pred_cam': model_out['pred_cam'],
+        }
+        return out
+    
+class HMR2023TextureSampler(HMR2Predictor):
+    def __init__(self, cfg) -> None:
+        super().__init__(cfg)
+
+        # Model's all set up. Now, load tex_bmap and tex_fmap
+        # Texture map atlas
+        bmap = np.load(os.path.join(CACHE_DIR, 'phalp/3D/bmap_256.npy'))
+        fmap = np.load(os.path.join(CACHE_DIR, 'phalp/3D/fmap_256.npy'))
+        self.register_buffer('tex_bmap', torch.tensor(bmap, dtype=torch.float))
+        self.register_buffer('tex_fmap', torch.tensor(fmap, dtype=torch.long))
+
+        self.img_size = 256         #self.cfg.MODEL.IMAGE_SIZE
+        self.focal_length = 5000.   #self.cfg.EXTRA.FOCAL_LENGTH
+
+        import neural_renderer as nr
+        self.neural_renderer = nr.Renderer(dist_coeffs=None, orig_size=self.img_size,
+                                          image_size=self.img_size,
+                                          light_intensity_ambient=1,
+                                          light_intensity_directional=0,
+                                          anti_aliasing=False)
+
+    def forward(self, x):
+        batch = {
+            'img': x[:,:3,:,:],
+            'mask': (x[:,3,:,:]).clip(0,1),
+        }
+        model_out = self.model(batch)
+
+        # from hmr2.models.prohmr_texture import unproject_uvmap_to_mesh
+
+        def unproject_uvmap_to_mesh(bmap, fmap, verts, faces):
+            # bmap:  256,256,3
+            # fmap:  256,256
+            # verts: B,V,3
+            # faces: F,3
+            valid_mask = (fmap >= 0)
+
+            fmap_flat = fmap[valid_mask]      # N
+            bmap_flat = bmap[valid_mask,:]    # N,3
+
+            face_vids = faces[fmap_flat, :]  # N,3
+            face_verts = verts[:, face_vids, :] # B,N,3,3
+
+            bs = face_verts.shape
+            map_verts = torch.einsum('bnij,ni->bnj', face_verts, bmap_flat) # B,N,3
+
+            return map_verts, valid_mask
+
+        pred_verts = model_out['pred_vertices'] + model_out['pred_cam_t'].unsqueeze(1)
+        device = pred_verts.device
+        face_tensor = torch.tensor(self.smpl.faces.astype(np.int64), dtype=torch.long, device=device)
+        map_verts, valid_mask = unproject_uvmap_to_mesh(self.tex_bmap, self.tex_fmap, pred_verts, face_tensor) # B,N,3
+
+        # Project map_verts to image using K,R,t
+        # map_verts_view = einsum('bij,bnj->bni', R, map_verts) + t # R=I t=0
+        focal = self.focal_length / (self.img_size / 2)
+        map_verts_proj = focal * map_verts[:, :, :2] / map_verts[:, :, 2:3] # B,N,2
+        map_verts_depth = map_verts[:, :, 2] # B,N
+
+        # Render Depth. Annoying but we need to create this
+        K = torch.eye(3, device=device)
+        K[0, 0] = K[1, 1] = self.focal_length
+        K[1, 2] = K[0, 2] = self.img_size / 2  # Because the neural renderer only support squared images
+        K = K.unsqueeze(0)
+        R = torch.eye(3, device=device).unsqueeze(0)
+        t = torch.zeros(3, device=device).unsqueeze(0)
+        rend_depth = self.neural_renderer(pred_verts,
+                                        face_tensor[None].expand(pred_verts.shape[0], -1, -1).int(),
+                                        # textures=texture_atlas_rgb,
+                                        mode='depth',
+                                        K=K, R=R, t=t)
+
+        rend_depth_at_proj = torch.nn.functional.grid_sample(rend_depth[:,None,:,:], map_verts_proj[:,None,:,:]) # B,1,1,N
+        rend_depth_at_proj = rend_depth_at_proj.squeeze(1).squeeze(1) # B,N
+
+        img_rgba = torch.cat([batch['img'], batch['mask'][:,None,:,:]], dim=1) # B,4,H,W
+        img_rgba_at_proj = torch.nn.functional.grid_sample(img_rgba, map_verts_proj[:,None,:,:]) # B,4,1,N
+        img_rgba_at_proj = img_rgba_at_proj.squeeze(2) # B,4,N
+
+        visibility_mask = map_verts_depth <= (rend_depth_at_proj + 1e-4) # B,N
+        img_rgba_at_proj[:,3,:][~visibility_mask] = 0
+
+        # Paste image back onto square uv_image
+        uv_image = torch.zeros((batch['img'].shape[0], 4, 256, 256), dtype=torch.float, device=device)
+        uv_image[:, :, valid_mask] = img_rgba_at_proj
+
+        out = {
+            'uv_image':  uv_image,
+            'uv_vector' : self.hmar_old.process_uv_image(uv_image),
+            'pose_smpl': model_out['pred_smpl_params'],
+            'pred_cam':  model_out['pred_cam'],
+        }
+        return out
+
+class HMR2_4dhuman(PHALP):
+    def __init__(self, cfg):
+        super().__init__(cfg)
+
+    def setup_hmr(self):
+        self.HMAR = HMR2023TextureSampler(self.cfg)
+
+    def get_detections(self, image, frame_name, t_, additional_data=None, measurments=None):
+        (
+            pred_bbox, pred_bbox, pred_masks, pred_scores, pred_classes, 
+            ground_truth_track_id, ground_truth_annotations
+        ) =  super().get_detections(image, frame_name, t_, additional_data, measurments)
+
+        # Pad bounding boxes 
+        pred_bbox_padded = expand_bbox_to_aspect_ratio(pred_bbox, self.cfg.expand_bbox_shape)
+
+        return (
+            pred_bbox, pred_bbox_padded, pred_masks, pred_scores, pred_classes,
+            ground_truth_track_id, ground_truth_annotations
+        )
+    
+
+@dataclass
+class Human4DConfig(FullConfig):
+    # override defaults if needed
+    expand_bbox_shape: Optional[Tuple[int]] = (192,256)
+    pass
+
+cs = ConfigStore.instance()
+cs.store(name="config", node=Human4DConfig)
+
+@hydra.main(version_base="1.2", config_name="config")
+def main(cfg: DictConfig) -> Optional[float]:
+    """Main function for running the PHALP tracker."""
+
+    phalp_tracker = HMR2_4dhuman(cfg)
+
+    phalp_tracker.track()
+
+if __name__ == "__main__":
+    main()

