Upload config.yaml with huggingface_hub

config.yaml

@@ -0,0 +1,195 @@
+run_id: 0213_16-3d-2d-1d-lr1e5-3d_loss_1e1-decay1e3-train_data-qwen_freeze-bz_4_8
+run_root_dir: /path/to/navsim_exp
+seed: 42
+trackers:
+- jsonl
+- wandb
+wandb_entity: kmzy99
+wandb_project: vla
+is_debug: false
+ver_1225: 1
+act_tok: 8
+pretrain_model_2d: null
+enable_image_aug: 0
+doing_v_pre: 1
+w_depth: 1
+doing_s2: 0
+w_video_latent: 0
+vit_pre: 0
+rgb_query_loss: 1
+gs_query_loss: 1
+framework:
+  name: QwenOFT
+  qwenvl:
+    base_vlm: /path/to/Qwen3-VL-2B-WorldAction  # overridden by BASE_VLM in env.sh at inference
+    attn_implementation: flash_attention_2
+    vl_hidden_dim: 2048
+  dino:
+    dino_backbone: dinov2_vits14
+  action_model:
+    action_model_type: DiT-B
+    hidden_size: 1536
+    add_pos_embed: true
+    max_seq_len: 1024
+    action_dim: 4
+    state_dim: 4
+    future_action_window_size: 8
+    action_horizon: 8
+    past_action_window_size: 0
+    repeated_diffusion_steps: 8
+    noise_beta_alpha: 1.5
+    noise_beta_beta: 1.0
+    noise_s: 0.999
+    num_timestep_buckets: 1000
+    num_inference_timesteps: 10
+    num_target_vision_tokens: 32
+    diffusion_model_cfg:
+      cross_attention_dim: 1536
+      dropout: 0.2
+      final_dropout: true
+      interleave_self_attention: true
+      norm_type: ada_norm
+      num_layers: 24
+      output_dim: 1536
+      positional_embeddings: null
+    mlp_head: 0
+  video_model:
+    model_name: /path/to/Wan2.1-Fun-V1.1-1.3B-InP/snapshots/xxx
+    config_path: starVLA/model/modules/video_model/config/wan2.1/wan_civitai.yaml
+    random_hw_adapt: true
+    training_with_video_token_length: true
+    enable_bucket: true
+    weighting_scheme: none
+    uniform_sampling: true
+    train_sampling_steps: 1000
+    train_mode: inpaint
+    trainable_modules: .
+    ckpt_1d: null
+  gs_model:
+    model_path: /path/to/gs_model.pth
+    model_type: STORM-B/8
+    input_size:
+    - 144
+    - 256
+    gs_dim: 3
+    decoder_type: dummy
+    disable_grad_checkpointing: 0
+    use_sky_token: 1
+    use_affine_token: 1
+    num_motion_tokens: 16
+    use_latest_gsplat: 0
+    enable_depth_loss: 1
+    enable_sky_opacity_loss: 1
+    sky_depth: 300.0
+    enable_sky_depth_loss: 0
+    sky_opacity_loss_coeff: 0.1
+    enable_flow_reg_loss: 1
+    flow_reg_coeff: 0.005
+    enable_perceptual_loss: 1
+    perceptual_loss_start_iter: 5000
+    perceptual_weight: 0.05
+  reward_model:
+    action_model_type: DiT-B
+    hidden_size: 1024
+    add_pos_embed: true
+    max_seq_len: 1024
+    action_dim: 1
+    state_dim: 1
+    future_action_window_size: 8
+    action_horizon: 8
+    past_action_window_size: 0
+    repeated_diffusion_steps: 1
+    noise_beta_alpha: 1.5
+    noise_beta_beta: 1.0
+    noise_s: 0.999
+    num_timestep_buckets: 1000
+    num_inference_timesteps: 10
+    num_target_vision_tokens: 32
+    diffusion_model_cfg:
+      cross_attention_dim: 1024
+      dropout: 0.2
+      final_dropout: true
+      interleave_self_attention: true
+      norm_type: ada_norm
+      num_layers: 16
+      output_dim: 1024
+      positional_embeddings: null
+    mlp_head: 0
+datasets:
+  vlm_data:
+    dataset_py: vlm_datasets
+    dataformat: llava_json
+    dataset_use: asv2_conversation_en,asv2_detailed_description_en,asv2_region_captioning_en,coco_internvl_longcap_en,coco_karpathy_train_567_en,coco_negative_gpt4o_en,coco_poetry_zh,coco_rem_en_zh,cocorem_exist_yorn_en,cocotextv2_en,cocotextv2_gpt4o_en,okvqa_en,refcoco_grounding_aug_en,refcoco_grounding_en,tallyqa_coco_en,toloka_grounding_aug_en,vqav2_en,vsr_en
+    eval_dataset: aokvqa_cauldron_llava_format
+    data_flatten: false
+    base_interval: 2
+    max_pixels: 50176
+    min_pixels: 784
+    model_max_length: 2048
+    model_type: qwen2.5vl
+    per_device_batch_size: 4
+  vla_data:
+    load_act_data: 1
+    dataset_py: navsim_dataset
+    datalist_path: /path/to/navtrain_meta.json
+    split: train
+    per_device_batch_size: 4
+    load_all_data_for_training: true
+    w_neg_traj: null
+    act_norm: 1
+  video_data:
+    load_2d_data: 1
+    rgb_meta_dir: /path/to/navsim_video
+    sample_n_frames: 9
+    sample_size:
+    - 144
+    - 256
+    text_input: 0
+  gs_data:
+    load_3d_data: 0
+    debug: 0
+    gs_meta_dir: /path/to/navsim_storm
+  reward_data:
+    load_reward_data: 0
+    reward_meta_dir: /path/to/navsim_reward
+trainer:
+  epochs: 100
+  max_train_steps: 200000
+  num_warmup_steps: 5000
+  save_interval: 10000
+  eval_interval: 200000
+  learning_rate:
+    base: 1.0e-05
+    qwen_vl_interface: 1.0e-05
+    rgb_model: 1.0e-05
+    action_model: 1.0e-05
+    gs_model: 1.0e-05
+    reward_model: 1.0e-05
+  lr_scheduler_type: cosine_with_min_lr
+  scheduler_specific_kwargs:
+    min_lr: 5.0e-07
+  freeze_modules: rgb_model.vae,rgb_model.clip_image_encoder,rgb_model.text_encoder,qwen_vl_interface.model.visual
+  loss_scale:
+    vla: 1.0
+    vlm: 0.1
+  repeated_diffusion_steps: 4
+  max_grad_norm: 1.0
+  warmup_ratio: 0.1
+  weight_decay: 0.0
+  logging_frequency: 100
+  gradient_clipping: 1.0
+  gradient_accumulation_steps: 1
+  optimizer:
+    name: AdamW
+    betas:
+    - 0.9
+    - 0.95
+    eps: 1.0e-08
+    weight_decay: 0.001
+  is_resume: false
+  resume_ckpt: none
+  resume_epoch: null
+  resume_step: null
+  enable_gradient_checkpointing: true
+  enable_mixed_precision_training: true
+output_dir: /path/to/navsim_exp/your-run-id