solarian-diffusion 0.1.0__py3-none-any.whl

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (159) hide show
  1. diffusion_policy/__init__.py +0 -0
  2. diffusion_policy/codecs/__init__.py +0 -0
  3. diffusion_policy/codecs/imagecodecs_numcodecs.py +1408 -0
  4. diffusion_policy/common/__init__.py +0 -0
  5. diffusion_policy/common/checkpoint_util.py +64 -0
  6. diffusion_policy/common/json_logger.py +117 -0
  7. diffusion_policy/common/normalize_util.py +251 -0
  8. diffusion_policy/common/pose_repr_util.py +112 -0
  9. diffusion_policy/common/pytorch_util.py +82 -0
  10. diffusion_policy/common/q_critic_util.py +135 -0
  11. diffusion_policy/common/replay_buffer.py +672 -0
  12. diffusion_policy/common/sampler.py +204 -0
  13. diffusion_policy/config/legacy/task/blockpush_lowdim_seed.yaml +34 -0
  14. diffusion_policy/config/legacy/task/blockpush_lowdim_seed_abs.yaml +34 -0
  15. diffusion_policy/config/legacy/task/can_image.yaml +64 -0
  16. diffusion_policy/config/legacy/task/can_image_abs.yaml +64 -0
  17. diffusion_policy/config/legacy/task/can_lowdim.yaml +45 -0
  18. diffusion_policy/config/legacy/task/can_lowdim_abs.yaml +46 -0
  19. diffusion_policy/config/legacy/task/kitchen_lowdim.yaml +33 -0
  20. diffusion_policy/config/legacy/task/kitchen_lowdim_abs.yaml +38 -0
  21. diffusion_policy/config/legacy/task/lift_abs.yaml +72 -0
  22. diffusion_policy/config/legacy/task/lift_image.yaml +64 -0
  23. diffusion_policy/config/legacy/task/lift_image_abs.yaml +63 -0
  24. diffusion_policy/config/legacy/task/lift_lowdim.yaml +46 -0
  25. diffusion_policy/config/legacy/task/lift_lowdim_abs.yaml +47 -0
  26. diffusion_policy/config/legacy/task/pusht_image.yaml +40 -0
  27. diffusion_policy/config/legacy/task/pusht_lowdim.yaml +34 -0
  28. diffusion_policy/config/legacy/task/real_pusht_image.yaml +47 -0
  29. diffusion_policy/config/legacy/task/square_abs.yaml +72 -0
  30. diffusion_policy/config/legacy/task/square_image.yaml +64 -0
  31. diffusion_policy/config/legacy/task/square_image_abs.yaml +64 -0
  32. diffusion_policy/config/legacy/task/square_lowdim.yaml +46 -0
  33. diffusion_policy/config/legacy/task/square_lowdim_abs.yaml +47 -0
  34. diffusion_policy/config/legacy/task/tool_hang_abs.yaml +71 -0
  35. diffusion_policy/config/legacy/task/tool_hang_image.yaml +63 -0
  36. diffusion_policy/config/legacy/task/tool_hang_image_abs.yaml +63 -0
  37. diffusion_policy/config/legacy/task/tool_hang_lowdim.yaml +45 -0
  38. diffusion_policy/config/legacy/task/tool_hang_lowdim_abs.yaml +46 -0
  39. diffusion_policy/config/legacy/task/transport_image.yaml +75 -0
  40. diffusion_policy/config/legacy/task/transport_image_abs.yaml +75 -0
  41. diffusion_policy/config/legacy/task/transport_lowdim.yaml +49 -0
  42. diffusion_policy/config/legacy/task/transport_lowdim_abs.yaml +49 -0
  43. diffusion_policy/config/legacy/task/umi.yaml +45 -0
  44. diffusion_policy/config/legacy/task/umi_image.yaml +36 -0
  45. diffusion_policy/config/legacy/task/umi_teleop_image.yaml +41 -0
  46. diffusion_policy/config/legacy/train_bet_lowdim_workspace.yaml +132 -0
  47. diffusion_policy/config/legacy/train_diffusion_transformer_hybrid_workspace.yaml +141 -0
  48. diffusion_policy/config/legacy/train_diffusion_transformer_lowdim_kitchen_workspace.yaml +146 -0
  49. diffusion_policy/config/legacy/train_diffusion_transformer_lowdim_pusht_workspace.yaml +146 -0
  50. diffusion_policy/config/legacy/train_diffusion_transformer_lowdim_workspace.yaml +146 -0
  51. diffusion_policy/config/legacy/train_diffusion_transformer_real_hybrid_workspace.yaml +144 -0
  52. diffusion_policy/config/legacy/train_diffusion_transformer_umi_teleop_hybrid_workspace.yaml +156 -0
  53. diffusion_policy/config/legacy/train_diffusion_unet_ddim_hybrid_workspace.yaml +140 -0
  54. diffusion_policy/config/legacy/train_diffusion_unet_ddim_lowdim_workspace.yaml +146 -0
