HaWoR / hawor /model_config.yaml
ThunderVVV's picture
update
4c44a98
task_name: train
tags:
- dev
train: true
auto_resume: 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: false
POSES_BETAS_SIMULTANEOUS: true
FILTER_NO_POSES: false
TRAIN:
MULTI_SET: true
HOT3D_TRAIN:
WEIGHT: 1.0
ROOT: hot3d_trainset_export
PKL: train_set_new.pkl
ARCTIC_TRAIN:
WEIGHT: 1.0
ROOT: arctic_trainset_export
PKL: train_set_new.pkl
DEXYCB_TRAIN:
WEIGHT: 1.0
ROOT: dex_ycb_trainset_export
PKL: train_set_new.pkl
HO3D_TRAIN:
WEIGHT: 1.0
ROOT: ho3d_trainset_export
PKL: train_set_new.pkl
VAL:
HOT3D_TEST:
WEIGHT: 1.0
BETAS_REG: true
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
trainer:
_target_: pytorch_lightning.Trainer
default_root_dir:
accelerator: gpu
devices: 4
deterministic: false
num_sanity_val_steps: 0
log_every_n_steps: 1000
val_check_interval: 0
precision: 16
max_steps: 250000
limit_val_batches: 0
paths:
root_dir:
data_dir:
log_dir: logs/
output_dir:
work_dir:
extras:
ignore_warnings: false
enforce_tags: true
print_config: true
exp_name:
MANO:
DATA_DIR: _DATA/data/
MODEL_PATH: _DATA/data//mano
GENDER: neutral
NUM_HAND_JOINTS: 15
MEAN_PARAMS: _DATA/data//mano_mean_params.npz
CREATE_BODY_POSE: false
EXTRA:
FOCAL_LENGTH: 5000
NUM_LOG_IMAGES: 4
NUM_LOG_SAMPLES_PER_IMAGE: 8
PELVIS_IND: 0
GENERAL:
TOTAL_STEPS: 250000
LOG_STEPS: 1000
VAL_STEPS: 0
CHECKPOINT_STEPS: 1000
CHECKPOINT_SAVE_TOP_K: 1
NUM_WORKERS: 25
PREFETCH_FACTOR: 2
TRAIN:
LR: 1.0e-05
WEIGHT_DECAY: 0.0001
BATCH_SIZE: 8
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: vit
TORCH_COMPILE: 0
PRETRAINED_WEIGHTS:
MANO_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
ST_MODULE: true
MOTION_MODULE: true
ST_HDIM: 512
MOTION_HDIM: 384
ST_NLAYER: 6
MOTION_NLAYER: 6
LOSS_WEIGHTS:
KEYPOINTS_3D: 0.05
KEYPOINTS_2D: 0.01
GLOBAL_ORIENT: 0.001
HAND_POSE: 0.001
BETAS: 0.0005