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 +1,38 @@
1
+ import torch
2
+ import torch.nn as nn
3
+
4
+ class DictOfTensorMixin(nn.Module):
5
+ def __init__(self, params_dict=None):
6
+ super().__init__()
7
+ if params_dict is None:
8
+ params_dict = nn.ParameterDict()
9
+ self.params_dict = params_dict
10
+
11
+ @property
12
+ def device(self):
13
+ return next(iter(self.parameters())).device
14
+
15
+ def _load_from_state_dict(self, state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs):
16
+ def dfs_add(dest, keys, value: torch.Tensor):
17
+ if len(keys) == 1:
18
+ dest[keys[0]] = value
19
+ return
20
+
21
+ if keys[0] not in dest:
22
+ dest[keys[0]] = nn.ParameterDict()
23
+ dfs_add(dest[keys[0]], keys[1:], value)
24
+
25
+ def load_dict(state_dict, prefix):
26
+ out_dict = nn.ParameterDict()
27
+ for key, value in state_dict.items():
28
+ value: torch.Tensor
29
+ if key.startswith(prefix):
30
+ param_keys = key[len(prefix):].split('.')[1:]
31
+ # if len(param_keys) == 0:
32
+ # import pdb; pdb.set_trace()
33
+ dfs_add(out_dict, param_keys, value.clone())
34
+ return out_dict
35
+
36
+ self.params_dict = load_dict(state_dict, prefix + 'params_dict')
37
+ self.params_dict.requires_grad_(False)
38
+ return
@@ -0,0 +1,46 @@
1
+ from diffusers.optimization import (
2
+ Union, SchedulerType, Optional,
3
+ Optimizer, TYPE_TO_SCHEDULER_FUNCTION
4
+ )
5
+
6
+ def get_scheduler(
7
+ name: Union[str, SchedulerType],
8
+ optimizer: Optimizer,
9
+ num_warmup_steps: Optional[int] = None,
10
+ num_training_steps: Optional[int] = None,
11
+ **kwargs
12
+ ):
13
+ """
14
+ Added kwargs vs diffuser's original implementation
15
+
16
+ Unified API to get any scheduler from its name.
17
+
18
+ Args:
19
+ name (`str` or `SchedulerType`):
20
+ The name of the scheduler to use.
21
+ optimizer (`torch.optim.Optimizer`):
22
+ The optimizer that will be used during training.
23
+ num_warmup_steps (`int`, *optional*):
24
+ The number of warmup steps to do. This is not required by all schedulers (hence the argument being
25
+ optional), the function will raise an error if it's unset and the scheduler type requires it.
26
+ num_training_steps (`int``, *optional*):
27
+ The number of training steps to do. This is not required by all schedulers (hence the argument being
28
+ optional), the function will raise an error if it's unset and the scheduler type requires it.
29
+ """
30
+ name = SchedulerType(name)
31
+ schedule_func = TYPE_TO_SCHEDULER_FUNCTION[name]
32
+ if name == SchedulerType.CONSTANT:
33
+ return schedule_func(optimizer, **kwargs)
34
+
35
+ # All other schedulers require `num_warmup_steps`
36
+ if num_warmup_steps is None:
37
+ raise ValueError(f"{name} requires `num_warmup_steps`, please provide that argument.")
38
+
39
+ if name == SchedulerType.CONSTANT_WITH_WARMUP:
40
+ return schedule_func(optimizer, num_warmup_steps=num_warmup_steps, **kwargs)
41
+
42
+ # All other schedulers require `num_training_steps`
43
+ if num_training_steps is None:
44
+ raise ValueError(f"{name} requires `num_training_steps`, please provide that argument.")