humans4d/train.py

@@ -0,0 +1,111 @@
+from typing import Optional, Tuple
+import pyrootutils
+
+root = pyrootutils.setup_root(
+    search_from=__file__,
+    indicator=[".git", "pyproject.toml"],
+    pythonpath=True,
+    dotenv=True,
+)
+
+import os
+from pathlib import Path
+
+import hydra
+import pytorch_lightning as pl
+from omegaconf import DictConfig, OmegaConf
+from pytorch_lightning import Trainer
+from pytorch_lightning.loggers import TensorBoardLogger
+from pytorch_lightning.plugins.environments import SLURMEnvironment
+
+from yacs.config import CfgNode
+from hmr2.configs import dataset_config
+from hmr2.datasets import HMR2DataModule
+from hmr2.models.hmr2 import HMR2
+from hmr2.utils.pylogger import get_pylogger
+from hmr2.utils.misc import task_wrapper, log_hyperparameters
+
+# HACK reset the signal handling so the lightning is free to set it
+# Based on https://github.com/facebookincubator/submitit/issues/1709#issuecomment-1246758283
+import signal
+signal.signal(signal.SIGUSR1, signal.SIG_DFL)
+
+log = get_pylogger(__name__)
+
+
+@pl.utilities.rank_zero.rank_zero_only
+def save_configs(model_cfg: CfgNode, dataset_cfg: CfgNode, rootdir: str):
+    """Save config files to rootdir."""
+    Path(rootdir).mkdir(parents=True, exist_ok=True)
+    OmegaConf.save(config=model_cfg, f=os.path.join(rootdir, 'model_config.yaml'))
+    with open(os.path.join(rootdir, 'dataset_config.yaml'), 'w') as f:
+        f.write(dataset_cfg.dump())
+
+@task_wrapper
+def train(cfg: DictConfig) -> Tuple[dict, dict]:
+
+    # Load dataset config
+    dataset_cfg = dataset_config()
+
+    # Save configs
+    save_configs(cfg, dataset_cfg, cfg.paths.output_dir)
+
+    # Setup training and validation datasets
+    datamodule = HMR2DataModule(cfg, dataset_cfg)
+
+    # Setup model
+    model = HMR2(cfg)
+
+    # Setup Tensorboard logger
+    logger = TensorBoardLogger(os.path.join(cfg.paths.output_dir, 'tensorboard'), name='', version='', default_hp_metric=False)
+    loggers = [logger]
+
+    # Setup checkpoint saving
+    checkpoint_callback = pl.callbacks.ModelCheckpoint(
+        dirpath=os.path.join(cfg.paths.output_dir, 'checkpoints'), 
+        every_n_train_steps=cfg.GENERAL.CHECKPOINT_STEPS, 
+        save_last=True,
+        save_top_k=cfg.GENERAL.CHECKPOINT_SAVE_TOP_K,
+    )
+    rich_callback = pl.callbacks.RichProgressBar()
+    lr_monitor = pl.callbacks.LearningRateMonitor(logging_interval='step')
+    callbacks = [
+        checkpoint_callback, 
+        lr_monitor,
+        # rich_callback
+    ]
+
+    log.info(f"Instantiating trainer <{cfg.trainer._target_}>")
+    trainer: Trainer = hydra.utils.instantiate(
+        cfg.trainer, 
+        callbacks=callbacks, 
+        logger=loggers, 
+        plugins=(SLURMEnvironment(requeue_signal=signal.SIGUSR2) if (cfg.get('launcher',None) is not None) else None), # Submitit uses SIGUSR2
+    )
+
+    object_dict = {
+        "cfg": cfg,
+        "datamodule": datamodule,
+        "model": model,
+        "callbacks": callbacks,
+        "logger": logger,
+        "trainer": trainer,
+    }
+
+    if logger:
+        log.info("Logging hyperparameters!")
+        log_hyperparameters(object_dict)
+
+    # Train the model
+    trainer.fit(model, datamodule=datamodule, ckpt_path='last')
+    log.info("Fitting done")
+
+
+@hydra.main(version_base="1.2", config_path=str(root/"hmr2/configs_hydra"), config_name="train.yaml")
+def main(cfg: DictConfig) -> Optional[float]:
+    # train the model
+    train(cfg)
+
+
+if __name__ == "__main__":
+    main()

