config.yaml

# Frozen VLA Training Configuration (Qwen3-VL Version)
# PAI Environment: PyTorch 2.6.0 + Python 3.10 + CUDA 12.6 + A10 24GB
#
# 注意：
#   - Qwen3-VL-2B-Instruct 已下载到 /mnt/workspace/Qwen3-VL-2B-Instruct/
#   - 使用本地路径加载，避免训练时联网下载
#   - BridgeData 在 /mnt/workspace/Dataset/stage_frozen01/full/

# Data paths
data_dir: /mnt/workspace/Dataset/stage_frozen01/full
output_dir: /mnt/workspace/checkpoints
log_dir: /mnt/workspace/logs

# Model config
model:
  llm: "/mnt/workspace/Qwen3-VL-2B-Instruct"              # PAI 本地已下载路径
  mlp_hidden_dim: 512
  mlp_depth: 2
  action_dim: 7                                         # EEF delta: [dx,dy,dz,ax,ay,az,gripper]
  use_processor: true                                  # 使用模型自带 AutoProcessor 处理图像

# Training config
training:
  batch_size: 32
  gradient_accumulation_steps: 2   # 等效 batch_size=64，省显存
  num_epochs: 5
  learning_rate: 1.0e-4
  weight_decay: 1.0e-2
  warmup_steps: 500
  max_grad_norm: 1.0
  save_every_n_epochs: 1
  save_every_n_steps: 500        # 每 500 步额外存一次（防崩，只保留最近3个）
  eval_every_n_epochs: 1
  
  # Optimizer
  optimizer: "adamw"
  scheduler: "cosine"
  
  # Memory optimization
  mixed_precision: "bf16"          # bf16 比 fp16 更稳定，A10 支持
  gradient_checkpointing: false   # 冻结模型不需要
  compile: false                  # PyTorch 2.x compile，可加速但首次编译慢

# Data config
data:
  image_size: 448                 # 当 use_processor=false 时生效（Qwen3-VL 默认 448）
  fps: 5                          # BridgeData 采样率
  num_workers: 1                  # DataLoader workers（VideoCache 每个 worker 独立，太多爆 RAM）
  pin_memory: true
  shuffle: true
  action_normalize: true          # 对 action 做均值方差标准化
  
  # 采样策略
  frame_sampling: "all"           # "all"=每帧都训练, "uniform"=均匀采样N帧
  max_frames_per_episode: null    # 仅 frame_sampling=uniform 时生效
  max_cache_episodes: 150         # VideoCache 上限（RAM 有限时降低此值）

# Logging
logging:
  wandb_project: "lingarm-vla"
  wandb_run_name: "qwen3vl-frozen-mlp-bridge2.6k"
  log_interval: 50               # 每50步打印一次日志
  
# Device
device: "cuda"
seed: 42