Fannovel16-ComfyUI-MotionDiff/motiondiff_modules/mGPT/configs/inference.yaml

NAME: Webui # Experiment name
DEBUG: False # Debug mode
ACCELERATOR: 'cpu' # Devices optioncal: “cpu”, “gpu”, “tpu”, “ipu”, “hpu”, “mps, “auto”
DEVICE: [0] # Index of gpus eg. [0] or [0,1,2,3]

# Training configuration
TRAIN:
  #---------------------------------
  STAGE: lm_instruct
  DATASETS: ['humanml3d'] # Training datasets
  SPLIT: 'train' # Training split name
  NUM_WORKERS: 32 # Number of workers
  BATCH_SIZE: 16 # Size of batches
  START_EPOCH: 0 # Start epochMMOTIONENCODER
  END_EPOCH: 99999 # End epoch
  ABLATION:
    pkeep: 0.5
  OPTIM:
    TYPE: AdamW # Optimizer type
    LR: 2e-4 # Learning rate
    WEIGHT_DECAY: 0.0
    LR_SCHEDULER: [100, 200, 300, 400]
    GAMMA: 0.8

  LR_SCHEDULER:
    target: CosineAnnealingLR
    params:
      T_max: ${eval:${LOGGER.VAL_EVERY_STEPS} * 100}
      eta_min: 1e-6

# Evaluating Configuration
EVAL:
  DATASETS: ['humanml3d'] # Evaluating datasets
  BATCH_SIZE: 32 # Evaluating Batch size
  NUM_WORKERS: 8 # Validation Batch size
  SPLIT: test

# Test Configuration
TEST:
  CHECKPOINTS: checkpoints/MotionGPT-base/motiongpt_s3_h3d.ckpt
  DATASETS: ['humanml3d'] # training datasets
  SPLIT: test
  BATCH_SIZE: 32 # training Batch size
  MEAN: False
  NUM_SAMPLES: 1
  FACT: 1

# Datasets Configuration
DATASET:
  target: .mGPT.data.HumanML3D.HumanML3DDataModule
  JOINT_TYPE: 'humanml3d' # join type
  CODE_PATH: 'VQBEST'
  TASK_ROOT: .deps/mGPT_instructions
  TASK_PATH: ''
  SMPL_PATH: .deps/smpl
  TRANSFORM_PATH: .deps/transforms/
  WORD_VERTILIZER_PATH: .deps/glove/
  NFEATS: 263
  KIT:
    ROOT: .datasets/kit-ml # KIT directory
    SPLIT_ROOT: .datasets/kit-ml # KIT splits directory
    MEAN_STD_PATH: .deps/t2m/
    MAX_MOTION_LEN: 196
    MIN_MOTION_LEN: 24
    MAX_TEXT_LEN: 20
    PICK_ONE_TEXT: true
    FRAME_RATE: 12.5
    UNIT_LEN: 4
  HUMANML3D:
    ROOT: .datasets/humanml3d # HumanML3D directory
    SPLIT_ROOT: .datasets/humanml3d # HumanML3D splits directory
    ASSETS_ROOT: .assets/meta/
    MEAN_STD_PATH: .deps/t2m/
    MAX_MOTION_LEN: 196
    MIN_MOTION_LEN: 40
    MAX_TEXT_LEN: 20
    PICK_ONE_TEXT: true
    FRAME_RATE: 20.0
    UNIT_LEN: 4
    STD_TEXT: False

ABLATION:
  # For MotionGPT
  use_length: False
  predict_ratio: 0.2
  inbetween_ratio: 0.25
  image_size: 256
  # For Motion-latent-diffusion
  VAE_TYPE: 'actor' # vae ablation: actor or mcross
  VAE_ARCH: 'encoder_decoder' # mdiffusion vae architecture
  PE_TYPE: 'actor' # mdiffusion mld or actor
  DIFF_PE_TYPE: 'actor' # mdiffusion mld or actor
  SKIP_CONNECT: False # skip connection for denoiser va
  MLP_DIST: False # use linear to expand mean and std rather expand token nums
  IS_DIST: False # Mcross distribution kl
  PREDICT_EPSILON: True # noise or motion

METRIC:
  TYPE: ['TM2TMetrics']
  TM2T:
   t2m_path: .deps/t2m/ # path for tm2t evaluator
  TASK: 't2m'
  FORCE_IN_METER: True
  DIST_SYNC_ON_STEP: True
  MM_NUM_SAMPLES: 100 # Number of samples for multimodal test
  MM_NUM_REPEATS: 30 # Number of repeats for multimodal test
  MM_NUM_TIMES: 10 # Number of times to repeat the multimodal test
  DIVERSITY_TIMES: 300 # Number of times to repeat the diversity test

# Losses Configuration
LOSS:
  TYPE: t2mgpt # Losses type
  LAMBDA_FEATURE: 1.0
  LAMBDA_VELOCITY: 0.5
  LAMBDA_COMMIT: 0.02
  LAMBDA_CLS: 1.0
  LAMBDA_M2T2M: 1.0
  LAMBDA_T2M2T: 10.0
  ABLATION:
    RECONS_LOSS: 'l1_smooth'
  LAMBDA_REC: 1.0 # Lambda for reconstruction losses
  LAMBDA_JOINT: 1.0 # Lambda for joint losses

  LAMBDA_LATENT: 1e-5 # Lambda for latent losses
  LAMBDA_KL: 1e-5 # Lambda for kl losses
  LAMBDA_GEN: 1.0 # Lambda for text-motion generation losses
  LAMBDA_CROSS: 1.0 # Lambda for cross-reconstruction losses
  LAMBDA_CYCLE: 1.0 # Lambda for cycle losses
  LAMBDA_PRIOR: 0.0 # Lambda for diffusion prior losses

# Model Configuration
model:
  target: .mGPT.models.mgpt.MotionGPT
  params:
    condition: 'text'
    task: 't2m'
    lm:
      target: .mGPT.archs.mgpt_lm.MLM
      params:
        model_type: t5
        model_path: google/flan-t5-base
        stage: ${TRAIN.STAGE}
        motion_codebook_size: 512
        ablation: ${ABLATION}
    motion_vae:
      target: .mGPT.archs.mgpt_vq.VQVae
      params:
        quantizer: 'ema_reset'
        code_num: 512
        code_dim: 512
        output_emb_width: 512
        down_t: 2
        stride_t: 2
        width: 512
        depth: 3
        dilation_growth_rate: 3
        norm: None
        activation: 'relu'
        nfeats: ${DATASET.NFEATS}
        ablation: ${ABLATION}

    # Related parameters
    stage: ${TRAIN.STAGE}
    debug: ${DEBUG}
    codebook_size: 512
    metrics_dict: ${METRIC.TYPE}

# Logger configuration
LOGGER:
  LOG_EVERY_STEPS: 5
  VAL_EVERY_STEPS: 10
  TENSORBOARD: True
  wandb:
    params:
      project: null