nodes/load_detectron_node.py

@@ -0,0 +1,27 @@
+import os
+
+from folder_paths import models_dir
+
+from ..utils.process_humans import load_detectron
+
+
+
+DETECTRON_PATH = os.path.join(models_dir, 'detectron')
+os.makedirs(DETECTRON_PATH, exist_ok=True)
+
+
+
+class LoadDetectronNode:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": { "checkpoint": (os.listdir(DETECTRON_PATH), )}}
+
+    RETURN_TYPES = ("DETECTRON",)
+    FUNCTION = "load"
+
+    CATEGORY = "4dhumans"
+
+    def load(self, checkpoint):
+        checkpoint_path = os.path.join(DETECTRON_PATH, checkpoint)
+        model = load_detectron(checkpoint_path)
+        return (model,)

nodes/load_hmr_node.py

@@ -0,0 +1,30 @@
+import os
+
+import torch
+
+from folder_paths import models_dir
+
+from ..utils.process_humans import load_hmr_model
+
+
+HMR_PATH = os.path.join(models_dir, 'hmr')
+os.makedirs(HMR_PATH, exist_ok=True)
+
+SMPL_PATH = os.path.join(models_dir, 'smpl')
+os.makedirs(SMPL_PATH, exist_ok=True)
+
+
+class LoadHMRNode:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": { "checkpoint": (os.listdir(HMR_PATH), )}}
+
+    RETURN_TYPES = ("HMR",)
+    FUNCTION = "load"
+
+    CATEGORY = "4dhumans"
+
+    def load(self, checkpoint):
+        checkpoint_path = os.path.join(HMR_PATH, checkpoint)
+        model_info = load_hmr_model(checkpoint_path, SMPL_PATH)
+        return (model_info,)

nodes/process_humans_node.py

@@ -0,0 +1,27 @@
+from folder_paths import models_dir
+import comfy.model_management
+
+from ..utils.process_humans import process_humans
+
+
+class ProcessHumansNode:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": {
+            "images": ("IMAGE",),
+            "hmr": ("HMR", ),
+            "detectron": ("DETECTRON", ),
+        }}
+    RETURN_TYPES = ("SMPL", "IMAGE", "MASK", "IMAGE")
+    RETURN_NAMES = ("smpl", "images", "masks", "semantic_maps")
+    FUNCTION = "process"
+
+    CATEGORY = "4dhumans"
+
+    def process(self, images, hmr, detectron):
+        intermediate_device = comfy.model_management.intermediate_device()
+        torch_device = comfy.model_management.get_torch_device()
+
+        results = process_humans(images, hmr, detectron, figure_scale=None, device=torch_device)
+
+        return (results['smpl'], results['images'], results['masks'], results['semantic_maps'])

nodes/select_human_node.py

@@ -0,0 +1,21 @@
+from ..utils.process_humans import get_human
+
+
+class SelectHumanNode:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": {
+            "images": ("IMAGE",),
+            "smpl": ("SMPL", ),
+            "idx": ("INT", {"default": 0, "min": 0, "max": 999, "step": 1}),
+        }}
+    RETURN_TYPES = ("SMPL", "IMAGE", "MASK", "IMAGE")
+    RETURN_NAMES = ("smpl", "images", "masks", "semantic_maps")
+    FUNCTION = "process"
+
+    CATEGORY = "4dhumans"
+
+    def process(self, images, smpl, idx):
+        results = get_human(images, smpl, idx)
+
+        return (results['smpl'], results['images'], results['masks'], results['semantic_maps'])

