nodes/modelscope-scepter
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

new project from v0.0.1

Browse Source
This commit is contained in:
duanmurs@163.com
2023-12-28 18:23:39 +08:00
parent fda66e7ca6
commit 66f979f7e9
204 changed files with 37941 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
scepter/methods/examples/classification/example.yaml
+260
View File
@@ -0,0 +1,260 @@
ENV:
USE_PL: False
# SET GLOBAL SYSTEM
SOLVER:
# NAME DESCRIPTION: TYPE: default: 'TrainValSolver'
NAME: TrainValSolver
# RESUME_FROM DESCRIPTION: Resume from some state of training! TYPE: str default: ''
RESUME_FROM:
# MAX_EPOCHS DESCRIPTION: Max epochs for training. TYPE: int default: 10
MAX_EPOCHS: 200
# NUM_FOLDS DESCRIPTION: Num folds for training. TYPE: int default: 0
NUM_FOLDS: 1
# WORK_DIR DESCRIPTION: Save dir of the training log or model. TYPE: str default: ''
WORK_DIR: ./exp12/
LOG_FILE: std_log.txt
# EVAL_INTERVAL DESCRIPTION: Eval the model interval. TYPE: int default: 1
EVAL_INTERVAL: 1
ACCU_STEP: 1
# DO_FINAL_EVAL DESCRIPTION: If do final evaluation or not. TYPE: bool default: False
DO_FINAL_EVAL: True
# SAVE_EVAL_DATA DESCRIPTION: If save the evaluation data or not. TYPE: bool default: False
SAVE_EVAL_DATA: True
# EXTRA_KEYS DESCRIPTION: The extra keys for metric. TYPE: list default: []
EXTRA_KEYS: []
# TRAIN_DATA DESCRIPTION: Train data config. TYPE: default: ''
TRAIN_DATA:
# NAME DESCRIPTION: TYPE: default: 'ImageClassifyPublicDataset'
NAME: ImageClassifyPublicDataset
# DATASET DESCRIPTION: the public dataset name TYPE: str default: 'cifar10'
DATASET: cifar10
# DATA_ROOT DESCRIPTION: the download data save path TYPE: str default: ''
DATA_ROOT: cifar10
# MODE DESCRIPTION: test TYPE: str default: test
MODE: train
# PIN_MEMORY DESCRIPTION: pin_memory for data loader TYPE: bool default: False
PIN_MEMORY: True
# BATCH_SIZE DESCRIPTION: batch size for data TYPE: int default: 4
BATCH_SIZE: 96
# NUM_WORKERS DESCRIPTION: num workers for fetching data! TYPE: int default: 1
NUM_WORKERS: 4
# TRANSFORMS DESCRIPTION: TYPE: default:
TRANSFORMS:
# - DESCRIPTION: TYPE: default:
- # NAME DESCRIPTION: TYPE: default: 'RandomResizedCrop'
NAME: RandomResizedCrop
SIZE: 32
# RATIO DESCRIPTION: ratio TYPE: list default: [0.75, 1.3333333333333333]
RATIO: [0.75, 1.33]
# SCALE DESCRIPTION: scale TYPE: list default: [0.08, 1.0]
SCALE: [0.8, 1.0]
# INTERPOLATION DESCRIPTION: interpolation TYPE: str default: 'blilinear'
INTERPOLATION: bilinear
# INPUT_KEY DESCRIPTION: input key TYPE: str default: 'img'
INPUT_KEY: img
# OUTPUT_KEY DESCRIPTION: input key TYPE: str default: 'img'
OUTPUT_KEY: img
# BACKEND DESCRIPTION: backend, choose from pillow, cv2, torchvision TYPE: str default: 'pillow'
BACKEND: pillow
- # NAME DESCRIPTION: TYPE: default: 'RandomHorizontalFlip'
NAME: RandomHorizontalFlip
# P DESCRIPTION: P TYPE: float default: 0.5
P: 0.5
# INPUT_KEY DESCRIPTION: input key TYPE: str default: 'img'
INPUT_KEY: img
# OUTPUT_KEY DESCRIPTION: input key TYPE: str default: 'img'
OUTPUT_KEY: img
# BACKEND DESCRIPTION: backend, choose from pillow, cv2, torchvision TYPE: str default: 'pillow'
BACKEND: pillow
- # NAME DESCRIPTION: TYPE: default: 'ImageToTensor'
NAME: ImageToTensor
# INPUT_KEY DESCRIPTION: input key TYPE: str default: 'img'
INPUT_KEY: img
# OUTPUT_KEY DESCRIPTION: input key TYPE: str default: 'img'
OUTPUT_KEY: img
# BACKEND DESCRIPTION: backend, choose from pillow, cv2, torchvision TYPE: str default: 'pillow'
BACKEND: pillow
- # NAME DESCRIPTION: TYPE: default: 'Normalize'
NAME: Normalize
# MEAN DESCRIPTION: mean TYPE: list default: []
MEAN: [0.4914, 0.4822, 0.4465]
# STD DESCRIPTION: std TYPE: list default: []
STD: [0.2023, 0.1994, 0.2010]
# INPUT_KEY DESCRIPTION: input key TYPE: str default: 'img'
INPUT_KEY: img
# OUTPUT_KEY DESCRIPTION: input key TYPE: str default: 'img'
OUTPUT_KEY: img
# BACKEND DESCRIPTION: backend, choose from pillow, cv2, torchvision TYPE: str default: 'pillow'
BACKEND: pillow
- NAME: ToTensor
# KEYS DESCRIPTION: keys TYPE: list default: []
KEYS: ["img", "label"]
- # NAME DESCRIPTION: TYPE: default: 'Select'
NAME: Select
# KEYS DESCRIPTION: keys TYPE: list default: []
KEYS: ["img", "label"]
# META_KEYS DESCRIPTION: meta keys TYPE: list default: []
META_KEYS: []
# EVAL_DATA DESCRIPTION: Eval data config. TYPE: default: ''
EVAL_DATA:
# NAME DESCRIPTION: TYPE: default: 'ImageClassifyPublicDataset'
NAME: ImageClassifyPublicDataset
# DATASET DESCRIPTION: the public dataset name TYPE: str default: 'cifar10'
DATASET: cifar10
# DATA_ROOT DESCRIPTION: the download data save path TYPE: str default: ''
DATA_ROOT: ./local_data/cifar10
# MODE DESCRIPTION: test TYPE: str default: test
MODE: test
# PIN_MEMORY DESCRIPTION: pin_memory for data loader TYPE: bool default: False
PIN_MEMORY: True
# BATCH_SIZE DESCRIPTION: batch size for data TYPE: int default: 4
BATCH_SIZE: 96
# NUM_WORKERS DESCRIPTION: num workers for fetching data! TYPE: int default: 1
NUM_WORKERS: 4
# TRANSFORMS DESCRIPTION: TYPE: default:
TRANSFORMS:
# - DESCRIPTION: TYPE: default:
- # NAME DESCRIPTION: TYPE: default: 'Resize'
NAME: Resize
SIZE: 32
# INTERPOLATION DESCRIPTION: interpolation TYPE: str default: 'blilinear'
INTERPOLATION: bilinear
# INPUT_KEY DESCRIPTION: input key TYPE: str default: 'img'
INPUT_KEY: img