  55. diffusion_policy/config/legacy/train_diffusion_unet_hybrid_workspace.yaml +137 -0
  56. diffusion_policy/config/legacy/train_diffusion_unet_image_pretrained_workspace.yaml +149 -0
  57. diffusion_policy/config/legacy/train_diffusion_unet_image_workspace.yaml +150 -0
  58. diffusion_policy/config/legacy/train_diffusion_unet_lowdim_workspace.yaml +143 -0
  59. diffusion_policy/config/legacy/train_diffusion_unet_real_hybrid_workspace.yaml +141 -0
  60. diffusion_policy/config/legacy/train_diffusion_unet_real_image_workspace.yaml +152 -0
  61. diffusion_policy/config/legacy/train_diffusion_unet_real_pretrained_workspace.yaml +152 -0
  62. diffusion_policy/config/legacy/train_diffusion_unet_timm_workspace.yaml +145 -0
  63. diffusion_policy/config/legacy/train_diffusion_unet_umi_hybrid_workspace.yaml +141 -0
  64. diffusion_policy/config/legacy/train_diffusion_unet_umi_teleop_hybrid_workspace.yaml +153 -0
  65. diffusion_policy/config/legacy/train_diffusion_unet_umi_timm_workspace.yaml +145 -0
  66. diffusion_policy/config/legacy/train_diffusion_unet_umi_workspace.yaml +150 -0
  67. diffusion_policy/config/legacy/train_diffusion_unet_video_workspace.yaml +128 -0
  68. diffusion_policy/config/legacy/train_ibc_dfo_hybrid_workspace.yaml +111 -0
  69. diffusion_policy/config/legacy/train_ibc_dfo_lowdim_workspace.yaml +109 -0
  70. diffusion_policy/config/legacy/train_ibc_dfo_real_hybrid_workspace.yaml +111 -0
  71. diffusion_policy/config/legacy/train_robomimic_image_workspace.yaml +92 -0
  72. diffusion_policy/config/legacy/train_robomimic_lowdim_workspace.yaml +92 -0
  73. diffusion_policy/config/legacy/train_robomimic_real_image_workspace.yaml +92 -0
  74. diffusion_policy/config/rl_annotate.yaml +265 -0
  75. diffusion_policy/config/rl_annotate_per_frame.yaml +162 -0
  76. diffusion_policy/config/rl_annotate_per_frame_v4.yaml +162 -0
  77. diffusion_policy/config/rl_annotate_v2.yaml +141 -0
  78. diffusion_policy/config/rl_annotate_v3.yaml +165 -0
  79. diffusion_policy/config/task/lift.yaml +89 -0
  80. diffusion_policy/config/task/square.yaml +92 -0
  81. diffusion_policy/config/task/tool_hang.yaml +93 -0
  82. diffusion_policy/config/task/tool_hang_abs.yaml +71 -0
  83. diffusion_policy/config/task/umi.yaml +105 -0
  84. diffusion_policy/config/task/umi_bimanual.yaml +135 -0
  85. diffusion_policy/config/task/umi_cup_wild.yaml +106 -0
  86. diffusion_policy/config/task/umi_image.yaml +36 -0
  87. diffusion_policy/config/task/umi_pick_lego.yaml +106 -0
  88. diffusion_policy/config/task/umi_pick_single_lego_175.yaml +106 -0
  89. diffusion_policy/config/task/umi_pick_single_lego_bbox.yaml +135 -0
  90. diffusion_policy/config/task/umi_teleop_image.yaml +41 -0
  91. diffusion_policy/config/train_diffusion_transformer_umi_bimanual_workspace.yaml +185 -0
  92. diffusion_policy/config/train_diffusion_transformer_umi_workspace.yaml +185 -0
  93. diffusion_policy/config/train_diffusion_unet_image_workspace.yaml +152 -0
  94. diffusion_policy/config/train_diffusion_unet_timm_robomimic_workspace.yaml +152 -0
  95. diffusion_policy/config/train_diffusion_unet_timm_umi_cup_wild.yaml +260 -0
  96. diffusion_policy/config/train_diffusion_unet_timm_umi_pick_lego_pile_bbox.yaml +265 -0
  97. diffusion_policy/config/train_diffusion_unet_timm_umi_pick_lego_pile_bbox_awr.yaml +324 -0
  98. diffusion_policy/config/train_diffusion_unet_timm_umi_pick_lego_pile_bbox_parl.yaml +399 -0
  99. diffusion_policy/config/train_diffusion_unet_timm_umi_pick_single_lego.yaml +258 -0
  100. diffusion_policy/config/train_diffusion_unet_timm_umi_pick_single_lego_125.yaml +260 -0
  101. diffusion_policy/config/train_diffusion_unet_timm_umi_pick_single_lego_175.yaml +260 -0
  102. diffusion_policy/config/train_diffusion_unet_timm_umi_pick_single_lego_bbox.yaml +265 -0
  103. diffusion_policy/config/train_diffusion_unet_timm_umi_pick_two_legos_bbox.yaml +265 -0