requirements.txt

@@ -0,0 +1,4 @@
+# git+https://github.com/shubham-goel/4D-Humans
+git+https://github.com/facebookresearch/detectron2
+pyrender
+yacs

utils/process_humans.py

@@ -0,0 +1,298 @@
+import os
+import torch
+import numpy as np
+import platform
+import pyrender
+# from pathlib import Path
+from tqdm import tqdm
+
+
+if "PYOPENGL_PLATFORM" not in os.environ:
+    os.environ["PYOPENGL_PLATFORM"] = "egl"
+
+import comfy.utils
+
+from ..humans4d.hmr2.models import load_hmr2
+from ..humans4d.hmr2.utils import recursive_to
+from ..humans4d.hmr2.datasets.vitdet_dataset import ViTDetDataset
+from ..humans4d.hmr2.utils.renderer import cam_crop_to_full
+# from ..humans4d.hmr2.configs import get_config
+# from ..humans4d.hmr2.models import check_smpl_exists
+from ..humans4d.hmr2.utils.utils_detectron2 import DefaultPredictor_Lazy
+from detectron2.config import LazyConfig
+
+from .visualize_humans import SemanticRenderer
+from .. import REPO_PATH
+
+# For Windows, remove PYOPENGL_PLATFORM to enable default rendering backend
+sys_name = platform.system()
+if sys_name == "Windows":
+    os.environ.pop("PYOPENGL_PLATFORM")
+
+
+LIGHT_BLUE = (0.65098039, 0.74117647, 0.85882353)
+
+
+def predict_smpl(batch, model, model_cfg, figure_scale=None):
+    all_verts = []
+    all_cam_t = []
+    with torch.no_grad():
+        out = model(batch)
+
+    pred_cam = out["pred_cam"]
+    pred_smpl_parameter = out["pred_smpl_params"]
+    if figure_scale is not None:
+        pred_smpl_parameter['betas'][0][1] = float(figure_scale)
+    smpl_output = model.smpl(
+        **{k: v.float() for k, v in pred_smpl_parameter.items()},
+        pose2rot=False,
+    )
+    pred_vertices = smpl_output.vertices
+    out["pred_vertices"] = pred_vertices.reshape(
+        batch["img"].shape[0], -1, 3
+    )
+
+    box_center = batch["box_center"].float()
+    box_size = batch["box_size"].float()
+    img_size = batch["img_size"].float()
+
+    scaled_focal_length = (
+        model_cfg.EXTRA.FOCAL_LENGTH
+        / model_cfg.MODEL.IMAGE_SIZE
+        * img_size.max()
+    )
+    pred_cam_t_full = cam_crop_to_full(
+                pred_cam, box_center, box_size, img_size, scaled_focal_length
+            ).detach().cpu().numpy()
+    # Render the result
+    batch_size = batch["img"].shape[0]
+    for n in range(batch_size):
+        # Add all verts and cams to list
+        verts = out["pred_vertices"][n].detach().cpu().numpy()
+        cam_t = pred_cam_t_full[n]
+        all_verts.append(verts)
+        all_cam_t.append(cam_t)
+
+    misc_args = dict(
+        mesh_base_color=LIGHT_BLUE,
+        scene_bg_color=(1, 1, 1),
+        focal_length=scaled_focal_length,
+    )
+    
+    smpl_outs = {
+        k: v.detach().cpu().numpy() for k, v in pred_smpl_parameter.items()
+    }
+
+    results_dict_for_rendering = {
+        "verts": all_verts,
+        "cam_t": all_cam_t,
+        "render_res": img_size[n].cpu().numpy(),
+        "smpls": smpl_outs,
+        "scaled_focal_length": scaled_focal_length.cpu().numpy(),
+    }
+    return results_dict_for_rendering, misc_args
+
+
+def load_image(model_cfg, image, detector):
+    # Detect humans in image
+    # image = image.permute(1,2,0)*255
+    det_out = detector(image)
+    # image = image.numpy()
+    det_instances = det_out["instances"]
+    valid_idx = (det_instances.pred_classes == 0) & (det_instances.scores > 0.5)
+    boxes = det_instances.pred_boxes.tensor[valid_idx].cpu().numpy()
+    # Run HMR2.0 on all detected humans
+    dataset = ViTDetDataset(model_cfg, image, boxes)
+    return torch.utils.data.DataLoader(