# OUTPUT_KEY DESCRIPTION: input key TYPE: str default: 'img'
OUTPUT_KEY: img
# BACKEND DESCRIPTION: backend, choose from pillow, cv2, torchvision TYPE: str default: 'pillow'
BACKEND: pillow
- # NAME DESCRIPTION: TYPE: default: 'ImageToTensor'
NAME: ImageToTensor
# INPUT_KEY DESCRIPTION: input key TYPE: str default: 'img'
INPUT_KEY: img
# OUTPUT_KEY DESCRIPTION: input key TYPE: str default: 'img'
OUTPUT_KEY: img
# BACKEND DESCRIPTION: backend, choose from pillow, cv2, torchvision TYPE: str default: 'pillow'
BACKEND: pillow
- # NAME DESCRIPTION: TYPE: default: 'Normalize'
NAME: Normalize
# MEAN DESCRIPTION: mean TYPE: list default: []
MEAN: [0.4914, 0.4822, 0.4465]
# STD DESCRIPTION: std TYPE: list default: []
STD: [0.2023, 0.1994, 0.2010]
# INPUT_KEY DESCRIPTION: input key TYPE: str default: 'img'
INPUT_KEY: img
# OUTPUT_KEY DESCRIPTION: input key TYPE: str default: 'img'
OUTPUT_KEY: img
# BACKEND DESCRIPTION: backend, choose from pillow, cv2, torchvision TYPE: str default: 'pillow'
BACKEND: pillow
- NAME: ToTensor
# KEYS DESCRIPTION: keys TYPE: list default: []
KEYS: ["img", "label"]
- # NAME DESCRIPTION: TYPE: default: 'Select'
NAME: Select
# KEYS DESCRIPTION: keys TYPE: list default: []
KEYS: ["img", "label"]
# META_KEYS DESCRIPTION: meta keys TYPE: list default: []
META_KEYS: []
# TRAIN_HOOKS DESCRIPTION: TYPE: default: ''
TRAIN_HOOKS:
- # NAME DESCRIPTION: TYPE: default: 'LogHook'
NAME: LogHook
# LOG_INTERVAL DESCRIPTION: the interval for log print! TYPE: int default: 10
LOG_INTERVAL: 10
# EVAL_HOOKS DESCRIPTION: TYPE: default: ''
EVAL_HOOKS:
- # NAME DESCRIPTION: TYPE: default: 'LogHook'
NAME: LogHook
# LOG_INTERVAL DESCRIPTION: the interval for log print! TYPE: int default: 10
LOG_INTERVAL: 10
# TEST_HOOKS DESCRIPTION: TYPE: default: ''
MODEL:
# NAME DESCRIPTION: TYPE: default: 'Classifier'
NAME: Classifier
# ACT_NAME DESCRIPTION: the activation function for logits, select from [softmax, sigmoid]! TYPE: str default: 'softmax'
ACT_NAME: softmax
# FREEZE_BN DESCRIPTION: if freeze bn of not TYPE: bool default: False
FREEZE_BN: False
# BACKBONE DESCRIPTION: TYPE: default: ''
BACKBONE:
# NAME DESCRIPTION: TYPE: default: 'ResNet'
NAME: ResNet
# DEPTH DESCRIPTION: the depth of network for resnet! TYPE: int default: 18
DEPTH: 18
# PRETRAINED DESCRIPTION: if load the official pretrained model or not. TYPE: bool default: False
PRETRAINED: false
#
KERNEL_SIZE: 3
# USE_RELU DESCRIPTION: use relu or not! TYPE: bool default: True
USE_RELU: True
# USE_MAXPOOL DESCRIPTION: use maxpool or not! TYPE: bool default: True
USE_MAXPOOL: false
# FIRST_CONV_STRIDE DESCRIPTION: first conv stride 1 or 2! TYPE: int default: 1
FIRST_CONV_STRIDE: 1
# FIRST_MAX_POOL_STRIDE DESCRIPTION: first max pool stride 1 or 2! TYPE: int default: 1
FIRST_MAX_POOL_STRIDE: 1
# NECK DESCRIPTION: TYPE: default: ''
NECK:
# NAME DESCRIPTION: TYPE: default: 'GlobalAveragePooling'
NAME: GlobalAveragePooling
# DIM DESCRIPTION: GlobalAveragePooling dim! TYPE: int default: 2
DIM: 2
# HEAD DESCRIPTION: TYPE: default: ''
HEAD:
# NAME DESCRIPTION: TYPE: default: 'ClassifierHead'
NAME: ClassifierHead
# DIM DESCRIPTION: representation dim! TYPE: int default: 512
DIM: 512
# NUM_CLASSES DESCRIPTION: number of classes. TYPE: int default: 10
NUM_CLASSES: 10
# DROPOUT_RATE DESCRIPTION: dropout rate, default 0. TYPE: float default: 0.0
DROPOUT_RATE: 0.0
METRIC:
# NAME DESCRIPTION: TYPE: default: 'AccuracyMetric'
NAME: AccuracyMetric
# TOPK DESCRIPTION: topk accuracy! TYPE: int default: 1
TOPK: 1
# LOSS DESCRIPTION: TYPE: default: ''
LOSS:
# NAME DESCRIPTION: TYPE: default: 'CrossEntropy'
NAME: CrossEntropy
# REDUCE DESCRIPTION: reduce is False, returns a loss per batch element instead and ignores :attr: size_average. Default: True TYPE: NoneType default: None
# REDUCE: None
# SIZE_AVERAGE DESCRIPTION: Deprecated (see :attr: reduction). By default,the losses are averaged over each loss element in the batch. Note that forsome losses, there are multiple elements per sample. If the field :attr: size_averageis set to False, the losses are instead summed for each minibatch. Ignoredwhen :attr: reduce is False. Default: True TYPE: NoneType default: None
# SIZE_AVERAGE: None
# IGNORE_INDEX DESCRIPTION: Specifies a target value that is ignoredand does not contribute to the input gradient. When :attr: size_average isTrue, the loss is averaged over non-ignored targets. Note that:attr: ignore_index is only applicable when the target contains class indices. TYPE: int default: -100
# IGNORE_INDEX: -100
# REDUCTION DESCRIPTION: Specifies the reduction to apply to the output:'none' | 'mean' | 'sum'. 'none': no reduction willbe applied, 'mean': the weighted mean of the output is taken,'sum': the output will be summed. Note: :attr: size_averageand :attr:`reduce` are in the process of being deprecated, and inthe meantime, specifying either of those two args will override:attr:`reduction`. Default: 'mean' TYPE: str default: 'mean'
# REDUCTION: mean
# LABEL_SMOOTHING DESCRIPTION: A float in [0.0, 1.0]. Specifies the amountof smoothing when computing the loss, where 0.0 means no smoothing. TYPE: float default: 0.0
# LABEL_SMOOTHING: 0.0
# OPTIMIZER DESCRIPTION: TYPE: default: ''
OPTIMIZER:
# NAME DESCRIPTION: TYPE: default: 'SGD'
NAME: SGD
