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.
- diffusion_policy/__init__.py +0 -0
- diffusion_policy/codecs/__init__.py +0 -0
- diffusion_policy/codecs/imagecodecs_numcodecs.py +1408 -0
- diffusion_policy/common/__init__.py +0 -0
- diffusion_policy/common/checkpoint_util.py +64 -0
- diffusion_policy/common/json_logger.py +117 -0
- diffusion_policy/common/normalize_util.py +251 -0
- diffusion_policy/common/pose_repr_util.py +112 -0
- diffusion_policy/common/pytorch_util.py +82 -0
- diffusion_policy/common/q_critic_util.py +135 -0
- diffusion_policy/common/replay_buffer.py +672 -0
- diffusion_policy/common/sampler.py +204 -0
- diffusion_policy/config/legacy/task/blockpush_lowdim_seed.yaml +34 -0
- diffusion_policy/config/legacy/task/blockpush_lowdim_seed_abs.yaml +34 -0
- diffusion_policy/config/legacy/task/can_image.yaml +64 -0
- diffusion_policy/config/legacy/task/can_image_abs.yaml +64 -0
- diffusion_policy/config/legacy/task/can_lowdim.yaml +45 -0
- diffusion_policy/config/legacy/task/can_lowdim_abs.yaml +46 -0
- diffusion_policy/config/legacy/task/kitchen_lowdim.yaml +33 -0
- diffusion_policy/config/legacy/task/kitchen_lowdim_abs.yaml +38 -0
- diffusion_policy/config/legacy/task/lift_abs.yaml +72 -0
- diffusion_policy/config/legacy/task/lift_image.yaml +64 -0
- diffusion_policy/config/legacy/task/lift_image_abs.yaml +63 -0
- diffusion_policy/config/legacy/task/lift_lowdim.yaml +46 -0
- diffusion_policy/config/legacy/task/lift_lowdim_abs.yaml +47 -0
- diffusion_policy/config/legacy/task/pusht_image.yaml +40 -0
- diffusion_policy/config/legacy/task/pusht_lowdim.yaml +34 -0
- diffusion_policy/config/legacy/task/real_pusht_image.yaml +47 -0
- diffusion_policy/config/legacy/task/square_abs.yaml +72 -0
- diffusion_policy/config/legacy/task/square_image.yaml +64 -0
- diffusion_policy/config/legacy/task/square_image_abs.yaml +64 -0
- diffusion_policy/config/legacy/task/square_lowdim.yaml +46 -0
- diffusion_policy/config/legacy/task/square_lowdim_abs.yaml +47 -0
- diffusion_policy/config/legacy/task/tool_hang_abs.yaml +71 -0
- diffusion_policy/config/legacy/task/tool_hang_image.yaml +63 -0
- diffusion_policy/config/legacy/task/tool_hang_image_abs.yaml +63 -0
- diffusion_policy/config/legacy/task/tool_hang_lowdim.yaml +45 -0
- diffusion_policy/config/legacy/task/tool_hang_lowdim_abs.yaml +46 -0
- diffusion_policy/config/legacy/task/transport_image.yaml +75 -0
- diffusion_policy/config/legacy/task/transport_image_abs.yaml +75 -0
- diffusion_policy/config/legacy/task/transport_lowdim.yaml +49 -0
- diffusion_policy/config/legacy/task/transport_lowdim_abs.yaml +49 -0
- diffusion_policy/config/legacy/task/umi.yaml +45 -0
- diffusion_policy/config/legacy/task/umi_image.yaml +36 -0
- diffusion_policy/config/legacy/task/umi_teleop_image.yaml +41 -0
- diffusion_policy/config/legacy/train_bet_lowdim_workspace.yaml +132 -0
- diffusion_policy/config/legacy/train_diffusion_transformer_hybrid_workspace.yaml +141 -0
- diffusion_policy/config/legacy/train_diffusion_transformer_lowdim_kitchen_workspace.yaml +146 -0
- diffusion_policy/config/legacy/train_diffusion_transformer_lowdim_pusht_workspace.yaml +146 -0