+        dataset, batch_size=8, shuffle=False, num_workers=0
+    )
+
+def load_hmr_model(hmr2_checkpoint_path, smpl_dir_path):
+    model, model_cfg = load_hmr2(hmr2_checkpoint_path,smpl_dir_path)
+    return {
+        'model': model,
+        'model_cfg': model_cfg
+    }
+
+
+def load_detectron(detectron_checkpoint_path):
+    cfg_path = os.path.join(REPO_PATH, "configs", "cascade_mask_rcnn_vitdet_h_75ep.py")
+    detectron2_cfg = LazyConfig.load(str(cfg_path))
+    detectron2_cfg.train.init_checkpoint = detectron_checkpoint_path
+    for i in range(3):
+        detectron2_cfg.model.roi_heads.box_predictors[i].test_score_thresh = 0.25
+    detector = DefaultPredictor_Lazy(detectron2_cfg)
+    return detector
+
+
+def process_humans(images, hmr, detector, figure_scale, device):
+    images = images.permute(0, 3, 1, 2)
+    _, _, height, width = images.shape
+    model, model_cfg = hmr['model'], hmr['model_cfg']
+
+
+    model = model.to(device)
+    detector.model.to(device)
+    # This PyRender is only used for visualizing, we use Blender after to render different conditions.
+    renderer = SemanticRenderer(
+        model_cfg,
+        faces=model.smpl.faces,
+        lbs=model.smpl.lbs_weights,
+        viewport_size=(720, 720),
+    )
+    results = []
+    rendered_images = []
+    semantic_maps = []
+    masks = []
+    focal_length = None
+    mesh_base_color = None
+    scene_bg_color = None
+    pbar = comfy.utils.ProgressBar(len(images))
+    for image in tqdm(images, desc="Processing Human Images:"):
+        image = image.permute(1,2,0)*255
+        image = image.numpy()
+        renderer.renderer.delete()
+        renderer.renderer = pyrender.OffscreenRenderer(
+            viewport_width=width,
+            viewport_height=height,
+            point_size=1.0,
+        )
+        dataloader = load_image(model_cfg, image, detector)
+        
+
+        for batch in dataloader:
+            batch = recursive_to(batch, device)
+            results_dict_for_rendering, misc_args = predict_smpl(batch, model, model_cfg, figure_scale)
+            mesh_base_color = misc_args['mesh_base_color']
+            scene_bg_color = misc_args['scene_bg_color']
+            focal_length = misc_args['focal_length'].to('cpu')
+            
+            rendering_results = renderer.render_all_multiple(
+                results_dict_for_rendering["verts"], 
+                cam_t=results_dict_for_rendering["cam_t"], 
+                render_res=results_dict_for_rendering["render_res"], **misc_args
+            )
+            # Overlay image
+            valid_mask = rendering_results["Image"][:, :, -1][:, :, np.newaxis]
+            cam_view = (
+                valid_mask * rendering_results["Image"][:, :, [2, 1, 0]]
+                + (1 - valid_mask) * image.astype(np.float32)[:, :, ::-1] / 255
+            )
+            rendered_images.append(torch.from_numpy(cam_view[:, :, ::-1].copy()))
+            semantic_maps.append(torch.from_numpy(rendering_results.get("SemanticMap")))
+            masks.append(torch.from_numpy(rendering_results.get("Mask")[:, :, 0]))
+            results.append(results_dict_for_rendering) # still in np
+        pbar.update(1)
+
+    smpl = {
+        'results': results,
+        'width': width,
+        'height': height,
+        'model_cfg': model_cfg,
+        'faces': model.smpl.faces,
+        'lbs': model.smpl.lbs_weights,
+        'viewport_size': (720, 720),
+        'misc_args': { 'mesh_base_color': mesh_base_color, 'scene_bg_color': scene_bg_color, 'focal_length': focal_length }
+    }
+    