45
+
46
+ return schedule_func(optimizer, num_warmup_steps=num_warmup_steps, num_training_steps=num_training_steps, **kwargs)
@@ -0,0 +1,14 @@
1
+ import torch.nn as nn
2
+
3
+ class ModuleAttrMixin(nn.Module):
4
+ def __init__(self):
5
+ super().__init__()
6
+ self._dummy_variable = nn.Parameter(requires_grad=False)
7
+
8
+ @property
9
+ def device(self):
10
+ return next(iter(self.parameters())).device
11
+
12
+ @property
13
+ def dtype(self):
14
+ return next(iter(self.parameters())).dtype
@@ -0,0 +1,353 @@
1
+ from typing import Union, Dict
2
+
3
+ import unittest
4
+ import zarr
5
+ import numpy as np
6
+ import torch
7
+ import torch.nn as nn
8
+ from diffusion_policy.common.pytorch_util import dict_apply
9
+ from diffusion_policy.model.common.dict_of_tensor_mixin import DictOfTensorMixin
10
+
11
+
12
+ class LinearNormalizer(DictOfTensorMixin):
13
+ avaliable_modes = ['limits', 'gaussian']
14
+
15
+ @torch.no_grad()
16
+ def fit(self,
17
+ data: Union[Dict, torch.Tensor, np.ndarray, zarr.Array],
18
+ last_n_dims=1,
19
+ dtype=torch.float32,
20
+ mode='limits',
21
+ output_max=1.,
22
+ output_min=-1.,
23
+ range_eps=1e-4,
24
+ fit_offset=True):
25
+ if isinstance(data, dict):
26
+ for key, value in data.items():
27
+ self.params_dict[key] = _fit(value,
28
+ last_n_dims=last_n_dims,
29
+ dtype=dtype,
30
+ mode=mode,
31
+ output_max=output_max,
32
+ output_min=output_min,
33
+ range_eps=range_eps,
34
+ fit_offset=fit_offset)
35
+ else:
36
+ self.params_dict['_default'] = _fit(data,
37
+ last_n_dims=last_n_dims,
38
+ dtype=dtype,
39
+ mode=mode,
40
+ output_max=output_max,
41
+ output_min=output_min,
42
+ range_eps=range_eps,
43
+ fit_offset=fit_offset)
44
+
45
+ def __call__(self, x: Union[Dict, torch.Tensor, np.ndarray]) -> torch.Tensor:
46
+ return self.normalize(x)
47
+
48
+ def __getitem__(self, key: str):
49
+ return SingleFieldLinearNormalizer(self.params_dict[key])
50
+
51
+ def __setitem__(self, key: str , value: 'SingleFieldLinearNormalizer'):
52
+ self.params_dict[key] = value.params_dict
53
+
54
+ def _normalize_impl(self, x, forward=True):
55
+ if isinstance(x, dict):
56
+ result = dict()
57
+ for key, value in x.items():
58
+ params = self.params_dict[key]
59
+ result[key] = _normalize(value, params, forward=forward)
60
+ return result
61
+ else:
62
+ if '_default' not in self.params_dict:
63
+ raise RuntimeError("Not initialized")
64
+ params = self.params_dict['_default']
65
+ return _normalize(x, params, forward=forward)
66
+
67
+ def normalize(self, x: Union[Dict, torch.Tensor, np.ndarray]) -> torch.Tensor:
68
+ return self._normalize_impl(x, forward=True)
69
+
70
+ def unnormalize(self, x: Union[Dict, torch.Tensor, np.ndarray]) -> torch.Tensor:
71
+ return self._normalize_impl(x, forward=False)
72
+
73
+ def get_input_stats(self) -> Dict:
74
+ if len(self.params_dict) == 0:
75
+ raise RuntimeError("Not initialized")
76
+ if len(self.params_dict) == 1 and '_default' in self.params_dict:
77
+ return self.params_dict['_default']['input_stats']
78
+
79
+ result = dict()
80
+ for key, value in self.params_dict.items():
81
+ if key != '_default':
82
+ result[key] = value['input_stats']
83
+ return result
84
+
85
+
86
+ def get_output_stats(self, key='_default'):
87
+ input_stats = self.get_input_stats()
88
+ if 'min' in input_stats:
89
+ # no dict
90
+ return dict_apply(input_stats, self.normalize)
91
+
92
+ result = dict()
93
+ for key, group in input_stats.items():
94
+ this_dict = dict()
95
+ for name, value in group.items():
96
+ this_dict[name] = self.normalize({key:value})[key]
97
+ result[key] = this_dict
98
+ return result
99
+
100
+
101
+ class SingleFieldLinearNormalizer(DictOfTensorMixin):
102
+ avaliable_modes = ['limits', 'gaussian']
103
+
104
+ @torch.no_grad()
105
+ def fit(self,
106
+ data: Union[torch.Tensor, np.ndarray, zarr.Array],
