task_name: train
tags:
- dev
train: true
test: false
ckpt_path: null
seed: null
DATASETS:
  SUPPRESS_KP_CONF_THRESH: 0.3
  FILTER_NUM_KP: 4
  FILTER_NUM_KP_THRESH: 0.0
  FILTER_REPROJ_THRESH: 31000
  SUPPRESS_BETAS_THRESH: 3.0
  SUPPRESS_BAD_POSES: true
  POSES_BETAS_SIMULTANEOUS: true
  FILTER_NO_POSES: false
  TRAIN:
    H36M-TRAIN-WMASK:
      WEIGHT: 0.1
    MPII-TRAIN-WMASK:
      WEIGHT: 0.1
    COCO-TRAIN-2014-WMASK-PRUNED:
      WEIGHT: 0.1
    COCO-TRAIN-2014-VITPOSE-REPLICATE-PRUNED12:
      WEIGHT: 0.1
    MPI-INF-TRAIN-PRUNED:
      WEIGHT: 0.02
    AVA-TRAIN-MIDFRAMES-1FPS-WMASK:
      WEIGHT: 0.19
    AIC-TRAIN-WMASK:
      WEIGHT: 0.19
    INSTA-TRAIN-WMASK:
      WEIGHT: 0.2
  VAL:
    COCO-VAL:
      WEIGHT: 1.0
  MOCAP: CMU-MOCAP
  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: true
    FLIP_AUG_RATE: 0.5
    EXTREME_CROP_AUG_RATE: 0.1
    EXTREME_CROP_AUG_LEVEL: 1
  BETAS_REG: true
trainer:
  _target_: pytorch_lightning.Trainer
  default_root_dir: ${paths.output_dir}
  accelerator: gpu
  devices: 8
  deterministic: false
  num_sanity_val_steps: 0
  log_every_n_steps: ${GENERAL.LOG_STEPS}
  val_check_interval: ${GENERAL.VAL_STEPS}
  precision: 16
  max_steps: ${GENERAL.TOTAL_STEPS}
  move_metrics_to_cpu: true
  limit_val_batches: 1
  track_grad_norm: -1
  strategy: ddp
  num_nodes: 1
  sync_batchnorm: true
paths:
  root_dir: ${oc.env:PROJECT_ROOT}
  data_dir: ${paths.root_dir}/data/
  log_dir: /fsx/shubham/code/hmr2023/logs_hydra/
  output_dir: ${hydra:runtime.output_dir}
  work_dir: ${hydra:runtime.cwd}
extras:
  ignore_warnings: false
  enforce_tags: true
  print_config: true
exp_name: hmr2
SMPL:
  MODEL_PATH: data/smpl
  GENDER: neutral
  NUM_BODY_JOINTS: 23
  JOINT_REGRESSOR_EXTRA: data/SMPL_to_J19.pkl
  MEAN_PARAMS: data/smpl_mean_params.npz
EXTRA:
  FOCAL_LENGTH: 5000
  NUM_LOG_IMAGES: 4
  NUM_LOG_SAMPLES_PER_IMAGE: 8
  PELVIS_IND: 39
MODEL:
  IMAGE_SIZE: 256
  IMAGE_MEAN:
  - 0.485
  - 0.456
  - 0.406
  IMAGE_STD:
  - 0.229
  - 0.224
  - 0.225
  BACKBONE:
    TYPE: vit
    NUM_LAYERS: 50
    OUT_CHANNELS: 2048
  ADD_NECK: false
  FLOW:
    DIM: 144
    NUM_LAYERS: 4
    CONTEXT_FEATURES: 2048
    LAYER_HIDDEN_FEATURES: 1024
    LAYER_DEPTH: 2
  FC_HEAD:
    NUM_FEATURES: 1024
  SMPL_HEAD:
    TYPE: transformer_decoder
    IN_CHANNELS: 2048
    TRANSFORMER_DECODER:
      depth: 6
      heads: 8
      mlp_dim: 1024
      dim_head: 64
      dropout: 0.0
      emb_dropout: 0.0
      norm: layer
      context_dim: 1280
GENERAL:
  TOTAL_STEPS: 1000000
  LOG_STEPS: 1000
  VAL_STEPS: 1000
  CHECKPOINT_STEPS: 10000
  CHECKPOINT_SAVE_TOP_K: 1
  NUM_WORKERS: 6
  PREFETCH_FACTOR: 2
TRAIN:
  LR: 1.0e-05
  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
LOSS_WEIGHTS:
  KEYPOINTS_3D: 0.05
  KEYPOINTS_2D: 0.01
  GLOBAL_ORIENT: 0.001
  BODY_POSE: 0.001
  BETAS: 0.0005
  ADVERSARIAL: 0.0005
local: {}