robopark commited on
Commit
2ea8102
·
verified ·
1 Parent(s): 3b48004

Add Click Mouse and Pinch Tongs epoch-2000 BC checkpoints

Browse files
pretrained_bc_policy_click_mouse_img/.hydra/config.yaml ADDED
@@ -0,0 +1,124 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ _target_: agent.pretrain.train_flow_matching_img_agent.TrainFlowMatchingImgAgent
2
+ name: ${env_name}_pre_flow_matching_unet_img_ta${horizon_steps}_td${flow_steps}
3
+ logdir: log_dir/dexjoco-pretrain/pretrained_bc_policy_click_mouse_img
4
+ train_dataset_path: data_dir/dexjoco/click_mouse-img/ph_pretrain/train.npz
5
+ normalization_path: log_dir/dexjoco-pretrain/pretrained_bc_policy_click_mouse_img/normalization.npz
6
+ seed: 42
7
+ device: cuda:0
8
+ env_name: click_mouse
9
+ obs_dim: 23
10
+ action_dim: 22
11
+ use_6d_rot: false
12
+ abs_action: false
13
+ flow_steps: 10
14
+ horizon_steps: 16
15
+ cond_steps: 1
16
+ img_cond_steps: 1
17
+ act_steps: 16
18
+ env:
19
+ n_envs: 10
20
+ name: ${env_name}
21
+ best_reward_threshold_for_success: 1
22
+ max_episode_steps: 500
23
+ save_video: false
24
+ use_image_obs: true
25
+ env_type: dexjoco
26
+ wrappers:
27
+ dexjoco_image:
28
+ normalization_path: ${normalization_path}
29
+ low_dim_keys:
30
+ - tcp_pose
31
+ - gripper_pose
32
+ image_keys:
33
+ - front
34
+ - wrist
35
+ shape_meta: ${shape_meta}
36
+ task_name: ${env_name}
37
+ policy_mode: true
38
+ randomize: false
39
+ render_mode: rgb_array
40
+ realtime: false
41
+ max_episode_steps: ${env.max_episode_steps}
42
+ success_steps_before_termination: 1
43
+ multi_step:
44
+ n_obs_steps: ${cond_steps}
45
+ n_action_steps: ${act_steps}
46
+ max_episode_steps: ${env.max_episode_steps}
47
+ reset_within_step: true
48
+ shape_meta:
49
+ obs:
50
+ rgb:
51
+ shape:
52
+ - 96
53
+ - 96
54
+ - 6
55
+ state:
56
+ shape:
57
+ - 23
58
+ action:
59
+ shape:
60
+ - 22
61
+ wandb: null
62
+ train:
63
+ n_epochs: 3000
64
+ batch_size: 256
65
+ learning_rate: 0.0001
66
+ weight_decay: 1.0e-06
67
+ lr_scheduler:
68
+ first_cycle_steps: ${train.n_epochs}
69
+ warmup_steps: 100
70
+ min_lr: 1.0e-05
71
+ epoch_start_ema: 20
72
+ update_ema_freq: 10
73
+ save_model_freq: 100
74
+ val_freq: 100
75
+ n_eval_episodes: 100
76
+ render:
77
+ freq: 2
78
+ num: 0
79
+ model:
80
+ _target_: model.flow_matching.flow_matching.FlowMatchingModel
81
+ network:
82
+ _target_: model.flow_matching.vision_resnet_mlp_flow_matching.VisionResNetFlowMatchingMLP
83
+ num_img: 2
84
+ spatial_emb: 64
85
+ time_dim: 64
86
+ dropout: 0.1
87
+ visual_feature_dim: 128
88
+ fused_feature_dim: 128
89
+ img_cond_steps: ${img_cond_steps}
90
+ cond_dim: ${obs_dim}
91
+ horizon_steps: ${horizon_steps}
92
+ action_dim: ${action_dim}
93
+ img_height: 96
94
+ img_width: 96
95
+ num_kp: 32
96
+ unet_base_channels: 128
97
+ unet_channel_mults:
98
+ - 1
99
+ - 2
100
+ - 2
101
+ use_augmentation: true
102
+ crop_height: 84
103
+ crop_width: 84
104
+ horizon_steps: ${horizon_steps}
105
+ obs_dim: ${obs_dim}
106
+ action_dim: ${action_dim}
107
+ flow_steps: ${flow_steps}
108
+ integration_method: euler
109
+ schedule: linear
110
+ t_sampling: uniform
111
+ weight_method: none
112
+ device: ${device}
113
+ ema:
114
+ decay: 0.995
115
+ train_dataset:
116
+ _target_: agent.dataset.sequence.StitchedSequenceDataset
117
+ use_img: true
118
+ dataset_path: ${train_dataset_path}
119
+ horizon_steps: ${horizon_steps}
120
+ max_n_episodes: 99
121
+ cond_steps: ${cond_steps}
122
+ img_cond_steps: ${img_cond_steps}
123
+ device: cpu
124
+ image_keys: ${env.wrappers.dexjoco_image.image_keys}
pretrained_bc_policy_click_mouse_img/checkpoint/state_2000.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:87c10099e1e1897c17b1387f34e96be3a6da362899c2c8321d6d4c7c6f2dbcb1
3
+ size 207915841
pretrained_bc_policy_click_mouse_img/normalization.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6d7a2aa1d1da83146df134b424b3af8419271198552b97169de2803d708e35c1
