upload configs/config.yaml

configs/config.yaml

@@ -0,0 +1,154 @@
+stage_1:
+  ckpt_path: /cache/witwicky/checkpoints/track4world_da3.pth
+  encoder_input_size: 224
+  normalization_stat_path: null
+da3_finetune:
+  enabled: true
+  freeze_blocks_before: 13
+  n_action_steps: 9
+  n_views: 2
+  use_temporal_embed: false
+  action_input_rate: 0.4
+action_head:
+  type: mlp_resnet
+  input_dim: 1536
+  hidden_dim: 1536
+  n_dims: 7
+  chunk_size: 8
+  chunk_position_encoding: learned
+  num_blocks: 2
+  pool_mode: mean
+regularization:
+  lambda_feat: 0.0
+  layer_weight_min: 0.5
+  adaptive_lambda: false
+  lambda_depth: 3.0
+  depth_grad_weight: 1.0
+  depth_decode_chunk_size: 8
+  lambda_camera: 0.0
+  teacher_depth_fallback: true
+  lambda_path_b_deep_feat_reg: 0.01
+  deep_feat_reg_layer_weight_min: 0.5
+proprioception:
+  proprio_dim: 7
+predictor:
+  enabled: true
+  type: shallow12_ar
+  d_model: 1024
+  depth: 12
+  num_heads: 16
+  ffn_ratio: 4.0
+  dropout: 0.0
+  num_patches_per_view: 256
+  use_language: true
+  language_encoder_type: t5
+  language_dim: 768
+  language_len: 77
+  clip_model: openai/clip-vit-large-patch14
+  t5_model: google-t5/t5-base
+  condition_mode: concat
+  input_proj_norm: ln
+  lambda_feat_future: 1.0
+  lambda_sigreg: 0.0
+  feature_loss_type: l1
+  H_choices:
+  - 8
+  H_weights:
+  - 1.0
+  num_register_tokens: 0
+  deep_gradient_checkpointing: true
+training:
+  global_batch_size: 12
+  micro_batch_size: 4
+  base_lr: 2.0e-05
+  head_lr_mult: 10.0
+  predictor_lr_mult: 10.0
+  adam_eps: 1.0e-06
+  adam_beta1: 0.9
+  adam_beta2: 0.95
+  weight_decay: 0.0
+  lambda_action: 3.0
+  lambda_action_direct: 0.0
+  lambda_action_refine: 1.0
+  grad_accum_steps: 1
+  epochs: 500000
+  clip_grad: 1.0
+  warmup_steps: 0
+  max_steps: 150000
+  min_lr_ratio: 1.0
+  log_every: 50
+  ckpt_every: 2000
+  vis_every: 500
+  eval_every: 999999
+  global_seed: 42
+  num_workers: 8
+  persistent_workers: true
+  bf16: true
+  compile: false
+dataset:
+  type: openx
+  openx_root: /cache/witwicky/openx_lerobot
+  stats_dir: /cache/witwicky/openx_lerobot/_stats
+  preset: null
+  datasets:
+  - name: bridge
+    weight: 2.0
+  - name: utaustin_mutex
+    weight: 1.0
+  - name: stanford_hydra_dataset
+    weight: 2.0
+  - name: berkeley_autolab_ur5
+    weight: 2.0
+  - name: austin_sailor_dataset
+    weight: 1.0
+  - name: austin_sirius_dataset
+    weight: 1.0
+  - name: berkeley_fanuc_manipulation
+    weight: 2.0
+  - name: jaco_play
+    weight: 1.0
+  - name: fmb_dataset
+    weight: 2.0
+  - name: fractal20220817_data
+    weight: 0.5
+  - name: berkeley_cable_routing
+    weight: 1.0
+  - name: roboturk
+    weight: 2.0
+  - name: dlr_edan_shared_control
+    weight: 1.0
+  - name: austin_buds_dataset
+    weight: 1.0
+  - name: nyu_franka_play_dataset
+    weight: 2.0
+  - name: nyu_door_opening_surprising_effectiveness
+    weight: 1.0
+  - name: cmu_stretch
+    weight: 1.0
+  - name: furniture_bench_dataset
+    weight: 0.1
+  - name: bc_z
+    weight: 0.2
+  image_size:
+  - 224
+  - 224
+  future_steps: 8
+  chunk_size: 8
+  aggregate_chunk_actions: false
+  target_hz: null
+  uniform_action_sampling: true
+  include_current_action: true
+  n_views: 2
+  proprio_dim: 7
+  normalize_proprio: true
+  download_videos: true
+  video_backend: pyav
+  require_stats_timing_signature: true
+  action_stats_samples: 50000
+  proprio_stats_samples: 50000
+  train_crop_min_scale: 0.9
+  eval_crop_scale: 0.9
+  eval_ratio: 0.001
+  force_cache_sync: false
+  lazy_download: false
+use_dit: false