  104. diffusion_policy/config/train_diffusion_unet_timm_umi_workspace.yaml +224 -0
  105. diffusion_policy/config/train_diffusion_unet_timm_umi_workspace_lego.yaml +213 -0
  106. diffusion_policy/config/train_diffusion_unet_timm_umi_workspace_lego_resume.yaml +213 -0
  107. diffusion_policy/config/train_diffusion_unet_timm_umi_workspace_pick_lego.yaml +254 -0
  108. diffusion_policy/config/train_diffusion_unet_timm_umi_workspace_red_cup.yaml +219 -0
  109. diffusion_policy/config/train_diffusion_unet_umi_bimanual_workspace.yaml +203 -0
  110. diffusion_policy/config/train_iql_umi.yaml +186 -0
  111. diffusion_policy/config/train_iql_umi_v2.yaml +212 -0
  112. diffusion_policy/config/train_iql_umi_v3.yaml +212 -0
  113. diffusion_policy/dataset/__init__.py +0 -0
  114. diffusion_policy/dataset/base_dataset.py +76 -0
  115. diffusion_policy/dataset/umi_dataset.py +652 -0
  116. diffusion_policy/env_runner/__init__.py +0 -0
  117. diffusion_policy/env_runner/base_image_runner.py +9 -0
  118. diffusion_policy/env_runner/real_pusht_image_runner.py +10 -0
  119. diffusion_policy/model/__init__.py +0 -0
  120. diffusion_policy/model/common/__init__.py +0 -0
  121. diffusion_policy/model/common/dict_of_tensor_mixin.py +38 -0
  122. diffusion_policy/model/common/lr_scheduler.py +46 -0
  123. diffusion_policy/model/common/module_attr_mixin.py +14 -0
  124. diffusion_policy/model/common/normalizer.py +353 -0
  125. diffusion_policy/model/common/tensor_util.py +960 -0
  126. diffusion_policy/model/diffusion/__init__.py +0 -0
  127. diffusion_policy/model/diffusion/conditional_unet1d.py +238 -0
  128. diffusion_policy/model/diffusion/conv1d_components.py +46 -0
  129. diffusion_policy/model/diffusion/ema_model.py +88 -0
  130. diffusion_policy/model/diffusion/mask_generator.py +221 -0
  131. diffusion_policy/model/diffusion/positional_embedding.py +17 -0
  132. diffusion_policy/model/vision/__init__.py +0 -0
  133. diffusion_policy/model/vision/crop_randomizer.py +288 -0
  134. diffusion_policy/model/vision/multi_image_obs_encoder.py +195 -0
  135. diffusion_policy/model/vision/timm_obs_encoder.py +525 -0
  136. diffusion_policy/policy/__init__.py +0 -0
  137. diffusion_policy/policy/base_image_policy.py +27 -0
  138. diffusion_policy/policy/diffusion_unet_image_policy.py +264 -0
  139. diffusion_policy/policy/diffusion_unet_timm_policy.py +607 -0
  140. diffusion_policy/policy/parl_action_optimizer.py +675 -0
  141. diffusion_policy/workspace/__init__.py +0 -0
  142. diffusion_policy/workspace/base_workspace.py +146 -0
  143. diffusion_policy/workspace/train_diffusion_unet_image_workspace.py +1585 -0
  144. solarian_diffusion/__init__.py +32 -0
  145. solarian_diffusion/config.py +31 -0
  146. solarian_diffusion/policy_server.py +316 -0
  147. solarian_diffusion/rtc.py +272 -0
  148. solarian_diffusion/service.py +16 -0
  149. solarian_diffusion/trainer.py +213 -0
  150. solarian_diffusion-0.1.0.dist-info/METADATA +126 -0
  151. solarian_diffusion-0.1.0.dist-info/RECORD +159 -0
  152. solarian_diffusion-0.1.0.dist-info/WHEEL +5 -0
  153. solarian_diffusion-0.1.0.dist-info/entry_points.txt +8 -0
  154. solarian_diffusion-0.1.0.dist-info/licenses/LICENSE +201 -0
  155. solarian_diffusion-0.1.0.dist-info/licenses/NOTICE +45 -0
  156. solarian_diffusion-0.1.0.dist-info/top_level.txt +3 -0
  157. umi/__init__.py +0 -0
  158. umi/common/__init__.py +0 -0
  159. umi/common/pose_util.py +115 -0
@@ -0,0 +1,132 @@
1
+ defaults:
2
+ - _self_
3
+ - task: blockpush_lowdim_seed
4
+
5
+ name: train_bet_lowdim
6
+ _target_: diffusion_policy.workspace.train_bet_lowdim_workspace.TrainBETLowdimWorkspace
7
+
8
+ obs_dim: ${task.obs_dim}
9
+ action_dim: ${task.action_dim}
10
+ keypoint_dim: ${task.keypoint_dim}
11
+ task_name: ${task.name}
12
+ exp_name: "default"
13
+
14
+ horizon: 3
15
+ n_obs_steps: 3
16
+ n_action_steps: 1