- diffusion_policy/config/legacy/train_diffusion_transformer_lowdim_workspace.yaml +146 -0
- diffusion_policy/config/legacy/train_diffusion_transformer_real_hybrid_workspace.yaml +144 -0
- diffusion_policy/config/legacy/train_diffusion_transformer_umi_teleop_hybrid_workspace.yaml +156 -0
- diffusion_policy/config/legacy/train_diffusion_unet_ddim_hybrid_workspace.yaml +140 -0
- diffusion_policy/config/legacy/train_diffusion_unet_ddim_lowdim_workspace.yaml +146 -0
- diffusion_policy/config/legacy/train_diffusion_unet_hybrid_workspace.yaml +137 -0
- diffusion_policy/config/legacy/train_diffusion_unet_image_pretrained_workspace.yaml +149 -0
- diffusion_policy/config/legacy/train_diffusion_unet_image_workspace.yaml +150 -0
- diffusion_policy/config/legacy/train_diffusion_unet_lowdim_workspace.yaml +143 -0
- diffusion_policy/config/legacy/train_diffusion_unet_real_hybrid_workspace.yaml +141 -0
- diffusion_policy/config/legacy/train_diffusion_unet_real_image_workspace.yaml +152 -0
- diffusion_policy/config/legacy/train_diffusion_unet_real_pretrained_workspace.yaml +152 -0
- diffusion_policy/config/legacy/train_diffusion_unet_timm_workspace.yaml +145 -0
- diffusion_policy/config/legacy/train_diffusion_unet_umi_hybrid_workspace.yaml +141 -0
- diffusion_policy/config/legacy/train_diffusion_unet_umi_teleop_hybrid_workspace.yaml +153 -0
- diffusion_policy/config/legacy/train_diffusion_unet_umi_timm_workspace.yaml +145 -0
- diffusion_policy/config/legacy/train_diffusion_unet_umi_workspace.yaml +150 -0
- diffusion_policy/config/legacy/train_diffusion_unet_video_workspace.yaml +128 -0
- diffusion_policy/config/legacy/train_ibc_dfo_hybrid_workspace.yaml +111 -0
- diffusion_policy/config/legacy/train_ibc_dfo_lowdim_workspace.yaml +109 -0
- diffusion_policy/config/legacy/train_ibc_dfo_real_hybrid_workspace.yaml +111 -0
- diffusion_policy/config/legacy/train_robomimic_image_workspace.yaml +92 -0
- diffusion_policy/config/legacy/train_robomimic_lowdim_workspace.yaml +92 -0
- diffusion_policy/config/legacy/train_robomimic_real_image_workspace.yaml +92 -0
- diffusion_policy/config/rl_annotate.yaml +265 -0
- diffusion_policy/config/rl_annotate_per_frame.yaml +162 -0
- diffusion_policy/config/rl_annotate_per_frame_v4.yaml +162 -0
- diffusion_policy/config/rl_annotate_v2.yaml +141 -0
- diffusion_policy/config/rl_annotate_v3.yaml +165 -0
- diffusion_policy/config/task/lift.yaml +89 -0
- diffusion_policy/config/task/square.yaml +92 -0
- diffusion_policy/config/task/tool_hang.yaml +93 -0
- diffusion_policy/config/task/tool_hang_abs.yaml +71 -0
- diffusion_policy/config/task/umi.yaml +105 -0
- diffusion_policy/config/task/umi_bimanual.yaml +135 -0
- diffusion_policy/config/task/umi_cup_wild.yaml +106 -0
- diffusion_policy/config/task/umi_image.yaml +36 -0
- diffusion_policy/config/task/umi_pick_lego.yaml +106 -0
- diffusion_policy/config/task/umi_pick_single_lego_175.yaml +106 -0
- diffusion_policy/config/task/umi_pick_single_lego_bbox.yaml +135 -0
- diffusion_policy/config/task/umi_teleop_image.yaml +41 -0
- diffusion_policy/config/train_diffusion_transformer_umi_bimanual_workspace.yaml +185 -0
- diffusion_policy/config/train_diffusion_transformer_umi_workspace.yaml +185 -0