# LEARNING_RATE DESCRIPTION: the initial learning rate! TYPE: float default: 0.1
LEARNING_RATE: 0.01
# MOMENTUM DESCRIPTION: the momentum! TYPE: int default: 0
MOMENTUM: 0.9
# DAMPENING DESCRIPTION: the dampening! TYPE: int default: 0
DAMPENING: 0
# WEIGHT_DECAY DESCRIPTION: the weight decay! TYPE: int default: 0
WEIGHT_DECAY: 5e-4
# NESTEROV DESCRIPTION: the nesterov! TYPE: bool default: False
NESTEROV: False
# LR_SCHEDULER DESCRIPTION: TYPE: default: ''
LR_SCHEDULER:
# NAME DESCRIPTION: TYPE: default: 'CosineAnnealingLR'
NAME: CosineAnnealingLR
# T_MAX DESCRIPTION: the T max! TYPE: float default: 1.0
T_MAX: 200.0
# ETA_MIN DESCRIPTION: the eta min! TYPE: int default: 0
ETA_MIN: 0
# LAST_EPOCH DESCRIPTION: the last epoch! TYPE: int default: -1
LAST_EPOCH: -1
# METRICS DESCRIPTION: TYPE: default: ''
METRICS:
- # NAME DESCRIPTION: TYPE: default: 'AccuracyMetric'
NAME: AccuracyMetric
# TOPK DESCRIPTION: topk accuracy! TYPE: int default: 1
TOPK: 1
KEYS: ["logits", "label"]
scepter/methods/examples/generation/stable_diffusion_1.5_512.yaml
+218
View File
@@ -0,0 +1,218 @@
ENV:
BACKEND: nccl
SOLVER:
NAME: LatentDiffusionSolver
RESUME_FROM:
LOAD_MODEL_ONLY: True
USE_FSDP: False
SHARDING_STRATEGY:
USE_AMP: True
DTYPE: float16
CHANNELS_LAST: True
MAX_STEPS: 2000
MAX_EPOCHS: -1
NUM_FOLDS: 1
ACCU_STEP: 1
EVAL_INTERVAL: 100
#
WORK_DIR: ./cache/sd15_512_full
LOG_FILE: std_log.txt
#
FILE_SYSTEM:
NAME: "ModelscopeFs"
TEMP_DIR: "./cache/data"
#
MODEL:
NAME: LatentDiffusion
PARAMETERIZATION: eps
TIMESTEPS: 1000
MIN_SNR_GAMMA:
ZERO_TERMINAL_SNR: False
PRETRAINED_MODEL: ms://AI-ModelScope/stable-diffusion-v1-5@v1-5-pruned-emaonly.safetensors
IGNORE_KEYS: [ ]
SCALE_FACTOR: 0.18215
SIZE_FACTOR: 8
# DEFAULT_N_PROMPT: 'lowres, error, worst quality, low quality, jpeg artifacts, ugly, duplicate, morbid, mutilated, out of frame, extra fingers, mutated hands, poorly drawn hands, poorly drawn face, mutation, deformed, blurry, dehydrated, bad anatomy, bad proportions, extra limbs, cloned face, disfigured, gross proportions, malformed limbs, missing arms, missing legs, extra arms, extra legs, fused fingers, too many fingers, long neck, username, watermark, signature'
DEFAULT_N_PROMPT:
SCHEDULE_ARGS:
"NAME": "scaled_linear"
"BETA_MIN": 0.00085
"BETA_MAX": 0.012
USE_EMA: False
#
DIFFUSION_MODEL:
NAME: DiffusionUNet
IN_CHANNELS: 4
OUT_CHANNELS: 4
MODEL_CHANNELS: 320
NUM_HEADS: 8
NUM_RES_BLOCKS: 2
ATTENTION_RESOLUTIONS: [ 4, 2, 1 ]
CHANNEL_MULT: [ 1, 2, 4, 4 ]
CONV_RESAMPLE: True
DIMS: 2
USE_CHECKPOINT: False
USE_SCALE_SHIFT_NORM: False
RESBLOCK_UPDOWN: False
USE_SPATIAL_TRANSFORMER: True
TRANSFORMER_DEPTH: 1
CONTEXT_DIM: 768
DISABLE_MIDDLE_SELF_ATTN: False
USE_LINEAR_IN_TRANSFORMER: False
PRETRAINED_MODEL:
IGNORE_KEYS: []
#
FIRST_STAGE_MODEL:
NAME: AutoencoderKL
EMBED_DIM: 4
PRETRAINED_MODEL:
IGNORE_KEYS: []
BATCH_SIZE: 4
#
ENCODER:
NAME: Encoder
CH: 128
OUT_CH: 3
NUM_RES_BLOCKS: 2
IN_CHANNELS: 3
ATTN_RESOLUTIONS: [ ]
CH_MULT: [ 1, 2, 4, 4 ]
Z_CHANNELS: 4
DOUBLE_Z: True
DROPOUT: 0.0
RESAMP_WITH_CONV: True
#
DECODER:
NAME: Decoder
CH: 128
OUT_CH: 3
NUM_RES_BLOCKS: 2
IN_CHANNELS: 3
ATTN_RESOLUTIONS: [ ]
CH_MULT: [ 1, 2, 4, 4 ]
Z_CHANNELS: 4
DROPOUT: 0.0
RESAMP_WITH_CONV: True
GIVE_PRE_END: False
TANH_OUT: False
#
TOKENIZER:
NAME: ClipTokenizer
PRETRAINED_PATH: ms://AI-ModelScope/clip-vit-large-patch14
LENGTH: 77
CLEAN: True
#
COND_STAGE_MODEL:
NAME: FrozenCLIPEmbedder
FREEZE: True
LAYER: last
PRETRAINED_MODEL: ms://AI-ModelScope/clip-vit-large-patch14
#
LOSS:
NAME: ReconstructLoss
LOSS_TYPE: l2
#
SAMPLE_ARGS:
SAMPLER: ddim
SAMPLE_STEPS: 50
SEED: 2023
GUIDE_SCALE: 7.5
GUIDE_RESCALE:
DISCRETIZATION: trailing
IMAGE_SIZE: [512, 512]
RUN_TRAIN_N: False
#
OPTIMIZER:
NAME: AdamW
LEARNING_RATE: 0.0064
BETAS: [ 0.9, 0.999 ]
EPS: 1e-8
WEIGHT_DECAY: 1e-2
AMSGRAD: False
#
TRAIN_DATA:
NAME: ImageTextPairMSDataset
MODE: train
MS_DATASET_NAME: style_custom_dataset
MS_DATASET_NAMESPACE: damo
MS_DATASET_SUBNAME: 3D
PROMPT_PREFIX: ""
MS_DATASET_SPLIT: train_short
MS_REMAP_KEYS: { 'Image:FILE': 'Target:FILE' }
REPLACE_STYLE: False
PIN_MEMORY: True
BATCH_SIZE: 1
NUM_WORKERS: 4
SAMPLER:
NAME: LoopSampler
TRANSFORMS:
- NAME: LoadImageFromFile
RGB_ORDER: RGB
BACKEND: pillow
- NAME: Resize
SIZE: 512
INTERPOLATION: bilinear
INPUT_KEY: [ 'img' ]
OUTPUT_KEY: [ 'img' ]
BACKEND: pillow
- NAME: CenterCrop
SIZE: 512
INPUT_KEY: [ 'img' ]
OUTPUT_KEY: [ 'img' ]
BACKEND: pillow
- NAME: ImageToTensor
INPUT_KEY: [ 'img' ]
OUTPUT_KEY: [ 'img' ]
BACKEND: pillow
- NAME: Normalize
MEAN: [ 0.5, 0.5, 0.5 ]
STD: [ 0.5, 0.5, 0.5 ]
INPUT_KEY: [ 'img' ]
OUTPUT_KEY: [ 'image' ]
BACKEND: torchvision
- NAME: Select
KEYS: [ 'image', 'prompt' ]
META_KEYS: [ 'data_key' ]
#
EVAL_DATA:
NAME: ImageTextPairMSDataset
MODE: eval
MS_DATASET_NAME: style_custom_dataset
MS_DATASET_NAMESPACE: damo
MS_DATASET_SUBNAME: 3D
PROMPT_PREFIX: ""
MS_REMAP_KEYS: { 'Image': 'Target:FILE' }
MS_DATASET_SPLIT: test_short
OUTPUT_SIZE: [512, 512]
REPLACE_STYLE: False
PIN_MEMORY: True
BATCH_SIZE: 4
NUM_WORKERS: 4
FILE_SYSTEM:
NAME: "ModelscopeFs"
TEMP_DIR: "./cache/data"
#
TRANSFORMS:
-
NAME: Select
KEYS: ['prompt']
META_KEYS: ['image_size']
#
TRAIN_HOOKS:
-
NAME: BackwardHook
PRIORITY: 0
-
NAME: LogHook
LOG_INTERVAL: 50
-
NAME: CheckpointHook
INTERVAL: 1000
-
NAME: ProbeDataHook
PROB_INTERVAL: 100