17
+ n_latency_steps: 0
18
+ past_action_visible: False
19
+ keypoint_visible_rate: 1.0
20
+ obs_as_local_cond: False
21
+ obs_as_global_cond: False
22
+ pred_action_steps_only: False
23
+
24
+ policy:
25
+ _target_: diffusion_policy.policy.bet_lowdim_policy.BETLowdimPolicy
26
+
27
+ action_ae:
28
+ _target_: diffusion_policy.model.bet.action_ae.discretizers.k_means.KMeansDiscretizer
29
+ num_bins: 24
30
+ action_dim: ${action_dim}
31
+ predict_offsets: True
32
+
33
+ obs_encoding_net:
34
+ _target_: torch.nn.Identity
35
+ output_dim: ${obs_dim}
36
+
37
+ state_prior:
38
+ _target_: diffusion_policy.model.bet.latent_generators.mingpt.MinGPT
39
+
40
+ discrete_input: false
41
+ input_dim: ${obs_dim}
42
+
43
+ vocab_size: ${policy.action_ae.num_bins}
44
+
45
+ # Architecture details
46
+ n_layer: 4
47
+ n_head: 4
48
+ n_embd: 72
49
+
50
+ block_size: ${horizon} # Length of history/context
51
+ predict_offsets: True
52
+ offset_loss_scale: 1000.0 # actions are very small
53
+ focal_loss_gamma: 2.0
54
+ action_dim: ${action_dim}
55
+
56
+ horizon: ${horizon}
57
+ n_obs_steps: ${n_obs_steps}
58
+ n_action_steps: ${n_action_steps}
59
+
60
+ dataloader:
61
+ batch_size: 256
62
+ num_workers: 1
63
+ shuffle: True
64
+ pin_memory: True
65
+ persistent_workers: False
66
+
67
+ val_dataloader:
68
+ batch_size: 256
69
+ num_workers: 1
70
+ shuffle: False
71
+ pin_memory: True
72
+ persistent_workers: False
73
+
74
+ optimizer:
75
+ learning_rate: 0.0001 # 1e-4
76
+ weight_decay: 0.1
77
+ betas: [0.9, 0.95]
78
+
79
+ training:
80
+ device: "cuda:0"
81
+ seed: 42
82
+ debug: False
83
+ resume: True
84
+ # optimization
85
+ lr_scheduler: cosine
86
+ lr_warmup_steps: 500
87
+ num_epochs: 5000
88
+ gradient_accumulate_every: 1
89
+ grad_norm_clip: 1.0
90
+ enable_normalizer: True
91
+ # training loop control
92
+ # in epochs
93
+ rollout_every: 50
94
+ checkpoint_every: 50
95
+ val_every: 1
96
+ sample_every: 5
97
+ # steps per epoch
98
+ max_train_steps: null
99
+ max_val_steps: null
100
+ # misc
101
+ tqdm_interval_sec: 1.0
102
+
103
+ logging:
104
+ project: diffusion_policy_debug
105
+ resume: True
106
+ mode: online
107
+ name: ${now:%Y.%m.%d-%H.%M.%S}_${name}_${task_name}
108
+ tags: ["${name}", "${task_name}", "${exp_name}"]
109
+ id: null
110
+ group: null
111
+
112
+ checkpoint:
113
+ topk:
114
+ monitor_key: test_mean_score
115
+ mode: max
116
+ k: 5
117
+ format_str: 'epoch={epoch:04d}-test_mean_score={test_mean_score:.3f}.ckpt'
118
+ save_last_ckpt: True
119
+ save_last_snapshot: False
120
+
121
+ multi_run:
122
+ run_dir: data/outputs/${now:%Y.%m.%d}/${now:%H.%M.%S}_${name}_${task_name}
123
+ wandb_name_base: ${now:%Y.%m.%d-%H.%M.%S}_${name}_${task_name}
124
+
125
+ hydra:
126
+ job:
127
+ override_dirname: ${name}
128
+ run:
129
+ dir: data/outputs/${now:%Y.%m.%d}/${now:%H.%M.%S}_${name}_${task_name}
130
+ sweep:
131
+ dir: data/outputs/${now:%Y.%m.%d}/${now:%H.%M.%S}_${name}_${task_name}
132
+ subdir: ${hydra.job.num}
@@ -0,0 +1,141 @@
1
+ defaults:
2
+ - _self_
3
+ - task: lift_image_abs
4
+
5
+ name: train_diffusion_transformer_hybrid
6
+ _target_: diffusion_policy.workspace.train_diffusion_transformer_hybrid_workspace.TrainDiffusionTransformerHybridWorkspace
7
+
8
+ task_name: ${task.name}
9
+ shape_meta: ${task.shape_meta}
10
+ exp_name: "default"
11
+
12
+ horizon: 10
13
+ n_obs_steps: 2
14
+ n_action_steps: 8
15
+ n_latency_steps: 0
16
+ dataset_obs_steps: ${n_obs_steps}
17
+ past_action_visible: False
18
+ keypoint_visible_rate: 1.0
19
+ obs_as_cond: True
20
+
21
+ policy:
22
+ _target_: diffusion_policy.policy.diffusion_transformer_hybrid_image_policy.DiffusionTransformerHybridImagePolicy
23
+
24
+ shape_meta: ${shape_meta}
25
+
26
+ noise_scheduler:
27
+ _target_: diffusers.schedulers.scheduling_ddpm.DDPMScheduler
28
+ num_train_timesteps: 100
29
+ beta_start: 0.0001
30
+ beta_end: 0.02
31