+    return {
+        'smpl': smpl,
+        'images': torch.stack(rendered_images),
+        'masks': torch.stack(masks).to(torch.float16),
+        'semantic_maps': torch.stack(semantic_maps),
+    }
+
+
+def get_human(images, smpl, idx):
+    results = []
+    rendered_images = []
+    semantic_maps = []
+    masks = []
+
+    model_cfg, faces, lbs = smpl['model_cfg'], smpl['faces'], smpl['lbs']
+    viewport_size = smpl['viewport_size']
+    width, height = smpl['width'], smpl['height']
+    misc_args = smpl['misc_args']
+
+    renderer = SemanticRenderer(
+        model_cfg,
+        faces=faces,
+        lbs=lbs,
+        viewport_size=viewport_size,
+    )
+
+    renderer.renderer = pyrender.OffscreenRenderer(
+        viewport_width=width,
+        viewport_height=height,
+        point_size=1.0,
+    )
+
+    human_smpl = []
+    keys = ['verts']
+    smpl_keys = ['global_orient', 'body_pose', 'betas']
+    for results_dict_for_rendering in smpl['results']:
+        result = { 
+            'scaled_focal_length': results_dict_for_rendering['scaled_focal_length'], 
+            'smpls': {},
+            'cam_t': results_dict_for_rendering['cam_t'],
+            'render_res': results_dict_for_rendering['render_res']
+        }
+        good = True
+        for key in keys:
+            if idx < len(results_dict_for_rendering[key]):
+                result[key] = [results_dict_for_rendering[key][idx]]
+            else:
+                good = False
+        
+        for key in smpl_keys:
+            if idx < len(results_dict_for_rendering['smpls'][key]):
+                result['smpls'][key] = [results_dict_for_rendering['smpls'][key][idx]]
+            else:
+                good = False
+        
+        if good:
+            human_smpl.append(result)
+        else:
+            human_smpl.append(None)
+
+    for image_orig, results_dict_for_rendering in zip(images, human_smpl):
+        image = image_orig.permute(1,2,0)*255
+        image = image.numpy()
+        if results_dict_for_rendering is not None:
+            rendering_results = renderer.render_all_multiple(
+                results_dict_for_rendering["verts"], 
+                cam_t=results_dict_for_rendering["cam_t"], 
+                render_res=results_dict_for_rendering["render_res"], **misc_args
+            )
+            # Overlay image
+            valid_mask = rendering_results["Image"][:, :, -1][:, :, np.newaxis]
+            cam_view = (
+                valid_mask * rendering_results["Image"][:, :, [2, 1, 0]]
+                + (1 - valid_mask) * image.astype(np.float32)[:, :, ::-1] / 255
+            )
+            rendered_images.append(torch.from_numpy(cam_view[:, :, ::-1].copy()))
+            semantic_maps.append(torch.from_numpy(rendering_results.get("SemanticMap")))
+            masks.append(torch.from_numpy(rendering_results.get("Mask")[:, :, 0]))
+        else:
+            rendered_images.append(image_orig)
+            semantic_maps.append(torch.ones_like(image_orig))
+            masks.append(torch.zeros_like(image_orig))
+        results.append(results_dict_for_rendering) # still in np
+
+    smpl = {
+        'results': results,
+        'width': width,
+        'height': height,
+        'model_cfg': model_cfg,
+        'faces': faces,
+        'lbs': lbs,
+        'viewport_size': (720, 720),
+        'misc_args': misc_args
+    }
+    
+    return {
+        'smpl': smpl,
+        'images': torch.stack(rendered_images),
+        'masks': torch.stack(masks).to(torch.float16),
+        'semantic_maps': torch.stack(semantic_maps),
+    }