#
EVAL_HOOKS:
-
NAME: ProbeDataHook
PROB_INTERVAL: 100
scepter/methods/examples/generation/stable_diffusion_1.5_512_lora.yaml
+226
View File
@@ -0,0 +1,226 @@
ENV:
BACKEND: nccl
SOLVER:
NAME: LatentDiffusionSolver
RESUME_FROM:
LOAD_MODEL_ONLY: True
USE_FSDP: False
SHARDING_STRATEGY:
USE_AMP: True
DTYPE: float16
CHANNELS_LAST: True
MAX_STEPS: 2000
MAX_EPOCHS: -1
NUM_FOLDS: 1
ACCU_STEP: 1
EVAL_INTERVAL: 100
#
WORK_DIR: ./cache/sd15_512_lora
LOG_FILE: std_log.txt
#
FILE_SYSTEM:
NAME: "ModelscopeFs"
TEMP_DIR: "./cache/data"
TUNER:
-
NAME: SwiftLoRA
R: 64
LORA_ALPHA: 64
LORA_DROPOUT: 0.0
BIAS: "none"
TARGET_MODULES: model.*(to_q|to_k|to_v|to_out.0|net.0.proj|net.2)$
#
MODEL:
NAME: LatentDiffusion
PARAMETERIZATION: eps
TIMESTEPS: 1000
MIN_SNR_GAMMA:
ZERO_TERMINAL_SNR: False
PRETRAINED_MODEL: ms://AI-ModelScope/stable-diffusion-v1-5@v1-5-pruned-emaonly.safetensors
IGNORE_KEYS: [ ]
SCALE_FACTOR: 0.18215
SIZE_FACTOR: 8
# DEFAULT_N_PROMPT: 'lowres, error, worst quality, low quality, jpeg artifacts, ugly, duplicate, morbid, mutilated, out of frame, extra fingers, mutated hands, poorly drawn hands, poorly drawn face, mutation, deformed, blurry, dehydrated, bad anatomy, bad proportions, extra limbs, cloned face, disfigured, gross proportions, malformed limbs, missing arms, missing legs, extra arms, extra legs, fused fingers, too many fingers, long neck, username, watermark, signature'
DEFAULT_N_PROMPT:
SCHEDULE_ARGS:
"NAME": "scaled_linear"
"BETA_MIN": 0.00085
"BETA_MAX": 0.012
USE_EMA: False
#
DIFFUSION_MODEL:
NAME: DiffusionUNet
IN_CHANNELS: 4
OUT_CHANNELS: 4
MODEL_CHANNELS: 320
NUM_HEADS: 8
NUM_RES_BLOCKS: 2
ATTENTION_RESOLUTIONS: [ 4, 2, 1 ]
CHANNEL_MULT: [ 1, 2, 4, 4 ]
CONV_RESAMPLE: True
DIMS: 2
USE_CHECKPOINT: False
USE_SCALE_SHIFT_NORM: False
RESBLOCK_UPDOWN: False
USE_SPATIAL_TRANSFORMER: True
TRANSFORMER_DEPTH: 1
CONTEXT_DIM: 768
DISABLE_MIDDLE_SELF_ATTN: False
USE_LINEAR_IN_TRANSFORMER: False
PRETRAINED_MODEL:
IGNORE_KEYS: []
#
FIRST_STAGE_MODEL:
NAME: AutoencoderKL
EMBED_DIM: 4
PRETRAINED_MODEL:
IGNORE_KEYS: []
BATCH_SIZE: 4
#
ENCODER:
NAME: Encoder
CH: 128
OUT_CH: 3
NUM_RES_BLOCKS: 2
IN_CHANNELS: 3
ATTN_RESOLUTIONS: [ ]
CH_MULT: [ 1, 2, 4, 4 ]
Z_CHANNELS: 4
DOUBLE_Z: True
DROPOUT: 0.0
RESAMP_WITH_CONV: True
#
DECODER:
NAME: Decoder
CH: 128
OUT_CH: 3
NUM_RES_BLOCKS: 2
IN_CHANNELS: 3
ATTN_RESOLUTIONS: [ ]
CH_MULT: [ 1, 2, 4, 4 ]
Z_CHANNELS: 4
DROPOUT: 0.0
RESAMP_WITH_CONV: True
GIVE_PRE_END: False
TANH_OUT: False
#
TOKENIZER:
NAME: ClipTokenizer
PRETRAINED_PATH: ms://AI-ModelScope/clip-vit-large-patch14
LENGTH: 77
CLEAN: True
#
COND_STAGE_MODEL:
NAME: FrozenCLIPEmbedder
FREEZE: True
LAYER: last
PRETRAINED_MODEL: ms://AI-ModelScope/clip-vit-large-patch14
#
LOSS:
NAME: ReconstructLoss
LOSS_TYPE: l2
#
SAMPLE_ARGS:
SAMPLER: ddim
SAMPLE_STEPS: 50
SEED: 2023
GUIDE_SCALE: 7.5
GUIDE_RESCALE:
DISCRETIZATION: trailing
IMAGE_SIZE: [512, 512]
RUN_TRAIN_N: False
#
OPTIMIZER:
NAME: AdamW
LEARNING_RATE: 0.064
BETAS: [ 0.9, 0.999 ]
EPS: 1e-8
WEIGHT_DECAY: 1e-2
AMSGRAD: False
#
TRAIN_DATA:
NAME: ImageTextPairMSDataset
MODE: train
MS_DATASET_NAME: style_custom_dataset
MS_DATASET_NAMESPACE: damo
MS_DATASET_SUBNAME: 3D
PROMPT_PREFIX: ""
MS_DATASET_SPLIT: train_short
MS_REMAP_KEYS: { 'Image:FILE': 'Target:FILE' }
REPLACE_STYLE: False
PIN_MEMORY: True
BATCH_SIZE: 1
NUM_WORKERS: 4
SAMPLER:
NAME: LoopSampler
TRANSFORMS:
- NAME: LoadImageFromFile
RGB_ORDER: RGB
BACKEND: pillow
- NAME: Resize
SIZE: 512
INTERPOLATION: bilinear
INPUT_KEY: [ 'img' ]
OUTPUT_KEY: [ 'img' ]
BACKEND: pillow
- NAME: CenterCrop
SIZE: 512
INPUT_KEY: [ 'img' ]
OUTPUT_KEY: [ 'img' ]
BACKEND: pillow
- NAME: ImageToTensor
INPUT_KEY: [ 'img' ]
OUTPUT_KEY: [ 'img' ]
BACKEND: pillow
- NAME: Normalize
MEAN: [ 0.5, 0.5, 0.5 ]
STD: [ 0.5, 0.5, 0.5 ]
INPUT_KEY: [ 'img' ]
OUTPUT_KEY: [ 'image' ]
BACKEND: torchvision
- NAME: Select
KEYS: [ 'image', 'prompt' ]
META_KEYS: [ 'data_key' ]
#
EVAL_DATA:
NAME: ImageTextPairMSDataset
MODE: eval
MS_DATASET_NAME: style_custom_dataset
MS_DATASET_NAMESPACE: damo
MS_DATASET_SUBNAME: 3D
PROMPT_PREFIX: ""
MS_REMAP_KEYS: { 'Image': 'Target:FILE' }
MS_DATASET_SPLIT: test_short
OUTPUT_SIZE: [512, 512]
REPLACE_STYLE: False
PIN_MEMORY: True
BATCH_SIZE: 4
NUM_WORKERS: 4
FILE_SYSTEM:
NAME: "ModelscopeFs"
TEMP_DIR: "./cache/data"
#
TRANSFORMS:
-
NAME: Select
KEYS: ['prompt']
META_KEYS: ['image_size']
#
TRAIN_HOOKS:
-
NAME: BackwardHook
PRIORITY: 0
-
NAME: LogHook
LOG_INTERVAL: 50
-
NAME: CheckpointHook
INTERVAL: 1000
-
NAME: ProbeDataHook
PROB_INTERVAL: 100
#
EVAL_HOOKS:
-
NAME: ProbeDataHook
PROB_INTERVAL: 100
scepter/methods/examples/generation/stable_diffusion_2.1_768.yaml
+214
View File
@@ -0,0 +1,214 @@
ENV:
BACKEND: nccl
SOLVER:
NAME: LatentDiffusionSolver
RESUME_FROM:
LOAD_MODEL_ONLY: True
USE_FSDP: False
SHARDING_STRATEGY:
USE_AMP: True
DTYPE: float16
CHANNELS_LAST: True
MAX_STEPS: 2000
MAX_EPOCHS: -1
NUM_FOLDS: 1
ACCU_STEP: 1
EVAL_INTERVAL: 100
#
WORK_DIR: ./cache/sd21_768_full
LOG_FILE: std_log.txt
#
FILE_SYSTEM:
NAME: "ModelscopeFs"
TEMP_DIR: "./cache/data"
#
MODEL:
NAME: LatentDiffusion
PARAMETERIZATION: v
TIMESTEPS: 1000
MIN_SNR_GAMMA:
ZERO_TERMINAL_SNR: False
PRETRAINED_MODEL: ms://AI-ModelScope/stable-diffusion-2-1@v2-1_768-ema-pruned.safetensors
IGNORE_KEYS: [ ]
SCALE_FACTOR: 0.18215