+ beta_schedule: squaredcos_cap_v2
32
+ variance_type: fixed_small # Yilun's paper uses fixed_small_log instead, but easy to cause Nan
33
+ clip_sample: True # required when predict_epsilon=False
34
+ prediction_type: epsilon # or sample
35
+
36
+ horizon: ${horizon}
37
+ n_action_steps: ${eval:'${n_action_steps}+${n_latency_steps}'}
38
+ n_obs_steps: ${n_obs_steps}
39
+ num_inference_steps: 100
40
+
41
+ crop_shape: [76, 76]
42
+ obs_encoder_group_norm: True
43
+ eval_fixed_crop: True
44
+
45
+ n_layer: 8
46
+ n_cond_layers: 0 # >0: use transformer encoder for cond, otherwise use MLP
47
+ n_head: 4
48
+ n_emb: 256
49
+ p_drop_emb: 0.0
50
+ p_drop_attn: 0.3
51
+ causal_attn: True
52
+ time_as_cond: True # if false, use BERT like encoder only arch, time as input
53
+ obs_as_cond: ${obs_as_cond}
54
+
55
+ # scheduler.step params
56
+ # predict_epsilon: True
57
+
58
+ ema:
59
+ _target_: diffusion_policy.model.diffusion.ema_model.EMAModel
60
+ update_after_step: 0
61
+ inv_gamma: 1.0
62
+ power: 0.75
63
+ min_value: 0.0
64
+ max_value: 0.9999
65
+
66
+ dataloader:
67
+ batch_size: 64
68
+ num_workers: 8
69
+ shuffle: True
70
+ pin_memory: True
71
+ persistent_workers: False
72
+
73
+ val_dataloader:
74
+ batch_size: 64
75
+ num_workers: 8
76
+ shuffle: False
77
+ pin_memory: True
78
+ persistent_workers: False
79
+
80
+ optimizer:
81
+ transformer_weight_decay: 1.0e-3
82
+ obs_encoder_weight_decay: 1.0e-6
83
+ learning_rate: 1.0e-4
84
+ betas: [0.9, 0.95]
85
+
86
+ training:
87
+ device: "cuda:0"
88
+ seed: 42
89
+ debug: False
90
+ resume: True
91
+ # optimization
92
+ lr_scheduler: cosine
93
+ # Transformer needs LR warmup
94
+ lr_warmup_steps: 1000
95
+ num_epochs: 3050
96
+ gradient_accumulate_every: 1
97
+ # EMA destroys performance when used with BatchNorm
98
+ # replace BatchNorm with GroupNorm.
99
+ use_ema: True
100
+ # training loop control
101
+ # in epochs
102
+ rollout_every: 50
103
+ checkpoint_every: 50
104
+ val_every: 1
105
+ sample_every: 5
106
+ # steps per epoch
107
+ max_train_steps: null
108
+ max_val_steps: null
109
+ # misc
110
+ tqdm_interval_sec: 1.0
111
+
112
+ logging:
113
+ project: diffusion_policy_debug
114
+ resume: True
115
+ mode: online
116
+ name: ${now:%Y.%m.%d-%H.%M.%S}_${name}_${task_name}
117
+ tags: ["${name}", "${task_name}", "${exp_name}"]
118
+ id: null
119
+ group: null
120
+
121
+ checkpoint:
122
+ topk:
123
+ monitor_key: test_mean_score
124
+ mode: max
125
+ k: 5
126
+ format_str: 'epoch={epoch:04d}-test_mean_score={test_mean_score:.3f}.ckpt'
127
+ save_last_ckpt: True
128
+ save_last_snapshot: False
129
+
130
+ multi_run:
131
+ run_dir: data/outputs/${now:%Y.%m.%d}/${now:%H.%M.%S}_${name}_${task_name}
132
+ wandb_name_base: ${now:%Y.%m.%d-%H.%M.%S}_${name}_${task_name}
133
+
134
+ hydra:
135
+ job:
136
+ override_dirname: ${name}
137
+ run:
138
+ dir: data/outputs/${now:%Y.%m.%d}/${now:%H.%M.%S}_${name}_${task_name}
139
+ sweep:
140
+ dir: data/outputs/${now:%Y.%m.%d}/${now:%H.%M.%S}_${name}_${task_name}
141
+ subdir: ${hydra.job.num}
@@ -0,0 +1,146 @@
1
+ defaults:
2
+ - _self_
3
+ - task: kitchen_lowdim_abs
4
+
5
+ name: train_diffusion_transformer_lowdim
6
+ _target_: diffusion_policy.workspace.train_diffusion_transformer_lowdim_workspace.TrainDiffusionTransformerLowdimWorkspace
7
+
8
+ obs_dim: ${task.obs_dim}
9
+ action_dim: ${task.action_dim}
10
+ task_name: ${task.name}
11
+ exp_name: "default"
12
+
13
+ horizon: 16
14
+ n_obs_steps: 4
15
+ n_action_steps: 8
16
+ n_latency_steps: 0
17
+ past_action_visible: False
18
+ keypoint_visible_rate: 1.0
19
+ obs_as_cond: True
20
+ pred_action_steps_only: False
21
+
22
+ policy:
23
+ _target_: diffusion_policy.policy.diffusion_transformer_lowdim_policy.DiffusionTransformerLowdimPolicy
24
+
25
+ model:
26
+ _target_: diffusion_policy.model.diffusion.transformer_for_diffusion.TransformerForDiffusion
27