- diffusion_policy/config/train_diffusion_unet_image_workspace.yaml +152 -0
- diffusion_policy/config/train_diffusion_unet_timm_robomimic_workspace.yaml +152 -0
- diffusion_policy/config/train_diffusion_unet_timm_umi_cup_wild.yaml +260 -0
- diffusion_policy/config/train_diffusion_unet_timm_umi_pick_lego_pile_bbox.yaml +265 -0
- diffusion_policy/config/train_diffusion_unet_timm_umi_pick_lego_pile_bbox_awr.yaml +324 -0
- diffusion_policy/config/train_diffusion_unet_timm_umi_pick_lego_pile_bbox_parl.yaml +399 -0
- diffusion_policy/config/train_diffusion_unet_timm_umi_pick_single_lego.yaml +258 -0
- diffusion_policy/config/train_diffusion_unet_timm_umi_pick_single_lego_125.yaml +260 -0
- diffusion_policy/config/train_diffusion_unet_timm_umi_pick_single_lego_175.yaml +260 -0
- diffusion_policy/config/train_diffusion_unet_timm_umi_pick_single_lego_bbox.yaml +265 -0
- diffusion_policy/config/train_diffusion_unet_timm_umi_pick_two_legos_bbox.yaml +265 -0
- diffusion_policy/config/train_diffusion_unet_timm_umi_workspace.yaml +224 -0
- diffusion_policy/config/train_diffusion_unet_timm_umi_workspace_lego.yaml +213 -0
- diffusion_policy/config/train_diffusion_unet_timm_umi_workspace_lego_resume.yaml +213 -0
- diffusion_policy/config/train_diffusion_unet_timm_umi_workspace_pick_lego.yaml +254 -0
- diffusion_policy/config/train_diffusion_unet_timm_umi_workspace_red_cup.yaml +219 -0
- diffusion_policy/config/train_diffusion_unet_umi_bimanual_workspace.yaml +203 -0
- diffusion_policy/config/train_iql_umi.yaml +186 -0
- diffusion_policy/config/train_iql_umi_v2.yaml +212 -0
- diffusion_policy/config/train_iql_umi_v3.yaml +212 -0
- diffusion_policy/dataset/__init__.py +0 -0
- diffusion_policy/dataset/base_dataset.py +76 -0
- diffusion_policy/dataset/umi_dataset.py +652 -0
- diffusion_policy/env_runner/__init__.py +0 -0
- diffusion_policy/env_runner/base_image_runner.py +9 -0
- diffusion_policy/env_runner/real_pusht_image_runner.py +10 -0
- diffusion_policy/model/__init__.py +0 -0
- diffusion_policy/model/common/__init__.py +0 -0
- diffusion_policy/model/common/dict_of_tensor_mixin.py +38 -0
- diffusion_policy/model/common/lr_scheduler.py +46 -0
- diffusion_policy/model/common/module_attr_mixin.py +14 -0
- diffusion_policy/model/common/normalizer.py +353 -0
- diffusion_policy/model/common/tensor_util.py +960 -0
- diffusion_policy/model/diffusion/__init__.py +0 -0
- diffusion_policy/model/diffusion/conditional_unet1d.py +238 -0
- diffusion_policy/model/diffusion/conv1d_components.py +46 -0
- diffusion_policy/model/diffusion/ema_model.py +88 -0
- diffusion_policy/model/diffusion/mask_generator.py +221 -0
- diffusion_policy/model/diffusion/positional_embedding.py +17 -0
- diffusion_policy/model/vision/__init__.py +0 -0
- diffusion_policy/model/vision/crop_randomizer.py +288 -0
- diffusion_policy/model/vision/multi_image_obs_encoder.py +195 -0
- diffusion_policy/model/vision/timm_obs_encoder.py +525 -0
- diffusion_policy/policy/__init__.py +0 -0
- diffusion_policy/policy/base_image_policy.py +27 -0
- diffusion_policy/policy/diffusion_unet_image_policy.py +264 -0
- diffusion_policy/policy/diffusion_unet_timm_policy.py +607 -0
- diffusion_policy/policy/parl_action_optimizer.py +675 -0
