task_name: train
tags:
- dev
train: true
test: false
ckpt_path: true
seed: null
trainer:
  _target_: pytorch_lightning.Trainer
  default_root_dir: ${paths.output_dir}
  accelerator: gpu
  devices: 1
  deterministic: false
  num_sanity_val_steps: 0
  log_every_n_steps: ${GENERAL.LOG_STEPS}
  val_check_interval: ${GENERAL.VAL_STEPS}
  check_val_every_n_epoch: ${GENERAL.VAL_EPOCHS}
  precision: 16-mixed
  max_steps: ${GENERAL.TOTAL_STEPS}
  limit_val_batches: 80
paths:
  root_dir: ${oc.env:PROJECT_ROOT}
  data_dir: ${paths.root_dir}/data/
  log_dir: logs/
  output_dir: ${hydra:runtime.output_dir}
  work_dir: ${hydra:runtime.cwd}
extras:
  ignore_warnings: false
  enforce_tags: true
  print_config: true
exp_name: bestquad3dStage02
SMAL:
  DATA_DIR: data/smal
  MODEL_PATH: data/smal/my_smpl_00781_4_all.pkl
  SHAPE_PRIOR_PATH: data/smal/my_smpl_data_00781_4_all.pkl
  POSE_PRIOR_PATH: data/smal/walking_toy_symmetric_pose_prior_with_cov_35parts.pkl
  NUM_JOINTS: 34
EXTRA:
  FOCAL_LENGTH: 1000
  NUM_LOG_IMAGES: 4
  NUM_LOG_SAMPLES_PER_IMAGE: 4
  PELVIS_IND: 0
DATASETS:
  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: false
    FLIP_AUG_RATE: 0.0
    EXTREME_CROP_AUG_RATE: 0.0
    EXTREME_CROP_AUG_LEVEL: 1
  ANIMAL3D:
    ROOT_IMAGE: ./datasets/animal3d/
    JSON_FILE:
      TRAIN: ./datasets/animal3d/train.json
      TEST: ./datasets/animal3d/test.json
    WEIGHT: 1.0
  CONTROL_ANIMAL3D:
    ROOT_IMAGE: ./datasets/control_animal3dlatest/
    JSON_FILE:
      TRAIN: ./datasets/control_animal3dlatest/train.json
      TEST: ./datasets/control_animal3dlatest/test.json
    WEIGHT: 0.5
  QUADRUPED2D:
    ROOT_IMAGE: ./datasets/quadruped2d/
    JSON_FILE:
      TRAIN: ./datasets/quadruped2d/train3d_60filtered.json
      TEST: ./datasets/quadruped2d/test.json
    WEIGHT: 0.5
GENERAL:
  TOTAL_STEPS: 450000
  LOG_STEPS: 451
  VAL_STEPS: 451
  VAL_EPOCHS: 1
  CHECKPOINT_EPOCHS: 1
  CHECKPOINT_SAVE_TOP_K: 2
  NUM_WORKERS: 2
  PREFETCH_FACTOR: 2
LOSS_WEIGHTS:
  KEYPOINTS_3D: 0.05
  KEYPOINTS_2D: 0.01
  INTERMEDIATE_KP2D: 0.01
  INTERMEDIATE_KP3D: 0.01
  GLOBAL_ORIENT: 0.005
  POSE: 0.001
  BETAS: 0.0005
  TRANSL: 0.0005
  ADVERSARIAL: 0.0
  SUPCON: 0.0005
TRAIN:
  LR: 3.75e-06
  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
MODEL:
  IMAGE_SIZE: 256
  IMAGE_MEAN:
  - 0.485
  - 0.456
  - 0.406
  IMAGE_STD:
  - 0.229
  - 0.224
  - 0.225
  BACKBONE:
    TYPE: vith
    PRETRAINED_WEIGHTS: ./data/amr_vitbb.pth
    FREEZE: false
  USE_BIOCLIP_EMBEDDING: true
  BIOCLIP_EMBEDDING:
    EMBED_DIM: 1280
    TYPE: bioclip1
  USE_KEYPOINT_EMBEDDING: false
  KEYPOINT_EMBEDDING:
    NUM_KEYPOINTS: 26
    KEYPOINT_DIM: 2
    EMBED_DIM: 1280
    HIDDEN_DIM: 512
    TYPE: token
  SMAL_HEAD:
    TYPE: new_bio_pose_transformer_decoder
    IN_CHANNELS: 1280
    IEF_ITERS: 1
    DECODER_DIM: 1280
    NUM_DECODER_LAYERS: 6
    NUM_HEADS: 8
    MLP_RATIO: 4.0
    USE_KEYPOINT_2D_TOKENS: true
    USE_KEYPOINT_3D_TOKENS: true
    KEYPOINT_TOKEN_UPDATE: true
    KP2D_INJECT_IMAGE_FEAT: true
    TRANSFORMER_DECODER:
      depth: 6
      heads: 8
      mlp_dim: 1024
      dim_head: 64
      dropout: 0.0
      emb_dropout: 0.0
      norm: layer
      context_dim: 1280