+ input_dim: ${eval:'${action_dim} if ${obs_as_cond} else ${obs_dim} + ${action_dim}'}
28
+ output_dim: ${policy.model.input_dim}
29
+ horizon: ${horizon}
30
+ n_obs_steps: ${n_obs_steps}
31
+ cond_dim: ${eval:'${obs_dim} if ${obs_as_cond} else 0'}
32
+
33
+ n_layer: 8
34
+ n_head: 4
35
+ n_emb: 768
36
+ p_drop_emb: 0.0
37
+ p_drop_attn: 0.1
38
+
39
+ causal_attn: True
40
+ time_as_cond: True # if false, use BERT like encoder only arch, time as input
41
+ obs_as_cond: ${obs_as_cond}
42
+ n_cond_layers: 0 # >0: use transformer encoder for cond, otherwise use MLP
43
+
44
+ noise_scheduler:
45
+ _target_: diffusers.schedulers.scheduling_ddpm.DDPMScheduler
46
+ num_train_timesteps: 100
47
+ beta_start: 0.0001
48
+ beta_end: 0.02
49
+ beta_schedule: squaredcos_cap_v2
50
+ variance_type: fixed_small # Yilun's paper uses fixed_small_log instead, but easy to cause Nan
51
+ clip_sample: True # required when predict_epsilon=False
52
+ prediction_type: epsilon # or sample
53
+
54
+ horizon: ${horizon}
55
+ obs_dim: ${obs_dim}
56
+ action_dim: ${action_dim}
57
+ n_action_steps: ${n_action_steps}
58
+ n_obs_steps: ${n_obs_steps}
59
+ num_inference_steps: 100
60
+ obs_as_cond: ${obs_as_cond}
61
+ pred_action_steps_only: ${pred_action_steps_only}
62
+
63
+ # scheduler.step params
64
+ # predict_epsilon: True
65
+
66
+ ema:
67
+ _target_: diffusion_policy.model.diffusion.ema_model.EMAModel
68
+ update_after_step: 0
69
+ inv_gamma: 1.0
70
+ power: 0.75
71
+ min_value: 0.0
72
+ max_value: 0.9999
73
+
74
+ dataloader:
75
+ batch_size: 256
76
+ num_workers: 1
77
+ shuffle: True
78
+ pin_memory: True
79
+ persistent_workers: False
80
+
81
+ val_dataloader:
82
+ batch_size: 256
83
+ num_workers: 1
84
+ shuffle: False
85
+ pin_memory: True
86
+ persistent_workers: False
87
+
88
+ optimizer:
89
+ learning_rate: 1.0e-4
90
+ weight_decay: 1.0e-3
91
+ betas: [0.9, 0.95]
92
+
93
+ training:
94
+ device: "cuda:0"
95
+ seed: 42
96
+ debug: False
97
+ resume: True
98
+ # optimization
99
+ lr_scheduler: cosine
100
+ # Transformer needs LR warmup
101
+ lr_warmup_steps: 1000
102
+ num_epochs: 5000
103
+ gradient_accumulate_every: 1
104
+ use_ema: True
105
+ # training loop control
106
+ # in epochs
107
+ rollout_every: 50
108
+ checkpoint_every: 50
109
+ val_every: 1
110
+ sample_every: 5
111
+ # steps per epoch
112
+ max_train_steps: null
113
+ max_val_steps: null
114
+ # misc
115
+ tqdm_interval_sec: 1.0
116
+
117
+ logging:
118
+ project: diffusion_policy_debug
119
+ resume: True
120
+ mode: online
121
+ name: ${now:%Y.%m.%d-%H.%M.%S}_${name}_${task_name}
122
+ tags: ["${name}", "${task_name}", "${exp_name}"]
123
+ id: null
124
+ group: null
125
+
126
+ checkpoint:
127
+ topk:
128
+ monitor_key: test_mean_score
129
+ mode: max
130
+ k: 5
131
+ format_str: 'epoch={epoch:04d}-test_mean_score={test_mean_score:.3f}.ckpt'
132
+ save_last_ckpt: True
133
+ save_last_snapshot: False
134
+
135
+ multi_run:
136
+ run_dir: data/outputs/${now:%Y.%m.%d}/${now:%H.%M.%S}_${name}_${task_name}
137
+ wandb_name_base: ${now:%Y.%m.%d-%H.%M.%S}_${name}_${task_name}
138
+
139
+ hydra:
140
+ job:
141
+ override_dirname: ${name}
142
+ run:
143
+ dir: data/outputs/${now:%Y.%m.%d}/${now:%H.%M.%S}_${name}_${task_name}
144
+ sweep:
145
+ dir: data/outputs/${now:%Y.%m.%d}/${now:%H.%M.%S}_${name}_${task_name}
146
+ subdir: ${hydra.job.num}
@@ -0,0 +1,146 @@
1
+ defaults:
2
+ - _self_
3
+ - task: pusht_lowdim
4
+
5
+ name: train_diffusion_transformer_lowdim
6
+ _target_: diffusion_policy.workspace.train_diffusion_transformer_lowdim_workspace.TrainDiffusionTransformerLowdimWorkspace
7
+
8
+ obs_dim: ${task.obs_dim}
9
+ action_dim: ${task.action_dim}
10
+ task_name: ${task.name}
11
+ exp_name: "default"
12
+
13
+ horizon: 16
14
+ n_obs_steps: 2
15
+ n_action_steps: 8
16
+ n_latency_steps: 0
17
+ past_action_visible: False
18
+ keypoint_visible_rate: 1.0
19
+ obs_as_cond: True
20
+ pred_action_steps_only: False