- diffusion_policy/workspace/__init__.py +0 -0
- diffusion_policy/workspace/base_workspace.py +146 -0
- diffusion_policy/workspace/train_diffusion_unet_image_workspace.py +1585 -0
- solarian_diffusion/__init__.py +32 -0
- solarian_diffusion/config.py +31 -0
- solarian_diffusion/policy_server.py +316 -0
- solarian_diffusion/rtc.py +272 -0
- solarian_diffusion/service.py +16 -0
- solarian_diffusion/trainer.py +213 -0
- solarian_diffusion-0.1.0.dist-info/METADATA +126 -0
- solarian_diffusion-0.1.0.dist-info/RECORD +159 -0
- solarian_diffusion-0.1.0.dist-info/WHEEL +5 -0
- solarian_diffusion-0.1.0.dist-info/entry_points.txt +8 -0
- solarian_diffusion-0.1.0.dist-info/licenses/LICENSE +201 -0
- solarian_diffusion-0.1.0.dist-info/licenses/NOTICE +45 -0
- solarian_diffusion-0.1.0.dist-info/top_level.txt +3 -0
- umi/__init__.py +0 -0
- umi/common/__init__.py +0 -0
- umi/common/pose_util.py +115 -0
|
@@ -0,0 +1,264 @@
|
|
|
1
|
+
from typing import Dict
|
|
2
|
+
import torch
|
|
3
|
+
import torch.nn as nn
|
|
4
|
+
import torch.nn.functional as F
|
|
5
|
+
from einops import rearrange, reduce
|
|
6
|
+
from diffusers.schedulers.scheduling_ddpm import DDPMScheduler
|
|
7
|
+
|
|
8
|
+
from diffusion_policy.model.common.normalizer import LinearNormalizer
|
|
9
|
+
from diffusion_policy.policy.base_image_policy import BaseImagePolicy
|
|
10
|
+
from diffusion_policy.model.diffusion.conditional_unet1d import ConditionalUnet1D
|
|
11
|
+
from diffusion_policy.model.diffusion.mask_generator import LowdimMaskGenerator
|
|
12
|
+
from diffusion_policy.model.vision.multi_image_obs_encoder import MultiImageObsEncoder
|
|
13
|
+
from diffusion_policy.common.pytorch_util import dict_apply
|
|
14
|
+
|
|
15
|
+
class DiffusionUnetImagePolicy(BaseImagePolicy):
|
|
16
|
+
def __init__(self,
|
|
17
|
+
shape_meta: dict,
|
|
18
|
+
noise_scheduler: DDPMScheduler,
|
|
19
|
+
obs_encoder: MultiImageObsEncoder,
|
|
20
|
+
horizon,
|
|
21
|
+
n_action_steps,
|
|
22
|
+
n_obs_steps,
|
|
23
|
+
num_inference_steps=None,
|
|
24
|
+
obs_as_global_cond=True,
|
|
25
|
+
diffusion_step_embed_dim=256,
|
|
26
|
+
down_dims=(256,512,1024),
|
|
27
|
+
kernel_size=5,
|
|
28
|
+
n_groups=8,
|
|
29
|
+
cond_predict_scale=True,
|
|
30
|
+
input_pertub=0.1,
|
|
31
|
+
# parameters passed to step
|
|
32
|
+
**kwargs):
|
|
33
|
+
super().__init__()
|
|
34
|
+
|
|
35
|
+
# parse shapes
|
|
36
|
+
action_shape = shape_meta['action']['shape']
|
|
37
|
+
assert len(action_shape) == 1
|
|
38
|
+
action_dim = action_shape[0]
|
|
39
|
+
# get feature dim
|
|
40
|
+
obs_feature_dim = obs_encoder.output_shape()[0]
|
|
41
|
+
|
|
42
|
+
# create diffusion model
|
|
43
|
+
input_dim = action_dim + obs_feature_dim
|
|
44
|
+
global_cond_dim = None
|
|
45
|
+
if obs_as_global_cond:
|
|
46
|
+
input_dim = action_dim
|
|
47
|
+
global_cond_dim = obs_feature_dim * n_obs_steps
|
|
48
|
+
|
|
49
|
+
model = ConditionalUnet1D(
|
|
50
|
+
input_dim=input_dim,
|
|
51
|
+
local_cond_dim=None,
|
|
52
|
+
global_cond_dim=global_cond_dim,
|
|
53
|
+
diffusion_step_embed_dim=diffusion_step_embed_dim,
|
|
54
|
+
down_dims=down_dims,
|
|
55
|
+
kernel_size=kernel_size,
|
|
56
|
+
n_groups=n_groups,
