cheikh025 commited on
Commit
a5c303d
·
verified ·
1 Parent(s): 059197c

Upload promoted/0017_libero90_weak_task_oversampling/config.yaml with huggingface_hub

Browse files
promoted/0017_libero90_weak_task_oversampling/config.yaml ADDED
@@ -0,0 +1,204 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ output_dir: ./runs/reweighted_libero90_finetune/exp0017_weak_task_oversample
2
+ batch_size: 4
3
+ num_workers: 8
4
+ lr_scheduler_type: cosine
5
+ learning_rate: 3.0e-05
6
+ num_epochs: 1
7
+ max_steps: 8000
8
+ log_every: 10
9
+ save_every: 500
10
+ eval_every: 200
11
+ eval_num_inference_steps: 10
12
+ gradient_accumulation_steps: 4
13
+ mixed_precision: bf16
14
+ seed: 42
15
+ max_grad_norm: 1.0
16
+ weight_decay: 0.01
17
+ resume: ./runs/reweighted_libero90_finetune/exp0014_continued_dirresume/checkpoints/weights/step_014000.pt
18
+ wandb:
19
+ enabled: false
20
+ workspace: null
21
+ project: fast-wam
22
+ name: exp0017_libero90_weak_task_oversampling
23
+ group: null
24
+ mode: online
25
+ data:
26
+ train:
27
+ _target_: fastwam.datasets.lerobot.robot_video_dataset.RobotVideoDataset
28
+ dataset_dirs:
29
+ - ./data/libero_mujoco3.3.2/libero_spatial_no_noops_lerobot
30
+ - ./data/libero_mujoco3.3.2/libero_object_no_noops_lerobot
31
+ - ./data/libero_mujoco3.3.2/libero_goal_no_noops_lerobot
32
+ - ./data/libero_mujoco3.3.2/libero_goal_no_noops_lerobot
33
+ - ./data/libero_mujoco3.3.2/libero_goal_no_noops_lerobot
34
+ - ./data/libero_mujoco3.3.2/libero_goal_no_noops_lerobot
35
+ - ./data/libero_mujoco3.3.2/libero_goal_no_noops_lerobot
36
+ - ./data/libero_mujoco3.3.2/libero_10_no_noops_lerobot
37
+ - ./data/libero_mujoco3.3.2/libero_10_no_noops_lerobot
38
+ - ./data/libero_mujoco3.3.2/libero_10_no_noops_lerobot
39
+ - ./data/libero_mujoco3.3.2/libero_10_no_noops_lerobot
40
+ - ./data/libero_mujoco3.3.2/libero_10_no_noops_lerobot
41
+ - ./data/libero_mujoco3.3.2/libero_90_no_noops_lerobot
42
+ - ./data/libero_mujoco3.3.2/libero_90_weak_subset_lerobot
43
+ - ./data/libero_mujoco3.3.2/libero_90_weak_subset_lerobot
44
+ - ./data/libero_mujoco3.3.2/libero_90_weak_subset_lerobot
45
+ - ./data/libero_mujoco3.3.2/libero_90_weak_subset_lerobot
46
+ shape_meta:
47
+ images:
48
+ - key: image
49
+ raw_shape:
50
+ - 3
51
+ - 512
52
+ - 512
53
+ shape:
54
+ - 3
55
+ - 224
56
+ - 224
57
+ - key: wrist_image
58
+ raw_shape:
59
+ - 3
60
+ - 512
61
+ - 512
62
+ shape:
63
+ - 3
64
+ - 224
65
+ - 224
66
+ action:
67
+ - key: default
68
+ raw_shape: 7
69
+ shape: 7
70
+ state:
71
+ - key: default
72
+ raw_shape: 8
73
+ shape: 8
74
+ num_frames: 33
75
+ global_sample_stride: 1
76
+ action_video_freq_ratio: 4
77
+ video_size:
78
+ - 224
79
+ - 448
80
+ camera_key: null
81
+ val_set_proportion: 0.0
82
+ is_training_set: true
83
+ skip_padding_as_possible: false
84
+ concat_multi_camera: horizontal
85
+ processor:
86
+ _target_: fastwam.datasets.lerobot.processors.fastwam_processor.FastWAMProcessor