utils/visualize_humans.py

@@ -0,0 +1,294 @@
+import os
+import platform
+
+if "PYOPENGL_PLATFORM" not in os.environ:
+    os.environ["PYOPENGL_PLATFORM"] = "egl"
+import torch
+import numpy as np
+import pyrender
+import trimesh
+import cv2
+from yacs.config import CfgNode
+from typing import List, Optional
+from ..humans4d.hmr2.utils.renderer import Renderer
+
+# For Windows, remove PYOPENGL_PLATFORM to enable default rendering backend
+sys_name = platform.system()
+if sys_name == "Windows":
+    os.environ.pop("PYOPENGL_PLATFORM")
+
+def get_light_poses(n_lights=5, elevation=np.pi / 3, dist=12):
+    # get lights in a circle around origin at elevation
+    thetas = elevation * np.ones(n_lights)
+    phis = 2 * np.pi * np.arange(n_lights) / n_lights
+    poses = []
+    trans = make_translation(torch.tensor([0, 0, dist]))
+    for phi, theta in zip(phis, thetas):
+        rot = make_rotation(rx=-theta, ry=phi, order="xyz")
+        poses.append((rot @ trans).numpy())
+    return poses
+
+def make_translation(t):
+    return make_4x4_pose(torch.eye(3), t)
+
+def make_rotation(rx=0, ry=0, rz=0, order="xyz"):
+    Rx = rotx(rx)
+    Ry = roty(ry)
+    Rz = rotz(rz)
+    if order == "xyz":
+        R = Rz @ Ry @ Rx
+    elif order == "xzy":
+        R = Ry @ Rz @ Rx
+    elif order == "yxz":
+        R = Rz @ Rx @ Ry
+    elif order == "yzx":
+        R = Rx @ Rz @ Ry
+    elif order == "zyx":
+        R = Rx @ Ry @ Rz
+    elif order == "zxy":
+        R = Ry @ Rx @ Rz
+    return make_4x4_pose(R, torch.zeros(3))
+
+def make_4x4_pose(R, t):
+    """
+    :param R (*, 3, 3)
+    :param t (*, 3)
+    return (*, 4, 4)
+    """
+    dims = R.shape[:-2]
+    pose_3x4 = torch.cat([R, t.view(*dims, 3, 1)], dim=-1)
+    bottom = (
+        torch.tensor([0, 0, 0, 1], device=R.device)
+        .reshape(*(1,) * len(dims), 1, 4)
+        .expand(*dims, 1, 4)
+    )
+    return torch.cat([pose_3x4, bottom], dim=-2)
+
+def rotx(theta):
+    return torch.tensor(
+        [
+            [1, 0, 0],
+            [0, np.cos(theta), -np.sin(theta)],
+            [0, np.sin(theta), np.cos(theta)],
+        ],
+        dtype=torch.float32,
+    )
+
+def roty(theta):
+    return torch.tensor(
+        [
+            [np.cos(theta), 0, np.sin(theta)],
+            [0, 1, 0],
+            [-np.sin(theta), 0, np.cos(theta)],
+        ],
+        dtype=torch.float32,
+    )
+
+def rotz(theta):
+    return torch.tensor(
+        [
+            [np.cos(theta), -np.sin(theta), 0],
+            [np.sin(theta), np.cos(theta), 0],
+            [0, 0, 1],
+        ],
+        dtype=torch.float32,
+    )
+
+
+def create_raymond_lights() -> List[pyrender.Node]:
+    """
+    Return raymond light nodes for the scene.
+    """
+    thetas = np.pi * np.array([1.0 / 6.0, 1.0 / 6.0, 1.0 / 6.0])
+    phis = np.pi * np.array([0.0, 2.0 / 3.0, 4.0 / 3.0])
+
+    nodes = []
+
+    for phi, theta in zip(phis, thetas):
+        xp = np.sin(theta) * np.cos(phi)
+        yp = np.sin(theta) * np.sin(phi)
+        zp = np.cos(theta)
+
+        z = np.array([xp, yp, zp])
+        z = z / np.linalg.norm(z)
+        x = np.array([-z[1], z[0], 0.0])
+        if np.linalg.norm(x) == 0:
+            x = np.array([1.0, 0.0, 0.0])
+        x = x / np.linalg.norm(x)
+        y = np.cross(z, x)
+
+        matrix = np.eye(4)
+        matrix[:3, :3] = np.c_[x, y, z]
+        nodes.append(
+            pyrender.Node(
+                light=pyrender.DirectionalLight(color=np.ones(3), intensity=1.0),
+                matrix=matrix,
+            )
+        )
+
+    return nodes
+
+class SemanticRenderer(Renderer):
+
+    def __init__(
+        self, cfg: CfgNode, faces: np.array, lbs: np.array, viewport_size=(768, 768)
+    ):
+        """
+        Wrapper around the pyrender renderer to render SMPL meshes's semantic map for Champ.
+        Args:
+            cfg (CfgNode): Model config file.
+            faces (np.array): Array of shape (F, 3) containing the mesh faces.
+        """
+        self.cfg = cfg
+        self.focal_length = cfg.EXTRA.FOCAL_LENGTH
+        self.img_res = cfg.MODEL.IMAGE_SIZE
+
+        self.camera_center = [self.img_res // 2, self.img_res // 2]
+        self.faces = faces
+        self.lbs = lbs
+        # self.joint_names = smplx.joint_names.JOINT_NAMES
+
+        self.vertex_labels = np.argmax(self.lbs.cpu().numpy(), axis=1)
+        vertex_colors = cv2.applyColorMap(
+            (10 * self.vertex_labels).astype(np.uint8), cv2.COLORMAP_VIRIDIS
+        )
+        semantic_background_rgb = cv2.applyColorMap(np.uint8([0]), cv2.COLORMAP_VIRIDIS)
+        self.vertex_colors = np.squeeze(vertex_colors, axis=1)
+        self.semantic_background_rgb = (
+            np.squeeze(semantic_background_rgb.astype(np.float32), axis=1) / 255
+        )
+        self.renderer = pyrender.OffscreenRenderer(