107
+ last_n_dims=1,
108
+ dtype=torch.float32,
109
+ mode='limits',
110
+ output_max=1.,
111
+ output_min=-1.,
112
+ range_eps=1e-4,
113
+ fit_offset=True):
114
+ self.params_dict = _fit(data,
115
+ last_n_dims=last_n_dims,
116
+ dtype=dtype,
117
+ mode=mode,
118
+ output_max=output_max,
119
+ output_min=output_min,
120
+ range_eps=range_eps,
121
+ fit_offset=fit_offset)
122
+
123
+ @classmethod
124
+ def create_fit(cls, data: Union[torch.Tensor, np.ndarray, zarr.Array], **kwargs):
125
+ obj = cls()
126
+ obj.fit(data, **kwargs)
127
+ return obj
128
+
129
+ @classmethod
130
+ def create_manual(cls,
131
+ scale: Union[torch.Tensor, np.ndarray],
132
+ offset: Union[torch.Tensor, np.ndarray],
133
+ input_stats_dict: Dict[str, Union[torch.Tensor, np.ndarray]]):
134
+ def to_tensor(x):
135
+ if not isinstance(x, torch.Tensor):
136
+ x = torch.from_numpy(x)
137
+ x = x.flatten()
138
+ return x
139
+
140
+ # check
141
+ for x in [offset] + list(input_stats_dict.values()):
142
+ assert x.shape == scale.shape
143
+ assert x.dtype == scale.dtype
144
+
145
+ params_dict = nn.ParameterDict({
146
+ 'scale': to_tensor(scale),
147
+ 'offset': to_tensor(offset),
148
+ 'input_stats': nn.ParameterDict(
149
+ dict_apply(input_stats_dict, to_tensor))
150
+ })
151
+ return cls(params_dict)
152
+
153
+ @classmethod
154
+ def create_identity(cls, dtype=torch.float32):
155
+ scale = torch.tensor([1], dtype=dtype)
156
+ offset = torch.tensor([0], dtype=dtype)
157
+ input_stats_dict = {
158
+ 'min': torch.tensor([-1], dtype=dtype),
159
+ 'max': torch.tensor([1], dtype=dtype),
160
+ 'mean': torch.tensor([0], dtype=dtype),
161
+ 'std': torch.tensor([1], dtype=dtype)
162
+ }
163
+ return cls.create_manual(scale, offset, input_stats_dict)
164
+
165
+ def normalize(self, x: Union[torch.Tensor, np.ndarray]) -> torch.Tensor:
166
+ return _normalize(x, self.params_dict, forward=True)
167
+
168
+ def unnormalize(self, x: Union[torch.Tensor, np.ndarray]) -> torch.Tensor:
169
+ return _normalize(x, self.params_dict, forward=False)
170
+
171
+ def get_input_stats(self):
172
+ return self.params_dict['input_stats']
173
+
174
+ def get_output_stats(self):
175
+ return dict_apply(self.params_dict['input_stats'], self.normalize)
176
+
177
+ def __call__(self, x: Union[torch.Tensor, np.ndarray]) -> torch.Tensor:
178
+ return self.normalize(x)
179
+
180
+
181
+
182
+ def _fit(data: Union[torch.Tensor, np.ndarray, zarr.Array],
183
+ last_n_dims=1,
184
+ dtype=torch.float32,
185
+ mode='limits',
186
+ output_max=1.,
187
+ output_min=-1.,
188
+ range_eps=1e-4,
189
+ fit_offset=True):
190
+ assert mode in ['limits', 'gaussian']
191
+ assert last_n_dims >= 0
192
+ assert output_max > output_min
193
+
194
+ # convert data to torch and type
195
+ if isinstance(data, zarr.Array):
196
+ data = data[:]
197
+ if isinstance(data, np.ndarray):
198
+ data = torch.from_numpy(data)
199
+ if dtype is not None:
200
+ data = data.type(dtype)
201
+
202
+ # convert shape
203
+ dim = 1
204
+ if last_n_dims > 0:
205
+ dim = np.prod(data.shape[-last_n_dims:])
206
+ data = data.reshape(-1,dim)
207
+
208
+ # compute input stats min max mean std
209
+ input_min, _ = data.min(axis=0)
210
+ input_max, _ = data.max(axis=0)
211
+ input_mean = data.mean(axis=0)
212
+ input_std = data.std(axis=0)
213
+
214
+ # compute scale and offset
215
+ if mode == 'limits':
216
+ if fit_offset:
217
+ # unit scale
218
+ input_range = input_max - input_min
219
+ ignore_dim = input_range < range_eps
220
+ input_range[ignore_dim] = output_max - output_min
221
+ scale = (output_max - output_min) / input_range
222
+ offset = output_min - scale * input_min
223
+ offset[ignore_dim] = (output_max + output_min) / 2 - input_min[ignore_dim]
224
+ # ignore dims scaled to mean of output max and min
225
+ else:
226
+ # use this when data is pre-zero-centered.
