data/config.yaml

task_name: train
tags:
- dev
train: true
test: false
ckpt_path: null
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: AniMerPlus
SMAL:
  DATA_DIR: data/
  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
  FOCAL_LENGTH: 1000
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
GENERAL:
  TOTAL_STEPS: 1400000
  LOG_STEPS: 10000
  VAL_STEPS: 3000
  VAL_EPOCHS: 100
  CHECKPOINT_EPOCHS: 20
  CHECKPOINT_SAVE_TOP_K: 1
  NUM_WORKERS: 64
  PREFETCH_FACTOR: 2
LOSS_WEIGHTS:
  AVES:
    KEYPOINTS_3D: 0.05
    KEYPOINTS_2D: 0.01
    GLOBAL_ORIENT: 0.001
    POSE: 0.001
    BETAS: 0.0005
    BONE: 0.002
    ADVERSARIAL: 0.0
    MASK: 0.02
    POSE_RE: 0.05
    BETAS_RE: 0.001
    BONE_RE: 0.1
  SMAL:
    KEYPOINTS_3D: 0.05
    KEYPOINTS_2D: 0.01
    GLOBAL_ORIENT: 0.001
    POSE: 0.001
    BETAS: 0.0005
  SUPCON: 0.0005
TRAIN:
  LR: 1.25e-06
  WEIGHT_DECAY: 0.0001
  BATCH_SIZE: 16
  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: vithmoe
    PRETRAINED_WEIGHTS: data/vitmoe.pth
  AVES_HEAD:
    TYPE: transformer_decoder
    IN_CHANNELS: 2048
    IEF_ITERS: 1
    TRANSFORMER_DECODER:
      depth: 6
      heads: 8
      mlp_dim: 1024
      dim_head: 64
      dropout: 0.0
      emb_dropout: 0.0
      norm: layer
      context_dim: 1280
  SMAL_HEAD:
    TYPE: transformer_decoder
    IN_CHANNELS: 2048
    IEF_ITERS: 1
    TRANSFORMER_DECODER:
      depth: 6
      heads: 8
      mlp_dim: 1024
      dim_head: 64
      dropout: 0.0
      emb_dropout: 0.0
      norm: layer
      context_dim: 1280
  CLASS_TOKEN_HEAD:
    embed_dim: 1280
    hidden_dim: 4096
    output_dim: 256
    num_layers: 3
    last_bn: true
AVES:
  FOCAL_LENGTH: 2167
  MODEL_PATH: data/aves/aves_high_res.pt
  POSE_PRIOR_PATH: data/aves/aves_high_res.pt
  NUM_JOINTS: 24