+            viewport_width=viewport_size[0],
+            viewport_height=viewport_size[1],
+            point_size=1.0,
+        )
+
+    def vertices_to_trimesh(
+        self,
+        vertices,
+        camera_translation,
+        mesh_base_color=(1.0, 1.0, 0.9),
+        rot_axis=[1, 0, 0],
+        rot_angle=0,
+    ):
+        vertex_colors = np.array([(*mesh_base_color, 1.0)] * vertices.shape[0])
+        mesh = trimesh.Trimesh(
+            vertices.copy() + camera_translation,
+            self.faces.copy(),
+            vertex_colors=vertex_colors,
+        )
+        # mesh = trimesh.Trimesh(vertices.copy(), self.faces.copy())
+
+        rot = trimesh.transformations.rotation_matrix(np.radians(rot_angle), rot_axis)
+        mesh.apply_transform(rot)
+
+        rot = trimesh.transformations.rotation_matrix(np.radians(180), [1, 0, 0])
+        mesh.apply_transform(rot)
+        return mesh
+
+    def render_all_multiple(
+        self,
+        vertices: List[np.array],
+        cam_t: List[np.array],
+        rot_axis=[1, 0, 0],
+        rot_angle=0,
+        mesh_base_color=(1.0, 1.0, 0.9),
+        scene_bg_color=(0, 0, 0),
+        render_res=[256, 256],
+        focal_length=None,
+    ):
+
+        renderer = self.renderer
+        trimesh_list = [
+            self.vertices_to_trimesh(
+                vvv, ttt.copy(), mesh_base_color, rot_axis, rot_angle
+            )
+            for vvv, ttt in zip(vertices, cam_t)
+        ]
+        for trimesh in trimesh_list:
+            trimesh.visual.vertex_colors = self.vertex_colors
+        mesh_list = [pyrender.Mesh.from_trimesh(trimesh) for trimesh in trimesh_list]
+
+        scene = pyrender.Scene(
+            bg_color=[*scene_bg_color, 0.0], ambient_light=(0.3, 0.3, 0.3)
+        )
+        for i, mesh in enumerate(mesh_list):
+            scene.add(mesh, f"mesh_{i}")
+
+        camera_pose = np.eye(4)
+        camera_center = [render_res[0] / 2.0, render_res[1] / 2.0]
+        focal_length = focal_length if focal_length is not None else self.focal_length
+        camera = pyrender.IntrinsicsCamera(
+            fx=focal_length,
+            fy=focal_length,
+            cx=camera_center[0],
+            cy=camera_center[1],
+            zfar=1e12,
+        )
+
+        # Create camera node and add it to pyRender scene
+        camera_node = pyrender.Node(camera=camera, matrix=camera_pose)
+        scene.add_node(camera_node)
+        self.add_point_lighting(scene, camera_node)
+        self.add_lighting(scene, camera_node)
+
+        light_nodes = create_raymond_lights()
+        for node in light_nodes:
+            scene.add_node(node)
+
+        color, rend_depth = renderer.render(
+            scene, flags=pyrender.RenderFlags.FLAT | pyrender.RenderFlags.RGBA
+        )
+        color = color.astype(np.float32) / 255.0
+        # renderer.delete()
+
+        valid_mask = (color[:, :, -1])[:, :, np.newaxis]
+        semantic_map = (
+            color[:, :, :3] * valid_mask
+            + (1 - valid_mask) * self.semantic_background_rgb
+        )
+        semantic_map = semantic_map.astype(np.float32)
+
+        return {
+            "Image": color,
+            "Mask": valid_mask.astype(bool),
+            "SemanticMap": semantic_map,
+            "Depth": rend_depth,
+        }
+
+    def add_lighting(self, scene, cam_node, color=np.ones(3), intensity=1.0):
+        # from phalp.visualize.py_renderer import get_light_poses
+        light_poses = get_light_poses()
+        light_poses.append(np.eye(4))
+        cam_pose = scene.get_pose(cam_node)
+        for i, pose in enumerate(light_poses):
+            matrix = cam_pose @ pose
+            node = pyrender.Node(
+                name=f"light-{i:02d}",
+                light=pyrender.DirectionalLight(color=color, intensity=intensity),
+                matrix=matrix,
+            )
+            if scene.has_node(node):
+                continue
+            scene.add_node(node)
+
+    def add_point_lighting(self, scene, cam_node, color=np.ones(3), intensity=1.0):
+        # from phalp.visualize.py_renderer import get_light_poses
+        light_poses = get_light_poses(dist=0.5)
+        light_poses.append(np.eye(4))
+        cam_pose = scene.get_pose(cam_node)
+        for i, pose in enumerate(light_poses):
+            matrix = cam_pose @ pose
+            # node = pyrender.Node(
+            #     name=f"light-{i:02d}",
+            #     light=pyrender.DirectionalLight(color=color, intensity=intensity),
+            #     matrix=matrix,
+            # )
+            node = pyrender.Node(
+                name=f"plight-{i:02d}",
+                light=pyrender.PointLight(color=color, intensity=intensity),
+                matrix=matrix,
+            )
+            if scene.has_node(node):
+                continue
+            scene.add_node(node)