3
+ size 1317
pretrained_bc_policy_pinch_tongs_img/.hydra/config.yaml ADDED
@@ -0,0 +1,124 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ _target_: agent.pretrain.train_flow_matching_img_agent.TrainFlowMatchingImgAgent
2
+ name: ${env_name}_pre_flow_matching_unet_img_ta${horizon_steps}_td${flow_steps}
3
+ logdir: log_dir/dexjoco-pretrain/pretrained_bc_policy_pinch_tongs_img
4
+ train_dataset_path: data_dir/dexjoco/pinch_tongs-img/ph_pretrain/train.npz
5
+ normalization_path: log_dir/dexjoco-pretrain/pretrained_bc_policy_pinch_tongs_img/normalization.npz
6
+ seed: 42
7
+ device: cuda:0
8
+ env_name: pinch_tongs
9
+ obs_dim: 23
10
+ action_dim: 22
11
+ use_6d_rot: false
12
+ abs_action: false
13
+ flow_steps: 10
14
+ horizon_steps: 16
15
+ cond_steps: 1
16
+ img_cond_steps: 1
17
+ act_steps: 16
18
+ env:
19
+ n_envs: 10
20
+ name: ${env_name}
21
+ best_reward_threshold_for_success: 1
22
+ max_episode_steps: 700
23
+ save_video: false
24
+ use_image_obs: true
25
+ env_type: dexjoco
26
+ wrappers:
27
+ dexjoco_image:
28
+ normalization_path: ${normalization_path}
29
+ low_dim_keys:
30
+ - tcp_pose
31
+ - gripper_pose
32
+ image_keys:
33
+ - front
34
+ - wrist
35
+ shape_meta: ${shape_meta}
36
+ task_name: ${env_name}
37
+ policy_mode: true
38
+ randomize: false
39
+ render_mode: rgb_array
40
+ realtime: false
41
+ max_episode_steps: ${env.max_episode_steps}
42
+ success_steps_before_termination: 1
43
+ multi_step:
44
+ n_obs_steps: ${cond_steps}
45
+ n_action_steps: ${act_steps}
46
+ max_episode_steps: ${env.max_episode_steps}
47
+ reset_within_step: true
48
+ shape_meta:
49
+ obs:
50
+ rgb:
51
+ shape:
52
+ - 96
53
+ - 96
54
+ - 6
55
+ state:
56
+ shape:
57
+ - 23
58
+ action:
59
+ shape:
60
+ - 22
61
+ wandb: null
62
+ train:
63
+ n_epochs: 3000
64
+ batch_size: 256
65
+ learning_rate: 0.0001
66
+ weight_decay: 1.0e-06
67
+ lr_scheduler:
68
+ first_cycle_steps: ${train.n_epochs}
69
+ warmup_steps: 100
70
+ min_lr: 1.0e-05
71
+ epoch_start_ema: 20
72
+ update_ema_freq: 10
73
+ save_model_freq: 100
74
+ val_freq: 100
75
+ n_eval_episodes: 100
76
+ render:
77
+ freq: 2
78
+ num: 0
79
+ model:
80
+ _target_: model.flow_matching.flow_matching.FlowMatchingModel
81
+ network:
82
+ _target_: model.flow_matching.vision_resnet_mlp_flow_matching.VisionResNetFlowMatchingMLP
83
+ num_img: 2
84
+ spatial_emb: 64
85
+ time_dim: 64
86
+ dropout: 0.1
87
+ visual_feature_dim: 128
88
+ fused_feature_dim: 128
89
+ img_cond_steps: ${img_cond_steps}
90
+ cond_dim: ${obs_dim}
91
+ horizon_steps: ${horizon_steps}
92
+ action_dim: ${action_dim}
93
+ img_height: 96
94
+ img_width: 96
95
+ num_kp: 32
96
+ unet_base_channels: 128
97
+ unet_channel_mults:
98
+ - 1
99
+ - 2
100
+ - 2
101
+ use_augmentation: true
102
+ crop_height: 84
103
+ crop_width: 84
104
+ horizon_steps: ${horizon_steps}
105
+ obs_dim: ${obs_dim}
106
+ action_dim: ${action_dim}
107
+ flow_steps: ${flow_steps}
108
+ integration_method: euler
109
+ schedule: linear
110
+ t_sampling: uniform
111
+ weight_method: none
112
+ device: ${device}
113
+ ema:
114
+ decay: 0.995
115
+ train_dataset:
116
+ _target_: agent.dataset.sequence.StitchedSequenceDataset
117
+ use_img: true
118
+ dataset_path: ${train_dataset_path}
119
+ horizon_steps: ${horizon_steps}
120
+ max_n_episodes: 100
121
+ cond_steps: ${cond_steps}
122
+ img_cond_steps: ${img_cond_steps}
123
+ device: cpu
124
+ image_keys: ${env.wrappers.dexjoco_image.image_keys}
pretrained_bc_policy_pinch_tongs_img/checkpoint/state_2000.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c1824c8a5c58f22d783ecd2f8b306b400754d78391a1b2bf60028bf1b7b16544
3
+ size 207915841
pretrained_bc_policy_pinch_tongs_img/normalization.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f5e4ac0d258dab14afdfe272c63e647fd751d0636ab6b89b8a79ec15df94b2f2
3
+ size 1303