227
+ assert output_max > 0
228
+ assert output_min < 0
229
+ # unit abs
230
+ output_abs = min(abs(output_min), abs(output_max))
231
+ input_abs = torch.maximum(torch.abs(input_min), torch.abs(input_max))
232
+ ignore_dim = input_abs < range_eps
233
+ input_abs[ignore_dim] = output_abs
234
+ # don't scale constant channels
235
+ scale = output_abs / input_abs
236
+ offset = torch.zeros_like(input_mean)
237
+ elif mode == 'gaussian':
238
+ ignore_dim = input_std < range_eps
239
+ scale = input_std.clone()
240
+ scale[ignore_dim] = 1
241
+ scale = 1 / scale
242
+
243
+ if fit_offset:
244
+ offset = - input_mean * scale
245
+ else:
246
+ offset = torch.zeros_like(input_mean)
247
+
248
+ # save
249
+ this_params = nn.ParameterDict({
250
+ 'scale': scale,
251
+ 'offset': offset,
252
+ 'input_stats': nn.ParameterDict({
253
+ 'min': input_min,
254
+ 'max': input_max,
255
+ 'mean': input_mean,
256
+ 'std': input_std
257
+ })
258
+ })
259
+ for p in this_params.parameters():
260
+ p.requires_grad_(False)
261
+ return this_params
262
+
263
+
264
+ def _normalize(x, params, forward=True):
265
+ assert 'scale' in params
266
+ if isinstance(x, np.ndarray):
267
+ x = torch.from_numpy(x)
268
+ scale = params['scale']
269
+ offset = params['offset']
270
+ x = x.to(device=scale.device, dtype=scale.dtype)
271
+ src_shape = x.shape
272
+ x = x.reshape(-1, scale.shape[0])
273
+ if forward:
274
+ x = x * scale + offset
275
+ else:
276
+ x = (x - offset) / scale
277
+ x = x.reshape(src_shape)
278
+ return x
279
+
280
+
281
+ def test():
282
+ data = torch.zeros((100,10,9,2)).uniform_()
283
+ data[...,0,0] = 0
284
+
285
+ normalizer = SingleFieldLinearNormalizer()
286
+ normalizer.fit(data, mode='limits', last_n_dims=2)
287
+ datan = normalizer.normalize(data)
288
+ assert datan.shape == data.shape
289
+ assert np.allclose(datan.max(), 1.)
290
+ assert np.allclose(datan.min(), -1.)
291
+ dataun = normalizer.unnormalize(datan)
292
+ assert torch.allclose(data, dataun, atol=1e-7)
293
+
294
+ input_stats = normalizer.get_input_stats()
295
+ output_stats = normalizer.get_output_stats()
296
+
297
+ normalizer = SingleFieldLinearNormalizer()
298
+ normalizer.fit(data, mode='limits', last_n_dims=1, fit_offset=False)
299
+ datan = normalizer.normalize(data)
300
+ assert datan.shape == data.shape
301
+ assert np.allclose(datan.max(), 1., atol=1e-3)
302
+ assert np.allclose(datan.min(), 0., atol=1e-3)
303
+ dataun = normalizer.unnormalize(datan)
304
+ assert torch.allclose(data, dataun, atol=1e-7)
305
+
306
+ data = torch.zeros((100,10,9,2)).uniform_()
307
+ normalizer = SingleFieldLinearNormalizer()
308
+ normalizer.fit(data, mode='gaussian', last_n_dims=0)
309
+ datan = normalizer.normalize(data)
310
+ assert datan.shape == data.shape
311
+ assert np.allclose(datan.mean(), 0., atol=1e-3)
312
+ assert np.allclose(datan.std(), 1., atol=1e-3)
313
+ dataun = normalizer.unnormalize(datan)
314
+ assert torch.allclose(data, dataun, atol=1e-7)
315
+
316
+
317
+ # dict
318
+ data = torch.zeros((100,10,9,2)).uniform_()
319
+ data[...,0,0] = 0
320
+
321
+ normalizer = LinearNormalizer()
322
+ normalizer.fit(data, mode='limits', last_n_dims=2)
323
+ datan = normalizer.normalize(data)
324
+ assert datan.shape == data.shape
325
+ assert np.allclose(datan.max(), 1.)
326
+ assert np.allclose(datan.min(), -1.)
327
+ dataun = normalizer.unnormalize(datan)
328
+ assert torch.allclose(data, dataun, atol=1e-7)
329
+
330
+ input_stats = normalizer.get_input_stats()
331
+ output_stats = normalizer.get_output_stats()
332
+
333
+ data = {
334
+ 'obs': torch.zeros((1000,128,9,2)).uniform_() * 512,
335
+ 'action': torch.zeros((1000,128,2)).uniform_() * 512
336
+ }
337
+ normalizer = LinearNormalizer()
338
+ normalizer.fit(data)
339
+ datan = normalizer.normalize(data)
340
+ dataun = normalizer.unnormalize(datan)
341
+ for key in data:
342
+ assert torch.allclose(data[key], dataun[key], atol=1e-4)
343
+
344
+ input_stats = normalizer.get_input_stats()
345
+ output_stats = normalizer.get_output_stats()
346
+
347
+ state_dict = normalizer.state_dict()
348
+ n = LinearNormalizer()
349
+ n.load_state_dict(state_dict)
350
+ datan = n.normalize(data)
351
+ dataun = n.unnormalize(datan)
352
+ for key in data:
353
+ assert torch.allclose(data[key], dataun[key], atol=1e-4)