SIZE_FACTOR: 8
DEFAULT_N_PROMPT:
SCHEDULE_ARGS:
"NAME": "scaled_linear"
"BETA_MIN": 0.00085
"BETA_MAX": 0.012
USE_EMA: False
#
DIFFUSION_MODEL:
NAME: DiffusionUNet
IN_CHANNELS: 4
OUT_CHANNELS: 4
MODEL_CHANNELS: 320
NUM_HEADS_CHANNELS: 64
NUM_RES_BLOCKS: 2
ATTENTION_RESOLUTIONS: [ 4, 2, 1 ]
CHANNEL_MULT: [ 1, 2, 4, 4 ]
CONV_RESAMPLE: True
DIMS: 2
USE_CHECKPOINT: False
USE_SCALE_SHIFT_NORM: False
RESBLOCK_UPDOWN: False
USE_SPATIAL_TRANSFORMER: True
TRANSFORMER_DEPTH: 1
CONTEXT_DIM: 1024
DISABLE_MIDDLE_SELF_ATTN: False
USE_LINEAR_IN_TRANSFORMER: True
PRETRAINED_MODEL:
#
FIRST_STAGE_MODEL:
NAME: AutoencoderKL
EMBED_DIM: 4
PRETRAINED_MODEL:
IGNORE_KEYS: [ ]
BATCH_SIZE: 4
#
ENCODER:
NAME: Encoder
CH: 128
OUT_CH: 3
NUM_RES_BLOCKS: 2
IN_CHANNELS: 3
ATTN_RESOLUTIONS: [ ]
CH_MULT: [ 1, 2, 4, 4 ]
Z_CHANNELS: 4
DOUBLE_Z: True
DROPOUT: 0.0
RESAMP_WITH_CONV: True
#
DECODER:
NAME: Decoder
CH: 128
OUT_CH: 3
NUM_RES_BLOCKS: 2
IN_CHANNELS: 3
ATTN_RESOLUTIONS: [ ]
CH_MULT: [ 1, 2, 4, 4 ]
Z_CHANNELS: 4
DROPOUT: 0.0
RESAMP_WITH_CONV: True
GIVE_PRE_END: False
TANH_OUT: False
#
TOKENIZER:
NAME: OpenClipTokenizer
LENGTH: 77
#
COND_STAGE_MODEL:
NAME: FrozenOpenCLIPEmbedder
ARCH: ViT-H-14
PRETRAINED_MODEL:
LAYER: penultimate
#
LOSS:
NAME: ReconstructLoss
LOSS_TYPE: l2
#
SAMPLE_ARGS:
SAMPLER: ddim
SAMPLE_STEPS: 50
SEED: 2023
GUIDE_SCALE: 7.5
GUIDE_RESCALE:
DISCRETIZATION: trailing
IMAGE_SIZE: [768, 768]
RUN_TRAIN_N: False
#
OPTIMIZER:
NAME: AdamW
LEARNING_RATE: 0.0064
BETAS: [ 0.9, 0.999 ]
EPS: 1e-8
WEIGHT_DECAY: 1e-2
AMSGRAD: False
#
TRAIN_DATA:
NAME: ImageTextPairMSDataset
MODE: train
MS_DATASET_NAME: style_custom_dataset
MS_DATASET_NAMESPACE: damo
MS_DATASET_SUBNAME: 3D
PROMPT_PREFIX: ""
MS_DATASET_SPLIT: train_short
MS_REMAP_KEYS: { 'Image:FILE': 'Target:FILE' }
REPLACE_STYLE: False
PIN_MEMORY: True
BATCH_SIZE: 1
NUM_WORKERS: 4
SAMPLER:
NAME: LoopSampler
TRANSFORMS:
- NAME: LoadImageFromFile
RGB_ORDER: RGB
BACKEND: pillow
- NAME: Resize
SIZE: 768
INTERPOLATION: bilinear
INPUT_KEY: [ 'img' ]
OUTPUT_KEY: [ 'img' ]
BACKEND: pillow
- NAME: CenterCrop
SIZE: 768
INPUT_KEY: [ 'img' ]
OUTPUT_KEY: [ 'img' ]
BACKEND: pillow
- NAME: ImageToTensor
INPUT_KEY: [ 'img' ]
OUTPUT_KEY: [ 'img' ]
BACKEND: pillow
- NAME: Normalize
MEAN: [ 0.5, 0.5, 0.5 ]
STD: [ 0.5, 0.5, 0.5 ]
INPUT_KEY: [ 'img' ]
OUTPUT_KEY: [ 'image' ]
BACKEND: torchvision
- NAME: Select
KEYS: [ 'image', 'prompt' ]
META_KEYS: [ 'data_key' ]
#
EVAL_DATA:
NAME: ImageTextPairMSDataset
MODE: eval
MS_DATASET_NAME: style_custom_dataset
MS_DATASET_NAMESPACE: damo
MS_DATASET_SUBNAME: 3D
PROMPT_PREFIX: ""
MS_REMAP_KEYS: { 'Image': 'Target:FILE' }
MS_DATASET_SPLIT: test_short
OUTPUT_SIZE: [768, 768]
REPLACE_STYLE: False
PIN_MEMORY: True
BATCH_SIZE: 4
NUM_WORKERS: 4
FILE_SYSTEM:
NAME: "ModelscopeFs"
TEMP_DIR: "./cache/data"
#
TRANSFORMS:
-
NAME: Select
KEYS: ['prompt']
META_KEYS: ['image_size']
#
TRAIN_HOOKS:
-
NAME: BackwardHook
PRIORITY: 0
-
NAME: LogHook
LOG_INTERVAL: 50
-
NAME: CheckpointHook
INTERVAL: 1000
-
NAME: ProbeDataHook
PROB_INTERVAL: 100
#
EVAL_HOOKS:
-
NAME: ProbeDataHook
PROB_INTERVAL: 100
scepter/methods/examples/generation/stable_diffusion_2.1_768_lora.yaml
+223
View File
@@ -0,0 +1,223 @@
ENV:
BACKEND: nccl
SOLVER:
NAME: LatentDiffusionSolver
RESUME_FROM:
LOAD_MODEL_ONLY: True
USE_FSDP: False
SHARDING_STRATEGY:
USE_AMP: True
DTYPE: float16
CHANNELS_LAST: True
MAX_STEPS: 2000
MAX_EPOCHS: -1
NUM_FOLDS: 1
ACCU_STEP: 1
EVAL_INTERVAL: 100
#
WORK_DIR: ./cache/sd21_768_lora
LOG_FILE: std_log.txt
#
FILE_SYSTEM:
NAME: "ModelscopeFs"
TEMP_DIR: "./cache/data"
#
TUNER:
-
NAME: SwiftLoRA
R: 64
LORA_ALPHA: 64
LORA_DROPOUT: 0.0
BIAS: "none"
TARGET_MODULES: model.*(to_q|to_k|to_v|to_out.0|net.0.proj|net.2)$
#
MODEL:
NAME: LatentDiffusion
PARAMETERIZATION: v
TIMESTEPS: 1000
MIN_SNR_GAMMA:
ZERO_TERMINAL_SNR: False
PRETRAINED_MODEL: ms://AI-ModelScope/stable-diffusion-2-1@v2-1_768-ema-pruned.safetensors
IGNORE_KEYS: [ ]
SCALE_FACTOR: 0.18215
SIZE_FACTOR: 8
DEFAULT_N_PROMPT:
SCHEDULE_ARGS:
"NAME": "scaled_linear"
"BETA_MIN": 0.00085
"BETA_MAX": 0.012
USE_EMA: False
#
DIFFUSION_MODEL:
NAME: DiffusionUNet
IN_CHANNELS: 4
OUT_CHANNELS: 4
MODEL_CHANNELS: 320
NUM_HEADS_CHANNELS: 64
NUM_RES_BLOCKS: 2
ATTENTION_RESOLUTIONS: [ 4, 2, 1 ]
CHANNEL_MULT: [ 1, 2, 4, 4 ]
CONV_RESAMPLE: True
DIMS: 2
USE_CHECKPOINT: False
USE_SCALE_SHIFT_NORM: False
RESBLOCK_UPDOWN: False
USE_SPATIAL_TRANSFORMER: True
TRANSFORMER_DEPTH: 1
CONTEXT_DIM: 1024
DISABLE_MIDDLE_SELF_ATTN: False
USE_LINEAR_IN_TRANSFORMER: True
PRETRAINED_MODEL:
#
FIRST_STAGE_MODEL:
NAME: AutoencoderKL
EMBED_DIM: 4
PRETRAINED_MODEL:
IGNORE_KEYS: [ ]
BATCH_SIZE: 4
#
ENCODER:
NAME: Encoder
CH: 128
OUT_CH: 3
NUM_RES_BLOCKS: 2
IN_CHANNELS: 3
ATTN_RESOLUTIONS: [ ]
CH_MULT: [ 1, 2, 4, 4 ]
Z_CHANNELS: 4
DOUBLE_Z: True
DROPOUT: 0.0
RESAMP_WITH_CONV: True
#
DECODER:
NAME: Decoder
CH: 128
OUT_CH: 3
NUM_RES_BLOCKS: 2
IN_CHANNELS: 3
ATTN_RESOLUTIONS: [ ]
CH_MULT: [ 1, 2, 4, 4 ]
Z_CHANNELS: 4
DROPOUT: 0.0
RESAMP_WITH_CONV: True
GIVE_PRE_END: False
TANH_OUT: False
#
TOKENIZER:
NAME: OpenClipTokenizer
LENGTH: 77
#
COND_STAGE_MODEL:
NAME: FrozenOpenCLIPEmbedder
ARCH: ViT-H-14
PRETRAINED_MODEL:
LAYER: penultimate
#
LOSS:
NAME: ReconstructLoss
LOSS_TYPE: l2
#
SAMPLE_ARGS:
SAMPLER: ddim
SAMPLE_STEPS: 50
SEED: 2023
GUIDE_SCALE: 7.5
GUIDE_RESCALE:
