Upload fastumi_pickandplace_qwenPI_0403_1_pick_from_moved/config.yaml with huggingface_hub
Browse files
fastumi_pickandplace_qwenPI_0403_1_pick_from_moved/config.yaml
ADDED
|
@@ -0,0 +1,70 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
datasets:
|
| 2 |
+
vla_data:
|
| 3 |
+
CoT_prompt: Your task is {instruction}. To identify the key objects for your task.
|
| 4 |
+
Locate their bounding boxes in [x1,y1,x2,y2] format.
|
| 5 |
+
data_mix: dynamic-0403-1-pick_from_moved
|
| 6 |
+
data_root_dir: playground/Datasets/FastUMI
|
| 7 |
+
dataset_py: lerobot_datasets
|
| 8 |
+
per_device_batch_size: 8
|
| 9 |
+
video_backend: torchvision_av
|
| 10 |
+
framework:
|
| 11 |
+
action_model:
|
| 12 |
+
action_dim: 10
|
| 13 |
+
add_pos_embed: true
|
| 14 |
+
diffusion_model_cfg:
|
| 15 |
+
attention_head_dim: 64
|
| 16 |
+
cross_attention_dim: 2048
|
| 17 |
+
dropout: 0.2
|
| 18 |
+
final_dropout: true
|
| 19 |
+
input_embedding_dim: 2048
|
| 20 |
+
interleave_self_attention: true
|
| 21 |
+
norm_type: ada_norm
|
| 22 |
+
num_attention_heads: 32
|
| 23 |
+
num_layers: 36
|
| 24 |
+
output_dim: 1024
|
| 25 |
+
positional_embeddings: null
|
| 26 |
+
future_action_window_size: 15
|
| 27 |
+
max_seq_len: 1024
|
| 28 |
+
noise_beta_alpha: 1.5
|
| 29 |
+
noise_beta_beta: 1.0
|
| 30 |
+
noise_s: 0.999
|
| 31 |
+
num_inference_timesteps: 4
|
| 32 |
+
num_target_vision_tokens: 32
|
| 33 |
+
num_timestep_buckets: 1000
|
| 34 |
+
past_action_window_size: 0
|
| 35 |
+
state_dim: 10
|
| 36 |
+
name: QwenPI
|
| 37 |
+
qwenvl:
|
| 38 |
+
attn_implementation: flash_attention_2
|
| 39 |
+
base_vlm: playground/Pretrained_models/Qwen2.5-VL-3B-Instruct-Action
|
| 40 |
+
num_vl_layers: 36
|
| 41 |
+
vl_hidden_dim: 2048
|
| 42 |
+
output_dir: ./results/Checkpoints/fastumi_pickandplace_qwenPI_0403_1_pick_from_moved
|
| 43 |
+
run_id: fastumi_pickandplace_qwenPI_0403_1_pick_from_moved
|
| 44 |
+
run_root_dir: ./results/Checkpoints
|
| 45 |
+
seed: 42
|
| 46 |
+
trainer:
|
| 47 |
+
eval_interval: 100
|
| 48 |
+
freeze_modules: null
|
| 49 |
+
gradient_accumulation_steps: 1
|
| 50 |
+
gradient_clipping: 1.0
|
| 51 |
+
is_resume: true
|
| 52 |
+
learning_rate:
|
| 53 |
+
action_model: 0.0001
|
| 54 |
+
base: 2.5e-05
|
| 55 |
+
qwen_vl_interface: 1.0e-05
|
| 56 |
+
logging_frequency: 50
|
| 57 |
+
lr_scheduler_type: cosine_with_min_lr
|
| 58 |
+
max_train_steps: 30000
|
| 59 |
+
num_warmup_steps: 5000
|
| 60 |
+
optimizer:
|
| 61 |
+
betas:
|
| 62 |
+
- 0.9
|
| 63 |
+
- 0.95
|
| 64 |
+
eps: 1.0e-08
|
| 65 |
+
weight_decay: 1.0e-08
|
| 66 |
+
save_interval: 10000
|
| 67 |
+
scheduler_specific_kwargs:
|
| 68 |
+
min_lr: 1.0e-06
|
| 69 |
+
wandb_entity: 2200011093-peking-university
|
| 70 |
+
wandb_project: starVLA_FastUMI_dynamic_0403_1_pick_from_moved
|