87
+ shape_meta:
88
+ images:
89
+ - key: image
90
+ raw_shape:
91
+ - 3
92
+ - 512
93
+ - 512
94
+ shape:
95
+ - 3
96
+ - 224
97
+ - 224
98
+ - key: wrist_image
99
+ raw_shape:
100
+ - 3
101
+ - 512
102
+ - 512
103
+ shape:
104
+ - 3
105
+ - 224
106
+ - 224
107
+ action:
108
+ - key: default
109
+ raw_shape: 7
110
+ shape: 7
111
+ state:
112
+ - key: default
113
+ raw_shape: 8
114
+ shape: 8
115
+ num_obs_steps: 33
116
+ num_output_cameras: 2
117
+ action_output_dim: 7
118
+ proprio_output_dim: 8
119
+ delta_action_dim_mask:
120
+ default:
121
+ - true
122
+ - true
123
+ - true
124
+ - true
125
+ - true
126
+ - true
127
+ - false
128
+ action_state_transforms: null
129
+ use_stepwise_action_norm: false
130
+ norm_default_mode: min/max
131
+ norm_exception_mode: null
132
+ action_state_merger:
133
+ _target_: fastwam.datasets.lerobot.transforms.action_state_merger.ConcatLeftAlign
134
+ train_transforms:
135
+ - _target_: fastwam.datasets.lerobot.transforms.image.ToTensor
136
+ - _target_: torchvision.transforms.Resize
137
+ size:
138
+ - 224
139
+ - 224
140
+ val_transforms:
141
+ - _target_: fastwam.datasets.lerobot.transforms.image.ToTensor
142
+ - _target_: torchvision.transforms.Resize
143
+ size:
144
+ - 224
145
+ - 224
146
+ text_embedding_cache_dir: ./data/text_embeds_cache/libero
147
+ context_len: 128
148
+ model:
149
+ _target_: fastwam.runtime.create_fastwam
150
+ model_id: Wan-AI/Wan2.2-TI2V-5B
151
+ tokenizer_model_id: Wan-AI/Wan2.1-T2V-1.3B
152
+ tokenizer_max_len: 128
153
+ load_text_encoder: false
154
+ proprio_dim: 8
155
+ redirect_common_files: true
156
+ mot_checkpoint_mixed_attn: false
157
+ action_dit_pretrained_path: checkpoints/ActionDiT_linear_interp_Wan22_alphascale_1024hdim.pt
158
+ skip_dit_load_from_pretrain: false
159
+ video_dit_config:
160
+ has_image_input: false
161
+ patch_size:
162
+ - 1
163
+ - 2
164
+ - 2
165
+ in_dim: 48
166
+ hidden_dim: 3072
167
+ ffn_dim: 14336
168
+ freq_dim: 256
169
+ text_dim: 4096
170
+ out_dim: 48
171
+ num_heads: 24
172
+ attn_head_dim: 128
173
+ num_layers: 30
174
+ eps: 1.0e-06
175
+ seperated_timestep: true
176
+ require_clip_embedding: false
177
+ require_vae_embedding: false
178
+ fuse_vae_embedding_in_latents: true
179
+ use_gradient_checkpointing: false
180
+ video_attention_mask_mode: first_frame_causal
181
+ action_conditioned: false
182
+ action_dim: 7
183
+ action_group_causal_mask_mode: group_diagonal
184
+ action_dit_config:
185
+ action_dim: 7
186
+ hidden_dim: 1024
187
+ ffn_dim: 4096
188
+ num_heads: 24
189
+ attn_head_dim: 128
190
+ num_layers: 30
191
+ text_dim: 4096
192
+ freq_dim: 256
193
+ eps: 1.0e-06
194
+ use_gradient_checkpointing: false
195
+ video_scheduler:
196
+ train_shift: 5.0
197
+ infer_shift: 5.0
198
+ num_train_timesteps: 1000
199
+ action_scheduler:
200
+ train_shift: 5.0
201
+ infer_shift: 5.0
202
+ num_train_timesteps: 1000
203
+ loss:
204
+ lambda_action: 1.0