21
+
22
+ policy:
23
+ _target_: diffusion_policy.policy.diffusion_transformer_lowdim_policy.DiffusionTransformerLowdimPolicy
24
+
25
+ model:
26
+ _target_: diffusion_policy.model.diffusion.transformer_for_diffusion.TransformerForDiffusion
27
+ input_dim: ${eval:'${action_dim} if ${obs_as_cond} else ${obs_dim} + ${action_dim}'}
28
+ output_dim: ${policy.model.input_dim}
29
+ horizon: ${horizon}
30
+ n_obs_steps: ${n_obs_steps}
31
+ cond_dim: ${eval:'${obs_dim} if ${obs_as_cond} else 0'}
32
+
33
+ n_layer: 8
34
+ n_head: 4
35
+ n_emb: 256
36
+ p_drop_emb: 0.0
37
+ p_drop_attn: 0.01
38
+
39
+ causal_attn: True
40
+ time_as_cond: True # if false, use BERT like encoder only arch, time as input
41
+ obs_as_cond: ${obs_as_cond}
42
+ n_cond_layers: 0 # >0: use transformer encoder for cond, otherwise use MLP
43
+
44
+ noise_scheduler:
45
+ _target_: diffusers.schedulers.scheduling_ddpm.DDPMScheduler
46
+ num_train_timesteps: 100
47
+ beta_start: 0.0001
48
+ beta_end: 0.02
49
+ beta_schedule: squaredcos_cap_v2
50
+ variance_type: fixed_small # Yilun's paper uses fixed_small_log instead, but easy to cause Nan
51
+ clip_sample: True # required when predict_epsilon=False
52
+ prediction_type: epsilon # or sample
53
+
54
+ horizon: ${horizon}
55
+ obs_dim: ${obs_dim}
56
+ action_dim: ${action_dim}
57
+ n_action_steps: ${n_action_steps}
58
+ n_obs_steps: ${n_obs_steps}
59
+ num_inference_steps: 100
60
+ obs_as_cond: ${obs_as_cond}
61
+ pred_action_steps_only: ${pred_action_steps_only}
62
+
63
+ # scheduler.step params
64
+ # predict_epsilon: True
65
+
66
+ ema:
67
+ _target_: diffusion_policy.model.diffusion.ema_model.EMAModel
68
+ update_after_step: 0
69
+ inv_gamma: 1.0
70
+ power: 0.75
71
+ min_value: 0.0
72
+ max_value: 0.9999
73
+
74
+ dataloader:
75
+ batch_size: 256
76
+ num_workers: 1
77
+ shuffle: True
78
+ pin_memory: True
79
+ persistent_workers: False
80
+
81
+ val_dataloader:
82
+ batch_size: 256
83
+ num_workers: 1
84
+ shuffle: False
85
+ pin_memory: True
86
+ persistent_workers: False
87
+
88
+ optimizer:
89
+ learning_rate: 1.0e-4
90
+ weight_decay: 1.0e-1
91
+ betas: [0.9, 0.95]
92
+
93
+ training:
94
+ device: "cuda:0"
95
+ seed: 42
96
+ debug: False
97
+ resume: True
98
+ # optimization
99
+ lr_scheduler: cosine
100
+ # Transformer needs LR warmup
101
+ lr_warmup_steps: 1000
102
+ num_epochs: 8000
103
+ gradient_accumulate_every: 1
104
+ use_ema: True
105
+ # training loop control
106
+ # in epochs
107
+ rollout_every: 50
108
+ checkpoint_every: 50
109
+ val_every: 1
110
+ sample_every: 5
111
+ # steps per epoch
112
+ max_train_steps: null
113
+ max_val_steps: null
114
+ # misc
115
+ tqdm_interval_sec: 1.0
116
+
117
+ logging:
118
+ project: diffusion_policy_debug
119
+ resume: True
120
+ mode: online
121
+ name: ${now:%Y.%m.%d-%H.%M.%S}_${name}_${task_name}
122
+ tags: ["${name}", "${task_name}", "${exp_name}"]
123
+ id: null
124
+ group: null
125
+
126
+ checkpoint:
127
+ topk:
128
+ monitor_key: test_mean_score
129
+ mode: max
130
+ k: 5
131
+ format_str: 'epoch={epoch:04d}-test_mean_score={test_mean_score:.3f}.ckpt'
132
+ save_last_ckpt: True
133
+ save_last_snapshot: False
134
+
135
+ multi_run:
136
+ run_dir: data/outputs/${now:%Y.%m.%d}/${now:%H.%M.%S}_${name}_${task_name}
137
+ wandb_name_base: ${now:%Y.%m.%d-%H.%M.%S}_${name}_${task_name}
138
+
139
+ hydra:
140
+ job:
141
+ override_dirname: ${name}
142
+ run:
143
+ dir: data/outputs/${now:%Y.%m.%d}/${now:%H.%M.%S}_${name}_${task_name}
144
+ sweep:
145
+ dir: data/outputs/${now:%Y.%m.%d}/${now:%H.%M.%S}_${name}_${task_name}
146
+ subdir: ${hydra.job.num}
@@ -0,0 +1,146 @@
1
+ defaults:
2
+ - _self_
3
+ - task: blockpush_lowdim_seed
4
+
5
+ name: train_diffusion_transformer_lowdim
6
+ _target_: diffusion_policy.workspace.train_diffusion_transformer_lowdim_workspace.TrainDiffusionTransformerLowdimWorkspace
7
+
8
+ obs_dim: ${task.obs_dim}