DISCRETIZATION: trailing
IMAGE_SIZE: [768, 768]
RUN_TRAIN_N: False
#
OPTIMIZER:
NAME: AdamW
LEARNING_RATE: 0.064
BETAS: [ 0.9, 0.999 ]
EPS: 1e-8
WEIGHT_DECAY: 1e-2
AMSGRAD: False
#
TRAIN_DATA:
NAME: ImageTextPairMSDataset
MODE: train
MS_DATASET_NAME: style_custom_dataset
MS_DATASET_NAMESPACE: damo
MS_DATASET_SUBNAME: 3D
PROMPT_PREFIX: ""
MS_DATASET_SPLIT: train_short
MS_REMAP_KEYS: { 'Image:FILE': 'Target:FILE' }
REPLACE_STYLE: False
PIN_MEMORY: True
BATCH_SIZE: 1
NUM_WORKERS: 4
SAMPLER:
NAME: LoopSampler
TRANSFORMS:
- NAME: LoadImageFromFile
RGB_ORDER: RGB
BACKEND: pillow
- NAME: Resize
SIZE: 768
INTERPOLATION: bilinear
INPUT_KEY: [ 'img' ]
OUTPUT_KEY: [ 'img' ]
BACKEND: pillow
- NAME: CenterCrop
SIZE: 768
INPUT_KEY: [ 'img' ]
OUTPUT_KEY: [ 'img' ]
BACKEND: pillow
- NAME: ImageToTensor
INPUT_KEY: [ 'img' ]
OUTPUT_KEY: [ 'img' ]
BACKEND: pillow
- NAME: Normalize
MEAN: [ 0.5, 0.5, 0.5 ]
STD: [ 0.5, 0.5, 0.5 ]
INPUT_KEY: [ 'img' ]
OUTPUT_KEY: [ 'image' ]
BACKEND: torchvision
- NAME: Select
KEYS: [ 'image', 'prompt' ]
META_KEYS: [ 'data_key' ]
#
EVAL_DATA:
NAME: ImageTextPairMSDataset
MODE: eval
MS_DATASET_NAME: style_custom_dataset
MS_DATASET_NAMESPACE: damo
MS_DATASET_SUBNAME: 3D
PROMPT_PREFIX: ""
MS_REMAP_KEYS: { 'Image': 'Target:FILE' }
MS_DATASET_SPLIT: test_short
OUTPUT_SIZE: [768, 768]
REPLACE_STYLE: False
PIN_MEMORY: True
BATCH_SIZE: 4
NUM_WORKERS: 4
FILE_SYSTEM:
NAME: "ModelscopeFs"
TEMP_DIR: "./cache/data"
#
TRANSFORMS:
-
NAME: Select
KEYS: ['prompt']
META_KEYS: ['image_size']
#
TRAIN_HOOKS:
-
NAME: BackwardHook
PRIORITY: 0
-
NAME: LogHook
LOG_INTERVAL: 50
-
NAME: CheckpointHook
INTERVAL: 1000
-
NAME: ProbeDataHook
PROB_INTERVAL: 100
#
EVAL_HOOKS:
-
NAME: ProbeDataHook
PROB_INTERVAL: 100
scepter/methods/examples/generation/stable_diffusion_xl_1024.yaml
+328
View File
@@ -0,0 +1,328 @@
ENV:
BACKEND: nccl
SOLVER:
NAME: LatentDiffusionSolver
RESUME_FROM:
LOAD_MODEL_ONLY: True
USE_FSDP: False
SHARDING_STRATEGY:
USE_AMP: True
DTYPE: float16
CHANNELS_LAST: True
MAX_STEPS: 2000
MAX_EPOCHS: -1
NUM_FOLDS: 1
ACCU_STEP: 1
EVAL_INTERVAL: 100
#
WORK_DIR: ./cache/sdxl_1024_full
LOG_FILE: std_log.txt
#
FILE_SYSTEM:
NAME: "ModelscopeFs"
TEMP_DIR: "./cache/data"
#
MODEL:
NAME: LatentDiffusionXL
PARAMETERIZATION: eps
TIMESTEPS: 1000
MIN_SNR_GAMMA:
ZERO_TERMINAL_SNR: False
PRETRAINED_MODEL: ms://AI-ModelScope/stable-diffusion-xl-base-1.0@sd_xl_base_1.0.safetensors
IGNORE_KEYS: [ ]
SCALE_FACTOR: 0.13025
SIZE_FACTOR: 8
# DEFAULT_N_PROMPT: 'lowres, error, worst quality, low quality, jpeg artifacts, ugly, duplicate, morbid, mutilated, out of frame, extra fingers, mutated hands, poorly drawn hands, poorly drawn face, mutation, deformed, blurry, dehydrated, bad anatomy, bad proportions, extra limbs, cloned face, disfigured, gross proportions, malformed limbs, missing arms, missing legs, extra arms, extra legs, fused fingers, too many fingers, long neck, username, watermark, signature'
DEFAULT_N_PROMPT:
SCHEDULE_ARGS:
"NAME": "scaled_linear"
"BETA_MIN": 0.00085
"BETA_MAX": 0.0120
USE_EMA: False
LOAD_REFINER: False
#
DIFFUSION_MODEL:
NAME: DiffusionUNetXL
PRETRAINED_MODEL:
IN_CHANNELS: 4
OUT_CHANNELS: 4
NUM_RES_BLOCKS: 2
MODEL_CHANNELS: 320
ATTENTION_RESOLUTIONS: [ 4, 2 ]
DROPOUT: 0
CHANNEL_MULT: [ 1, 2, 4 ]
CONV_RESAMPLE: True
DIMS: 2
NUM_CLASSES: sequential
USE_CHECKPOINT: False
NUM_HEADS: -1
NUM_HEADS_CHANNELS: 64
USE_SCALE_SHIFT_NORM: False
RESBLOCK_UPDOWN: False
USE_NEW_ATTENTION_ORDER: True
USE_SPATIAL_TRANSFORMER: True
TRANSFORMER_DEPTH: [ 1, 2, 10 ]
CONTEXT_DIM: 2048
DISABLE_MIDDLE_SELF_ATTN: False
USE_LINEAR_IN_TRANSFORMER: True
ADM_IN_CHANNELS: 2816
USE_SENTENCE_EMB: False
USE_WORD_MAPPING: False
#
FIRST_STAGE_MODEL:
NAME: AutoencoderKL
EMBED_DIM: 4
PRETRAINED_MODEL:
IGNORE_KEYS: []
BATCH_SIZE: 1
#
ENCODER:
NAME: Encoder
CH: 128
OUT_CH: 3
NUM_RES_BLOCKS: 2
IN_CHANNELS: 3
ATTN_RESOLUTIONS: [ ]
CH_MULT: [ 1, 2, 4, 4 ]
Z_CHANNELS: 4
DOUBLE_Z: True
DROPOUT: 0.0
RESAMP_WITH_CONV: True
#
DECODER:
NAME: Decoder
CH: 128
OUT_CH: 3
NUM_RES_BLOCKS: 2
IN_CHANNELS: 3
ATTN_RESOLUTIONS: [ ]
CH_MULT: [ 1, 2, 4, 4 ]
Z_CHANNELS: 4
DROPOUT: 0.0
RESAMP_WITH_CONV: True
GIVE_PRE_END: False
TANH_OUT: False
#
COND_STAGE_MODEL:
NAME: GeneralConditioner
PRETRAINED_MODEL:
EMBEDDERS:
-
NAME: FrozenCLIPEmbedder
PRETRAINED_MODEL: ms://AI-ModelScope/clip-vit-large-patch14
TOKENIZER_PATH: ms://AI-ModelScope/clip-vit-large-patch14
MAX_LENGTH: 77
FREEZE: True
LAYER: hidden
LAYER_IDX: 11
USE_FINAL_LAYER_NORM: False
IS_TRAINABLE: False
UCG_RATE: 0.0
INPUT_KEYS: [ "prompt" ]
LEGACY_UCG_VALUE:
-
NAME: FrozenOpenCLIPEmbedder2
ARCH: ViT-bigG-14
PRETRAINED_MODEL:
MAX_LENGTH: 77
FREEZE: True
ALWAYS_RETURN_POOLED: True
LEGACY: False
LAYER: penultimate
IS_TRAINABLE: False
UCG_RATE: 0.0
INPUT_KEYS: [ "prompt" ]
LEGACY_UCG_VALUE:
-
NAME: ConcatTimestepEmbedderND
OUT_DIM: 256
IS_TRAINABLE: False
UCG_RATE: 0.0
INPUT_KEYS: [ "original_size_as_tuple" ]
LEGACY_UCG_VALUE:
-
NAME: ConcatTimestepEmbedderND
OUT_DIM: 256
IS_TRAINABLE: False
UCG_RATE: 0.0
INPUT_KEYS: [ "crop_coords_top_left" ]
LEGACY_UCG_VALUE:
-
NAME: ConcatTimestepEmbedderND
OUT_DIM: 256
IS_TRAINABLE: False
UCG_RATE: 0.0
INPUT_KEYS: [ "target_size_as_tuple" ]
LEGACY_UCG_VALUE:
#
REFINER_MODEL:
NAME: DiffusionUNetXL
PRETRAINED_MODEL:
IN_CHANNELS: 4
OUT_CHANNELS: 4