|
|
57
|
+
cond_predict_scale=cond_predict_scale
|
|
58
|
+
)
|
|
59
|
+
|
|
60
|
+
self.obs_encoder = obs_encoder
|
|
61
|
+
self.model = model
|
|
62
|
+
self.noise_scheduler = noise_scheduler
|
|
63
|
+
self.mask_generator = LowdimMaskGenerator(
|
|
64
|
+
action_dim=action_dim,
|
|
65
|
+
obs_dim=0 if obs_as_global_cond else obs_feature_dim,
|
|
66
|
+
max_n_obs_steps=n_obs_steps,
|
|
67
|
+
fix_obs_steps=True,
|
|
68
|
+
action_visible=False
|
|
69
|
+
)
|
|
70
|
+
self.normalizer = LinearNormalizer()
|
|
71
|
+
self.horizon = horizon
|
|
72
|
+
self.obs_feature_dim = obs_feature_dim
|
|
73
|
+
self.action_dim = action_dim
|
|
74
|
+
self.n_action_steps = n_action_steps
|
|
75
|
+
self.n_obs_steps = n_obs_steps
|
|
76
|
+
self.obs_as_global_cond = obs_as_global_cond
|
|
77
|
+
self.input_pertub = input_pertub
|
|
78
|
+
self.kwargs = kwargs
|
|
79
|
+
|
|
80
|
+
if num_inference_steps is None:
|
|
81
|
+
num_inference_steps = noise_scheduler.config.num_train_timesteps
|
|
82
|
+
self.num_inference_steps = num_inference_steps
|
|
83
|
+
|
|
84
|
+
# ========= inference ============
|
|
85
|
+
def conditional_sample(self,
|
|
86
|
+
condition_data, condition_mask,
|
|
87
|
+
local_cond=None, global_cond=None,
|
|
88
|
+
generator=None,
|
|
89
|
+
# keyword arguments to scheduler.step
|
|
90
|
+
**kwargs
|
|
91
|
+
):
|
|
92
|
+
model = self.model
|
|
93
|
+
scheduler = self.noise_scheduler
|
|
94
|
+
|
|
95
|
+
trajectory = torch.randn(
|
|
96
|
+
size=condition_data.shape,
|
|
97
|
+
dtype=condition_data.dtype,
|
|
98
|
+
device=condition_data.device,
|
|
99
|
+
generator=generator)
|
|
100
|
+
|
|
101
|
+
# set step values
|
|
102
|
+
scheduler.set_timesteps(self.num_inference_steps)
|
|
103
|
+
|
|
104
|
+
for t in scheduler.timesteps:
|
|
105
|
+
# 1. apply conditioning
|
|
106
|
+
trajectory[condition_mask] = condition_data[condition_mask]
|
|
107
|
+
|
|
108
|
+
# 2. predict model output
|
|
109
|
+
model_output = model(trajectory, t,
|
|
110
|
+
local_cond=local_cond, global_cond=global_cond)
|
|
111
|
+
|
|
112
|
+
# 3. compute previous image: x_t -> x_t-1
|
|
113
|
+
trajectory = scheduler.step(
|
|
114
|
+
model_output, t, trajectory,
|
|
115
|
+
generator=generator,
|
|
116
|
+
**kwargs
|
|
117
|
+
).prev_sample
|
|
118
|
+
|
|
119
|
+
# finally make sure conditioning is enforced
|
|
120
|
+
trajectory[condition_mask] = condition_data[condition_mask]
|
|
121
|
+
|
|
122
|
+
return trajectory
|
|
123
|
+
|
|
124
|
+
|
|
125
|
+
def predict_action(self, obs_dict: Dict[str, torch.Tensor]) -> Dict[str, torch.Tensor]:
|
|
126
|
+
"""
|
|
127
|
+
obs_dict: must include "obs" key
|
|
128
|
+
result: must include "action" key
|
|
129
|
+
"""
|
|
130
|
+
assert 'past_action' not in obs_dict # not implemented yet
|
|
131
|
+
# normalize input
|
|
132
|
+
nobs = self.normalizer.normalize(obs_dict)
|
|
133
|
+
value = next(iter(nobs.values()))
|
|
134
|
+
B, To = value.shape[:2]
|
|
135
|
+
T = self.horizon
|
|
136
|
+
Da = self.action_dim
|
|
137
|
+
Do = self.obs_feature_dim
|
|
138
|
+
To = self.n_obs_steps
|
|
139
|
+
|
|
140
|
+
# build input
|
|
141
|
+
device = self.device
|
|
142
|
+
dtype = self.dtype
|
|
143