9
+ action_dim: ${task.action_dim}
10
+ task_name: ${task.name}
11
+ exp_name: "default"
12
+
13
+ horizon: 5
14
+ n_obs_steps: 3
15
+ n_action_steps: 1
16
+ n_latency_steps: 0
17
+ past_action_visible: False
18
+ keypoint_visible_rate: 1.0
19
+ obs_as_cond: True
20
+ pred_action_steps_only: False
21
+
22
+ policy:
23
+ _target_: diffusion_policy.policy.diffusion_transformer_lowdim_policy.DiffusionTransformerLowdimPolicy
24
+
25
+ model:
26
+ _target_: diffusion_policy.model.diffusion.transformer_for_diffusion.TransformerForDiffusion
27
+ input_dim: ${eval:'${action_dim} if ${obs_as_cond} else ${obs_dim} + ${action_dim}'}
28
+ output_dim: ${policy.model.input_dim}
29
+ horizon: ${horizon}
30
+ n_obs_steps: ${n_obs_steps}
31
+ cond_dim: ${eval:'${obs_dim} if ${obs_as_cond} else 0'}
32
+
33
+ n_layer: 8
34
+ n_head: 4
35
+ n_emb: 256
36
+ p_drop_emb: 0.0
37
+ p_drop_attn: 0.3
38
+
39
+ causal_attn: True
40
+ time_as_cond: True # if false, use BERT like encoder only arch, time as input
41
+ obs_as_cond: ${obs_as_cond}
42
+ n_cond_layers: 0 # >0: use transformer encoder for cond, otherwise use MLP
43
+
44
+ noise_scheduler:
45
+ _target_: diffusers.schedulers.scheduling_ddpm.DDPMScheduler
46
+ num_train_timesteps: 100
47
+ beta_start: 0.0001
48
+ beta_end: 0.02
49
+ beta_schedule: squaredcos_cap_v2
50
+ variance_type: fixed_small # Yilun's paper uses fixed_small_log instead, but easy to cause Nan
51
+ clip_sample: True # required when predict_epsilon=False
52
+ prediction_type: epsilon # or sample
53
+
54
+ horizon: ${horizon}
55
+ obs_dim: ${obs_dim}
56
+ action_dim: ${action_dim}
57
+ n_action_steps: ${n_action_steps}
58
+ n_obs_steps: ${n_obs_steps}
59
+ num_inference_steps: 100
60
+ obs_as_cond: ${obs_as_cond}
61
+ pred_action_steps_only: ${pred_action_steps_only}
62
+
63
+ # scheduler.step params
64
+ # predict_epsilon: True
65
+
66
+ ema:
67
+ _target_: diffusion_policy.model.diffusion.ema_model.EMAModel
68
+ update_after_step: 0
69
+ inv_gamma: 1.0
70
+ power: 0.75
71
+ min_value: 0.0
72
+ max_value: 0.9999
73
+
74
+ dataloader:
75
+ batch_size: 256
76
+ num_workers: 1
77
+ shuffle: True
78
+ pin_memory: True
79
+ persistent_workers: False
80
+
81
+ val_dataloader:
82
+ batch_size: 256
83
+ num_workers: 1
84
+ shuffle: False
85
+ pin_memory: True
86
+ persistent_workers: False
87
+
88
+ optimizer:
89
+ learning_rate: 1.0e-4
90
+ weight_decay: 1.0e-3
91
+ betas: [0.9, 0.95]
92
+
93
+ training:
94
+ device: "cuda:0"
95
+ seed: 42
96
+ debug: False
97
+ resume: True
98
+ # optimization
99
+ lr_scheduler: cosine
100
+ # Transformer needs LR warmup
101
+ lr_warmup_steps: 1000
102
+ num_epochs: 5000
103
+ gradient_accumulate_every: 1
104
+ use_ema: True
105
+ # training loop control
106
+ # in epochs
107
+ rollout_every: 50
108
+ checkpoint_every: 50
109
+ val_every: 1
110
+ sample_every: 5
111
+ # steps per epoch
112
+ max_train_steps: null
113
+ max_val_steps: null
114
+ # misc
115
+ tqdm_interval_sec: 1.0
116
+
117
+ logging:
118
+ project: diffusion_policy_debug
119
+ resume: True
120
+ mode: online
121
+ name: ${now:%Y.%m.%d-%H.%M.%S}_${name}_${task_name}
122
+ tags: ["${name}", "${task_name}", "${exp_name}"]
123
+ id: null
124
+ group: null
125
+
126
+ checkpoint:
127
+ topk:
128
+ monitor_key: test_mean_score
129
+ mode: max
130
+ k: 5
131
+ format_str: 'epoch={epoch:04d}-test_mean_score={test_mean_score:.3f}.ckpt'
132
+ save_last_ckpt: True
133
+ save_last_snapshot: False
134
+
135
+ multi_run:
136
+ run_dir: data/outputs/${now:%Y.%m.%d}/${now:%H.%M.%S}_${name}_${task_name}
137
+ wandb_name_base: ${now:%Y.%m.%d-%H.%M.%S}_${name}_${task_name}
138
+
139
+ hydra:
140
+ job:
141
+ override_dirname: ${name}
142
+ run:
143
+ dir: data/outputs/${now:%Y.%m.%d}/${now:%H.%M.%S}_${name}_${task_name}
144
+ sweep:
145
+ dir: data/outputs/${now:%Y.%m.%d}/${now:%H.%M.%S}_${name}_${task_name}
146
+ subdir: ${hydra.job.num}