Add BiCAT w/o CA AT (latest, step 300000)

BiCAT/CA_ablation/AT/README.md

@@ -0,0 +1,10 @@
+---
+tags:
+- model_hub_mixin
+- pytorch_model_hub_mixin
+---
+
+This model has been pushed to the Hub using the [PytorchModelHubMixin](https://huggingface.co/docs/huggingface_hub/package_reference/mixins#huggingface_hub.PyTorchModelHubMixin) integration:
+- Code: [More Information Needed]
+- Paper: [More Information Needed]
+- Docs: [More Information Needed]

BiCAT/CA_ablation/AT/config.json

@@ -0,0 +1,57 @@
+{
+  "action_dim": 14,
+  "action_horizon": 32,
+  "arm_dim": 7,
+  "clamp_normalized_actions": false,
+  "conv_kernel_sizes": [
+    5,
+    3
+  ],
+  "conv_strides": [
+    2,
+    2
+  ],
+  "decoder_nhead": 8,
+  "decoder_nlayer": 6,
+  "dropout": 0.1,
+  "fsq_levels": [
+    8,
+    8,
+    6,
+    5
+  ],
+  "latent_dim": 4,
+  "latent_horizon_per_arm": 8,
+  "model_dim": 256,
+  "name": "btt",
+  "num_heads": 8,
+  "right_arm_sign": [
+    -1.0,
+    1.0,
+    1.0,
+    1.0,
+    -1.0,
+    -1.0,
+    1.0
+  ],
+  "shared_action_max": [
+    2.945596,
+    3.3889313,
+    3.4514999,
+    1.9530925,
+    1.6339467,
+    3.0108311,
+    1.0
+  ],
+  "shared_action_min": [
+    -1.3947148,
+    -0.5742835,
+    -0.1190902,
+    -1.9893787,
+    -1.4216828,
+    -3.4687879,
+    0.0
+  ],
+  "use_causal_decoder": true,
+  "use_cross_attention": false
+}

BiCAT/CA_ablation/AT/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:454f2f1bbfc1d58885fd9f44fa05b8538604bab9c8dd1db4d8dca0c6bc1d2770
+size 28556932

BiCAT/CA_ablation/AT/norm_stats.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4b9e7e69a4cb1bf67acc655cba23f329c1a6125d8c74ee1b37c19bd2c54a56e5
+size 3957

BiCAT/CA_ablation/AT/policy_io.json

@@ -0,0 +1,15 @@
+{
+  "obs_features": [],
+  "action_feature": {
+    "key": "action",
+    "type": "action",
+    "shape": [
+      14
+    ],
+    "norm_mode": "min_max"
+  },
+  "n_obs_steps": 0,
+  "n_action_steps": 32,
+  "fps": 15,
+  "language_key": null
+}

BiCAT/CA_ablation/AT/train_config.yaml

@@ -0,0 +1,129 @@
+codec:
+  name: btt
+  kind: tokens
+  pretrained_path: null
+  action_dim: 14
+  arm_dim: 7
+  action_horizon: 32
+  latent_horizon_per_arm: 8
+  max_fit_chunks: 50000
+  model_dim: 256
+  num_heads: 8
+  decoder_nhead: 8
+  decoder_nlayer: 6
+  dropout: 0.1
+  use_causal_decoder: true
+  conv_kernel_sizes:
+  - 5
+  - 3
+  conv_strides:
+  - 2
+  - 2
+  latent_dim: 4
+  fsq_levels:
+  - 8
+  - 8
+  - 6
+  - 5
+  right_arm_sign:
+  - -1
+  - 1
+  - 1
+  - 1
+  - -1
+  - -1
+  - 1
+  shared_action_min:
+  - -1.3947148
+  - -0.5742835
+  - -0.1190902
+  - -1.9893787
+  - -1.4216828
+  - -3.4687879
+  - 0.0
+  shared_action_max:
+  - 2.945596
+  - 3.3889313
+  - 3.4514999
+  - 1.9530925
+  - 1.6339467
+  - 3.0108311
+  - 1.0
+  clamp_normalized_actions: false
+  use_cross_attention: false
+dataset:
+  name: robotwin_new_demo_clean
+  repo_id: local/robotwin_new_demo_clean
+  root: /scratch/xuanweiliu/bimanual_tokenizer/data/lerobot/local/robotwin_new_demo_clean
+  hf_datasets_cache: ${oc.env:HF_DATASETS_CACHE,null}
+  episodes: null
+  exclude_episodes_file: null
+  action_stats_file: null
+  observation_keys:
+  - observation.state
+  - observation.images.cam_high
+  - observation.images.cam_left_wrist
+  - observation.images.cam_right_wrist
+  video_backend: null
+  bypass_force_cache_sync: true
+trainer:
+  max_steps: 300000
+  batch_size: 512
+  global_batch_size: null
+  num_workers: 4
+  prefetch_factor: 4
+  use_episode_aware_sampler: true
+  drop_n_first_frames: 0
+  drop_n_last_frames: 0
+  weight_decay: 0.0
+  betas:
+  - 0.9
+  - 0.95
+  eps: 1.0e-08
+  lr_scheduler: constant
+  warmup_steps: 200
+  min_lr_ratio: 0.1
+  use_ema: true
+  ema_decay: 0.995
+  mixed_precision: bf16
+  gradient_accumulation_steps: 1
+  find_unused_parameters: false
+  max_grad_norm: 1.0
+  distributed_strategy: ddp
+  fsdp_min_num_params: 100000000
+  fsdp_reshard_after_forward: true
+  fsdp_cpu_offload: false
+  log_every: 200
+  val_every: 1000
+  checkpoint_every: 5000
+  checkpoint_dir: ${output_dir}/checkpoints
+  local_log_path: ${output_dir}/train_metrics.jsonl
+  val_ratio: 0.05
+  split_seed: ${seed}
+  val_batch_size: null
+  max_val_batches: 100
+  checkpoint_candidates:
+    best:
+      metric_key: val_reconst_loss
+      mode: min
+      k: 3
+    recent:
+      k: 2
+  use_wandb: true
+  wandb_project: tokenizer_pretrain
+  wandb_entity: evan_lxw-harbin-institute-of-technology
+  wandb_init_timeout: null
+  wandb_upload_eval_videos: true
+  wandb_max_eval_videos: 16
+  _trainer_type: codecs
+  lr: 5.0e-05
+  reconst_every: 1000
+  max_reconst_batches: 100
+  skip_codec_fit: false
+name: train_codecs
+seed: 42
+device: cuda
+run_name: robotwin_new_btt_wo_CA_300k
+output_dir: /scratch/Projects/CFP-05/CFP05-CF-002/BiCAT/sim/BiCAT/CA_ablation/AT
+run_mode: overwrite
+resume_checkpoint: null