outsider86 commited on
Commit
6792a26
·
verified ·
1 Parent(s): 1f2293a

Upload fastumi_airhockey_bounce_qwenDiscreteDiffusion_0523/config.yaml with huggingface_hub

Browse files
fastumi_airhockey_bounce_qwenDiscreteDiffusion_0523/config.yaml ADDED
@@ -0,0 +1,79 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ datasets:
2
+ vla_data:
3
+ action_mode: abs
4
+ data_mix: 3airhockey_dynamic_bounce
5
+ data_root_dir: playground/Datasets/FastUMI
6
+ dataset_py: lerobot_datasets
7
+ image_size:
8
+ - 224
9
+ - 224
10
+ per_device_batch_size: 8
11
+ video_backend: torchvision_av
12
+ framework:
13
+ action_model:
14
+ action_dim: 2
15
+ action_high: 1.0
16
+ action_low: -1.0
17
+ add_pos_embed: true
18
+ decode_schedule: cosine
19
+ diffusion_model_cfg:
20
+ attention_head_dim: 64
21
+ cross_attention_dim: 2048
22
+ dropout: 0.2
23
+ final_dropout: true
24
+ input_embedding_dim: 2048
25
+ interleave_self_attention: true
26
+ norm_type: ada_norm
27
+ num_attention_heads: 32
28
+ num_layers: 36
29
+ output_dim: 256
30
+ positional_embeddings: sinusoidal
31
+ future_action_window_size: 31
32
+ l1_loss_weight: 0.1
33
+ max_seq_len: 1024
34
+ no_mask_token_prob: 0.0
35
+ num_bins: 256
36
+ num_inference_steps: 8
37
+ num_target_vision_tokens: 32
38
+ past_action_window_size: 0
39
+ representation: bin
40
+ state_dim: 2
41
+ train_mask_schedule: cosine
42
+ use_simple_max: false
43
+ name: QwenDiscreteDiffusion
44
+ qwenvl:
45
+ attn_implementation: flash_attention_2
46
+ base_vlm: playground/Pretrained_models/Qwen2.5-VL-3B-Instruct-Action
47
+ num_vl_layers: 36
48
+ vl_hidden_dim: 2048
49
+ output_dir: ./results/Checkpoints/fastumi_airhockey_bounce_qwenDiscreteDiffusion_0523
50
+ run_id: fastumi_airhockey_bounce_qwenDiscreteDiffusion_0523
51
+ run_root_dir: ./results/Checkpoints
52
+ seed: 42
53
+ trainer:
54
+ eval_interval: 100
55
+ freeze_modules: null
56
+ gradient_accumulation_steps: 1
57
+ gradient_clipping: 1.0
58
+ is_resume: false
59
+ learning_rate:
60
+ action_model: 0.0001
61
+ base: 1.0e-05
62
+ qwen_vl_interface: 1.0e-05
63
+ logging_frequency: 50
64
+ lr_scheduler_type: cosine_with_min_lr
65
+ max_train_steps: 20000
66
+ num_warmup_steps: 5000
67
+ optimizer:
68
+ betas:
69
+ - 0.9
70
+ - 0.95
71
+ eps: 1.0e-08
72
+ weight_decay: 1.0e-08
73
+ repeated_diffusion_steps: 4
74
+ save_format: pt
75
+ save_interval: 2000
76
+ scheduler_specific_kwargs:
77
+ min_lr: 5.0e-07
78
+ wandb_entity: 2200011093-peking-university
79
+ wandb_project: starVLA_FastUMI_airhockey_bounce_0523_Discrete