NUM_RES_BLOCKS: 2
MODEL_CHANNELS: 384
ATTENTION_RESOLUTIONS: [ 4, 2 ]
DROPOUT: 0
CHANNEL_MULT: [ 1, 2, 4, 4 ]
CONV_RESAMPLE: True
DIMS: 2
NUM_CLASSES: sequential
USE_CHECKPOINT: False
NUM_HEADS: -1
NUM_HEADS_CHANNELS: 64
USE_SCALE_SHIFT_NORM: False
RESBLOCK_UPDOWN: False
USE_NEW_ATTENTION_ORDER: True
USE_SPATIAL_TRANSFORMER: True
TRANSFORMER_DEPTH: 4
CONTEXT_DIM: [ 1280, 1280, 1280, 1280 ]
DISABLE_MIDDLE_SELF_ATTN: False
USE_LINEAR_IN_TRANSFORMER: True
ADM_IN_CHANNELS: 2560
USE_SENTENCE_EMB: False
USE_WORD_MAPPING: False
#
REFINER_COND_MODEL:
NAME: GeneralConditioner
PRETRAINED_MODEL:
EMBEDDERS:
-
NAME: FrozenOpenCLIPEmbedder2
ARCH: ViT-bigG-14
PRETRAINED_MODEL:
MAX_LENGTH: 77
FREEZE: True
ALWAYS_RETURN_POOLED: True
LEGACY: False
LAYER: penultimate
IS_TRAINABLE: False
UCG_RATE: 0.0
INPUT_KEYS: [ "prompt" ]
LEGACY_UCG_VALUE:
-
NAME: ConcatTimestepEmbedderND
OUT_DIM: 256
IS_TRAINABLE: False
UCG_RATE: 0.0
INPUT_KEYS: [ "original_size_as_tuple" ]
LEGACY_UCG_VALUE:
-
NAME: ConcatTimestepEmbedderND
OUT_DIM: 256
IS_TRAINABLE: False
UCG_RATE: 0.0
INPUT_KEYS: [ "crop_coords_top_left" ]
LEGACY_UCG_VALUE:
-
NAME: ConcatTimestepEmbedderND
OUT_DIM: 256
IS_TRAINABLE: False
UCG_RATE: 0.0
INPUT_KEYS: [ "aesthetic_score" ]
LEGACY_UCG_VALUE:
#
LOSS:
NAME: ReconstructLoss
LOSS_TYPE: l2
#
SAMPLE_ARGS:
SAMPLER: ddim
SAMPLE_STEPS: 50
SEED: 2023
GUIDE_SCALE: 5.0
GUIDE_RESCALE:
DISCRETIZATION: trailing
IMAGE_SIZE: [1024, 1024]
RUN_TRAIN_N: False
#
OPTIMIZER:
NAME: AdamW
LEARNING_RATE: 0.0064
BETAS: [ 0.9, 0.999 ]
EPS: 1e-8
WEIGHT_DECAY: 1e-2
AMSGRAD: False
#
TRAIN_DATA:
NAME: ImageTextPairMSDataset
MODE: train
MS_DATASET_NAME: style_custom_dataset
MS_DATASET_NAMESPACE: damo
MS_DATASET_SUBNAME: 3D
PROMPT_PREFIX: ""
MS_DATASET_SPLIT: train_short
MS_REMAP_KEYS: { 'Image:FILE': 'Target:FILE' }
REPLACE_STYLE: False
PIN_MEMORY: True
BATCH_SIZE: 1
NUM_WORKERS: 4
SAMPLER:
NAME: LoopSampler
TRANSFORMS:
- NAME: LoadImageFromFile
RGB_ORDER: RGB
BACKEND: pillow
- NAME: FlexibleResize
INTERPOLATION: bicubic
SIZE: 1024
INPUT_KEY: [ 'img' ]
OUTPUT_KEY: [ 'img' ]
BACKEND: pillow
- NAME: FlexibleCropXL
INPUT_KEY: [ 'img' ]
OUTPUT_KEY: [ 'img' ]
BACKEND: pillow
- NAME: ImageToTensor
INPUT_KEY: [ 'img' ]
OUTPUT_KEY: [ 'img' ]
BACKEND: pillow
- NAME: Normalize
MEAN: [ 0.5, 0.5, 0.5 ]
STD: [ 0.5, 0.5, 0.5 ]
INPUT_KEY: [ 'img' ]
OUTPUT_KEY: [ 'img' ]
BACKEND: torchvision
- NAME: Select
KEYS: [ 'img', 'prompt', 'img_original_size_as_tuple', 'img_target_size_as_tuple', 'img_crop_coords_top_left' ]
META_KEYS: [ 'data_key', 'img_path' ]
- NAME: Rename
IN_KEYS: [ 'img', 'img_original_size_as_tuple', 'img_target_size_as_tuple', 'img_crop_coords_top_left' ]
OUT_KEYS: [ 'image', 'original_size_as_tuple', 'target_size_as_tuple', 'crop_coords_top_left' ]
#
EVAL_DATA:
NAME: ImageTextPairMSDataset
MODE: eval
MS_DATASET_NAME: style_custom_dataset
MS_DATASET_NAMESPACE: damo
MS_DATASET_SUBNAME: 3D
PROMPT_PREFIX: ""
MS_REMAP_KEYS: { 'Image': 'Target:FILE' }
MS_DATASET_SPLIT: test_short
OUTPUT_SIZE: [ 1024, 1024 ]
REPLACE_STYLE: False
PIN_MEMORY: True
BATCH_SIZE: 2
NUM_WORKERS: 4
FILE_SYSTEM:
NAME: "ModelscopeFs"
TEMP_DIR: "./cache/data"
#
TRANSFORMS:
- NAME: Select
KEYS: [ 'prompt' ]
META_KEYS: [ 'image_size' ]
#
TRAIN_HOOKS:
- NAME: BackwardHook
PRIORITY: 0
- NAME: LogHook
LOG_INTERVAL: 50
- NAME: CheckpointHook
INTERVAL: 1000
- NAME: ProbeDataHook
PROB_INTERVAL: 100
#
EVAL_HOOKS:
- NAME: ProbeDataHook
PROB_INTERVAL: 100
scepter/methods/examples/generation/stable_diffusion_xl_1024_lora.yaml
+337
View File
@@ -0,0 +1,337 @@
ENV:
BACKEND: nccl
SOLVER:
NAME: LatentDiffusionSolver
RESUME_FROM:
LOAD_MODEL_ONLY: True
USE_FSDP: False
SHARDING_STRATEGY:
USE_AMP: True
DTYPE: float16
CHANNELS_LAST: True
MAX_STEPS: 2000
MAX_EPOCHS: -1
NUM_FOLDS: 1
ACCU_STEP: 1
EVAL_INTERVAL: 100
#
WORK_DIR: ./cache/sdxl_1024_lora
LOG_FILE: std_log.txt
#
FILE_SYSTEM:
NAME: "ModelscopeFs"
TEMP_DIR: "./cache/data"
#
TUNER:
-
NAME: SwiftLoRA
R: 64
LORA_ALPHA: 64
LORA_DROPOUT: 0.0
BIAS: "none"
TARGET_MODULES: model.*(to_q|to_k|to_v|to_out.0|net.0.proj|net.2)$
#
MODEL:
NAME: LatentDiffusionXL
PARAMETERIZATION: eps
TIMESTEPS: 1000
MIN_SNR_GAMMA:
ZERO_TERMINAL_SNR: False
PRETRAINED_MODEL: ms://AI-ModelScope/stable-diffusion-xl-base-1.0@sd_xl_base_1.0.safetensors
IGNORE_KEYS: [ ]
SCALE_FACTOR: 0.13025
SIZE_FACTOR: 8
# DEFAULT_N_PROMPT: 'lowres, error, worst quality, low quality, jpeg artifacts, ugly, duplicate, morbid, mutilated, out of frame, extra fingers, mutated hands, poorly drawn hands, poorly drawn face, mutation, deformed, blurry, dehydrated, bad anatomy, bad proportions, extra limbs, cloned face, disfigured, gross proportions, malformed limbs, missing arms, missing legs, extra arms, extra legs, fused fingers, too many fingers, long neck, username, watermark, signature'
DEFAULT_N_PROMPT:
SCHEDULE_ARGS:
"NAME": "scaled_linear"
"BETA_MIN": 0.00085
"BETA_MAX": 0.0120
USE_EMA: False
LOAD_REFINER: False
#
DIFFUSION_MODEL:
NAME: DiffusionUNetXL
PRETRAINED_MODEL:
IN_CHANNELS: 4
OUT_CHANNELS: 4
NUM_RES_BLOCKS: 2
MODEL_CHANNELS: 320
ATTENTION_RESOLUTIONS: [ 4, 2 ]
DROPOUT: 0
CHANNEL_MULT: [ 1, 2, 4 ]
CONV_RESAMPLE: True
DIMS: 2
NUM_CLASSES: sequential
USE_CHECKPOINT: False
NUM_HEADS: -1
NUM_HEADS_CHANNELS: 64
USE_SCALE_SHIFT_NORM: False
RESBLOCK_UPDOWN: False
