configs/default.yaml

SEED_VALUE: 1234 # Seed value
DEBUG: True # Debug mode
FULL_CONFIG: false

TRAIN:
  SPLIT: 'train' # Training split name
  NUM_WORKERS: 8 # Number of workers
  BATCH_SIZE: 8 # Size of batches
  END_EPOCH: 2000 # End epoch

  RESUME: '' # Experiment path to be resumed training
  PRETRAINED_VAE: '' # Pretrained vae/vqvae model path
  PRETRAINED: '' # Pretrained model path

  OPTIM:
    target: AdamW
    params:
      lr: 2e-4
      betas: [0.9, 0.99]
      weight_decay: 0.0

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

EVAL:
  SPLIT: 'val' # Validation split name
  BATCH_SIZE: 16 # Validation Batch size
  NUM_WORKERS: 8 # Validation Batch size

TEST:
  CHECKPOINTS: '' # Pretrained model path
  SPLIT: 'test' # Testing split name
  BATCH_SIZE: 16 # Testing Batch size
  NUM_WORKERS: 8 # Testing Batch size

  SAVE_PREDICTIONS: False # Weather to save predictions
  COUNT_TIME: False # Weather to count time during test
  REPLICATION_TIMES: 20 # Number of times to replicate the test
  REP_I: 0 # For counting replication times

model:
  target: mGPT.models.mgpt.MotionGPT
  params:
    condition: 'text'
    task: 't2m'
    lm: ${lm.default}
    motion_vae: ${vq.default}

    # Related parameters
    stage: ${TRAIN.STAGE}
    debug: ${DEBUG}
    codebook_size: ${model.params.motion_vae.params.code_num}
    metrics_dict: ${METRIC.TYPE}

LOSS:
  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

  LAMBDA_VELOCITY: 0.5 # Lambda for velocity losses
  LAMBDA_COMMIT: 0.02 # Lambda for commitment losses

  ABLATION:
    RECONS_LOSS: 'l1_smooth'

METRIC:
  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
  TM2T: ${evaluator.tm2t}

DATASET:
  target: mGPT.data.HumanML3D.HumanML3DDataModule
  CODE_PATH: 'VQVAE'
  TASK_ROOT: ''
  TASK_PATH: ''
  NFEATS: 263
  KIT:
    MAX_MOTION_LEN: 196
    MIN_MOTION_LEN: 24
    MAX_TEXT_LEN: 20
    PICK_ONE_TEXT: true
    FRAME_RATE: 12.5
    UNIT_LEN: 4
  HUMANML3D:
    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

LOGGER:
  VAL_EVERY_STEPS: 10
  LOGGERS: ['tensorboard', 'wandb']
  TENSORBOARD:
    target: pytorch_lightning.loggers.TensorBoardLogger
    params:
      save_dir: ${FOLDER_EXP}
      name: 'tensorboard'
      version: ''
  WANDB:
    target: pytorch_lightning.loggers.WandbLogger
    params:
      project: null
      offline: False
      id: null
      version: ''
      name: ${NAME}
      save_dir: ${FOLDER_EXP}