liyang
ae28d55f81
Source: RGB-D -> Dyn-HaMR -> L20 retargeting -> FoundationPose -> reference repair -> SPIDER, documented in docs/PIPELINE_LATEST.md and docs/SETUP_AND_WEIGHTS.md. Adds FoundationPose and nvdiffrast upstream snapshots, requirements/pipeline_venv.txt and the FoundationPose weight manifest/downloader. Assets (Git LFS): weights/ (WiLoR detector, HandFlow denoiser, UniDepth-L), FoundationPose checkpoints, HaMeR checkpoint, Dyn-HaMR HMP model and BMC constraints, L20 URDF/meshes, the 20260915_171525 D405 recording and the two box CADs. MANO models are not redistributed (third_party/hamer/_DATA/data/mano/README.txt). Environments, caches and run outputs excluded. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
112 lines
2.0 KiB
YAML
112 lines
2.0 KiB
YAML
task_name: train
|
|
tags:
|
|
- dev
|
|
train: true
|
|
test: false
|
|
ckpt_path: null
|
|
seed: null
|
|
DATASETS:
|
|
TRAIN:
|
|
FREIHAND-TRAIN:
|
|
WEIGHT: 0.25
|
|
INTERHAND26M-TRAIN:
|
|
WEIGHT: 0.25
|
|
MTC-TRAIN:
|
|
WEIGHT: 0.1
|
|
RHD-TRAIN:
|
|
WEIGHT: 0.05
|
|
COCOW-TRAIN:
|
|
WEIGHT: 0.1
|
|
HALPE-TRAIN:
|
|
WEIGHT: 0.05
|
|
MPIINZSL-TRAIN:
|
|
WEIGHT: 0.05
|
|
HO3D-TRAIN:
|
|
WEIGHT: 0.05
|
|
H2O3D-TRAIN:
|
|
WEIGHT: 0.05
|
|
DEX-TRAIN:
|
|
WEIGHT: 0.05
|
|
VAL:
|
|
FREIHAND-TRAIN:
|
|
WEIGHT: 1.0
|
|
MOCAP: FREIHAND-MOCAP
|
|
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
|
|
extras:
|
|
ignore_warnings: false
|
|
enforce_tags: true
|
|
print_config: true
|
|
exp_name: hamer
|
|
MANO:
|
|
DATA_DIR: _DATA/data/
|
|
MODEL_PATH: data/mano
|
|
GENDER: neutral
|
|
NUM_HAND_JOINTS: 15
|
|
MEAN_PARAMS: 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: 1000000
|
|
LOG_STEPS: 1000
|
|
VAL_STEPS: 1000
|
|
CHECKPOINT_STEPS: 10000
|
|
CHECKPOINT_SAVE_TOP_K: 1
|
|
NUM_WORKERS: 8
|
|
PREFETCH_FACTOR: 2
|
|
TRAIN:
|
|
LR: 1.0e-05
|
|
WEIGHT_DECAY: 0.0001
|
|
BATCH_SIZE: 32
|
|
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
|
|
PRETRAINED_WEIGHTS: hamer_training_data/vitpose_backbone.pth
|
|
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
|
|
LOSS_WEIGHTS:
|
|
KEYPOINTS_3D: 0.05
|
|
KEYPOINTS_2D: 0.01
|
|
GLOBAL_ORIENT: 0.001
|
|
HAND_POSE: 0.001
|
|
BETAS: 0.0005
|
|
ADVERSARIAL: 0.0005
|