USE_NEW_ATTENTION_ORDER: True
USE_SPATIAL_TRANSFORMER: True
TRANSFORMER_DEPTH: [ 1, 2, 10 ]
CONTEXT_DIM: 2048
DISABLE_MIDDLE_SELF_ATTN: False
USE_LINEAR_IN_TRANSFORMER: True
ADM_IN_CHANNELS: 2816
USE_SENTENCE_EMB: False
USE_WORD_MAPPING: False
#
FIRST_STAGE_MODEL:
NAME: AutoencoderKL
EMBED_DIM: 4
PRETRAINED_MODEL:
IGNORE_KEYS: []
BATCH_SIZE: 1
#
ENCODER:
NAME: Encoder
CH: 128
OUT_CH: 3
NUM_RES_BLOCKS: 2
IN_CHANNELS: 3
ATTN_RESOLUTIONS: [ ]
CH_MULT: [ 1, 2, 4, 4 ]
Z_CHANNELS: 4
DOUBLE_Z: True
DROPOUT: 0.0
RESAMP_WITH_CONV: True
#
DECODER:
NAME: Decoder
CH: 128
OUT_CH: 3
NUM_RES_BLOCKS: 2
IN_CHANNELS: 3
ATTN_RESOLUTIONS: [ ]
CH_MULT: [ 1, 2, 4, 4 ]
Z_CHANNELS: 4
DROPOUT: 0.0
RESAMP_WITH_CONV: True
GIVE_PRE_END: False
TANH_OUT: False
#
COND_STAGE_MODEL:
NAME: GeneralConditioner
PRETRAINED_MODEL:
EMBEDDERS:
-
NAME: FrozenCLIPEmbedder
PRETRAINED_MODEL: ms://AI-ModelScope/clip-vit-large-patch14
TOKENIZER_PATH: ms://AI-ModelScope/clip-vit-large-patch14
MAX_LENGTH: 77
FREEZE: True
LAYER: hidden
LAYER_IDX: 11
USE_FINAL_LAYER_NORM: False
IS_TRAINABLE: False
UCG_RATE: 0.0
INPUT_KEYS: [ "prompt" ]
LEGACY_UCG_VALUE:
-
NAME: FrozenOpenCLIPEmbedder2
ARCH: ViT-bigG-14
PRETRAINED_MODEL:
MAX_LENGTH: 77
FREEZE: True
ALWAYS_RETURN_POOLED: True
LEGACY: False
LAYER: penultimate
IS_TRAINABLE: False
UCG_RATE: 0.0
INPUT_KEYS: [ "prompt" ]
LEGACY_UCG_VALUE:
-
NAME: ConcatTimestepEmbedderND
OUT_DIM: 256
IS_TRAINABLE: False
UCG_RATE: 0.0
INPUT_KEYS: [ "original_size_as_tuple" ]
LEGACY_UCG_VALUE:
-
NAME: ConcatTimestepEmbedderND
OUT_DIM: 256
IS_TRAINABLE: False
UCG_RATE: 0.0
INPUT_KEYS: [ "crop_coords_top_left" ]
LEGACY_UCG_VALUE:
-
NAME: ConcatTimestepEmbedderND
OUT_DIM: 256
IS_TRAINABLE: False
UCG_RATE: 0.0
INPUT_KEYS: [ "target_size_as_tuple" ]
LEGACY_UCG_VALUE:
#
REFINER_MODEL:
NAME: DiffusionUNetXL
PRETRAINED_MODEL:
IN_CHANNELS: 4
OUT_CHANNELS: 4
NUM_RES_BLOCKS: 2
MODEL_CHANNELS: 384
ATTENTION_RESOLUTIONS: [ 4, 2 ]
DROPOUT: 0
CHANNEL_MULT: [ 1, 2, 4, 4 ]
CONV_RESAMPLE: True
DIMS: 2
NUM_CLASSES: sequential
USE_CHECKPOINT: False
NUM_HEADS: -1
NUM_HEADS_CHANNELS: 64
USE_SCALE_SHIFT_NORM: False
RESBLOCK_UPDOWN: False
USE_NEW_ATTENTION_ORDER: True
USE_SPATIAL_TRANSFORMER: True
TRANSFORMER_DEPTH: 4
CONTEXT_DIM: [ 1280, 1280, 1280, 1280 ]
DISABLE_MIDDLE_SELF_ATTN: False
USE_LINEAR_IN_TRANSFORMER: True
ADM_IN_CHANNELS: 2560
USE_SENTENCE_EMB: False
USE_WORD_MAPPING: False
#
REFINER_COND_MODEL:
NAME: GeneralConditioner
PRETRAINED_MODEL:
EMBEDDERS:
-
NAME: FrozenOpenCLIPEmbedder2
ARCH: ViT-bigG-14
PRETRAINED_MODEL:
MAX_LENGTH: 77
FREEZE: True
ALWAYS_RETURN_POOLED: True
LEGACY: False
LAYER: penultimate
IS_TRAINABLE: False
UCG_RATE: 0.0
INPUT_KEYS: [ "prompt" ]
LEGACY_UCG_VALUE:
-
NAME: ConcatTimestepEmbedderND
OUT_DIM: 256
IS_TRAINABLE: False
UCG_RATE: 0.0
INPUT_KEYS: [ "original_size_as_tuple" ]
LEGACY_UCG_VALUE:
-
NAME: ConcatTimestepEmbedderND
OUT_DIM: 256
IS_TRAINABLE: False
UCG_RATE: 0.0
INPUT_KEYS: [ "crop_coords_top_left" ]
LEGACY_UCG_VALUE:
-
NAME: ConcatTimestepEmbedderND
OUT_DIM: 256
IS_TRAINABLE: False
UCG_RATE: 0.0
INPUT_KEYS: [ "aesthetic_score" ]
LEGACY_UCG_VALUE:
#
LOSS:
NAME: ReconstructLoss
LOSS_TYPE: l2
#
SAMPLE_ARGS:
SAMPLER: ddim
SAMPLE_STEPS: 50
SEED: 2023
GUIDE_SCALE: 5.0
GUIDE_RESCALE:
DISCRETIZATION: trailing
IMAGE_SIZE: [1024, 1024]
RUN_TRAIN_N: False
#
OPTIMIZER:
NAME: AdamW
LEARNING_RATE: 0.064
BETAS: [ 0.9, 0.999 ]
EPS: 1e-8
WEIGHT_DECAY: 1e-2
AMSGRAD: False
#
TRAIN_DATA:
NAME: ImageTextPairMSDataset
MODE: train
MS_DATASET_NAME: style_custom_dataset
MS_DATASET_NAMESPACE: damo
MS_DATASET_SUBNAME: 3D
PROMPT_PREFIX: ""
MS_DATASET_SPLIT: train_short
MS_REMAP_KEYS: { 'Image:FILE': 'Target:FILE' }
REPLACE_STYLE: False
PIN_MEMORY: True
BATCH_SIZE: 1
NUM_WORKERS: 4
SAMPLER:
NAME: LoopSampler
TRANSFORMS:
- NAME: LoadImageFromFile
RGB_ORDER: RGB
BACKEND: pillow
- NAME: FlexibleResize
INTERPOLATION: bicubic
SIZE: 1024
INPUT_KEY: [ 'img' ]
OUTPUT_KEY: [ 'img' ]
BACKEND: pillow
- NAME: FlexibleCropXL
INPUT_KEY: [ 'img' ]
OUTPUT_KEY: [ 'img' ]
BACKEND: pillow
- NAME: ImageToTensor
INPUT_KEY: [ 'img' ]
OUTPUT_KEY: [ 'img' ]
BACKEND: pillow
- NAME: Normalize
MEAN: [ 0.5, 0.5, 0.5 ]
STD: [ 0.5, 0.5, 0.5 ]
INPUT_KEY: [ 'img' ]
OUTPUT_KEY: [ 'img' ]
BACKEND: torchvision
- NAME: Select
KEYS: [ 'img', 'prompt', 'img_original_size_as_tuple', 'img_target_size_as_tuple', 'img_crop_coords_top_left' ]
META_KEYS: [ 'data_key', 'img_path' ]
- NAME: Rename
IN_KEYS: [ 'img', 'img_original_size_as_tuple', 'img_target_size_as_tuple', 'img_crop_coords_top_left' ]
OUT_KEYS: [ 'image', 'original_size_as_tuple', 'target_size_as_tuple', 'crop_coords_top_left' ]
#
EVAL_DATA:
NAME: ImageTextPairMSDataset
MODE: eval
MS_DATASET_NAME: style_custom_dataset
MS_DATASET_NAMESPACE: damo
MS_DATASET_SUBNAME: 3D
PROMPT_PREFIX: ""
MS_REMAP_KEYS: { 'Image': 'Target:FILE' }
MS_DATASET_SPLIT: test_short
OUTPUT_SIZE: [ 1024, 1024 ]
REPLACE_STYLE: False
PIN_MEMORY: True
BATCH_SIZE: 4
NUM_WORKERS: 4
FILE_SYSTEM:
NAME: "ModelscopeFs"
TEMP_DIR: "./cache/data"
#
TRANSFORMS:
- NAME: Select
KEYS: [ 'prompt' ]
META_KEYS: [ 'image_size' ]
#
TRAIN_HOOKS:
- NAME: BackwardHook
PRIORITY: 0
- NAME: LogHook
LOG_INTERVAL: 50
- NAME: CheckpointHook
INTERVAL: 1000
- NAME: ProbeDataHook
PROB_INTERVAL: 100
#
EVAL_HOOKS:
- NAME: ProbeDataHook
PROB_INTERVAL: 100