|
+
|
|
144
|
+
# handle different ways of passing observation
|
|
145
|
+
local_cond = None
|
|
146
|
+
global_cond = None
|
|
147
|
+
if self.obs_as_global_cond:
|
|
148
|
+
# condition through global feature
|
|
149
|
+
this_nobs = dict_apply(nobs, lambda x: x[:,:To,...].reshape(-1,*x.shape[2:]))
|
|
150
|
+
nobs_features = self.obs_encoder(this_nobs)
|
|
151
|
+
# reshape back to B, Do
|
|
152
|
+
global_cond = nobs_features.reshape(B, -1)
|
|
153
|
+
# empty data for action
|
|
154
|
+
cond_data = torch.zeros(size=(B, T, Da), device=device, dtype=dtype)
|
|
155
|
+
cond_mask = torch.zeros_like(cond_data, dtype=torch.bool)
|
|
156
|
+
else:
|
|
157
|
+
# condition through impainting
|
|
158
|
+
this_nobs = dict_apply(nobs, lambda x: x[:,:To,...].reshape(-1,*x.shape[2:]))
|
|
159
|
+
nobs_features = self.obs_encoder(this_nobs)
|
|
160
|
+
# reshape back to B, T, Do
|
|
161
|
+
nobs_features = nobs_features.reshape(B, To, -1)
|
|
162
|
+
cond_data = torch.zeros(size=(B, T, Da+Do), device=device, dtype=dtype)
|
|
163
|
+
cond_mask = torch.zeros_like(cond_data, dtype=torch.bool)
|
|
164
|
+
cond_data[:,:To,Da:] = nobs_features
|
|
165
|
+
cond_mask[:,:To,Da:] = True
|
|
166
|
+
|
|
167
|
+
# run sampling
|
|
168
|
+
nsample = self.conditional_sample(
|
|
169
|
+
cond_data,
|
|
170
|
+
cond_mask,
|
|
171
|
+
local_cond=local_cond,
|
|
172
|
+
global_cond=global_cond,
|
|
173
|
+
**self.kwargs)
|
|
174
|
+
|
|
175
|
+
# unnormalize prediction
|
|
176
|
+
naction_pred = nsample[...,:Da]
|
|
177
|
+
action_pred = self.normalizer['action'].unnormalize(naction_pred)
|
|
178
|
+
|
|
179
|
+
# get action
|
|
180
|
+
start = To - 1
|
|
181
|
+
end = start + self.n_action_steps
|
|
182
|
+
action = action_pred[:,start:end]
|
|
183
|
+
|
|
184
|
+
result = {
|
|
185
|
+
'action': action,
|
|
186
|
+
'action_pred': action_pred
|
|
187
|
+
}
|
|
188
|
+
return result
|
|
189
|
+
|
|
190
|
+
# ========= training ============
|
|
191
|
+
def set_normalizer(self, normalizer: LinearNormalizer):
|
|
192
|
+
self.normalizer.load_state_dict(normalizer.state_dict())
|
|
193
|
+
|
|
194
|
+
def compute_loss(self, batch):
|
|
195
|
+
# normalize input
|
|
196
|
+
assert 'valid_mask' not in batch
|
|
197
|
+
nobs = self.normalizer.normalize(batch['obs'])
|
|
198
|
+
nactions = self.normalizer['action'].normalize(batch['action'])
|
|
199
|
+
batch_size = nactions.shape[0]
|
|
200
|
+
horizon = nactions.shape[1]
|
|
201
|
+
|
|
202
|
+
# handle different ways of passing observation
|
|
203
|
+
local_cond = None
|
|
204
|
+
global_cond = None
|
|
205
|
+
trajectory = nactions
|
|
206
|
+
cond_data = trajectory
|
|
207
|
+
if self.obs_as_global_cond:
|
|
208
|
+
# reshape B, T, ... to B*T
|
|
209
|
+
this_nobs = dict_apply(nobs,
|
|
210
|
+
lambda x: x[:,:self.n_obs_steps,...].reshape(-1,*x.shape[2:]))
|
|
211
|
+
nobs_features = self.obs_encoder(this_nobs)
|
|
212
|
+
# reshape back to B, Do
|
|
213
|
+
global_cond = nobs_features.reshape(batch_size, -1)
|
|
214
|
+
else:
|
|
215
|
+
# reshape B, T, ... to B*T
|
|
216
|
+
this_nobs = dict_apply(nobs, lambda x: x.reshape(-1, *x.shape[2:]))
|
|
217
|
+
nobs_features = self.obs_encoder(this_nobs)
|
|
218
|
+
# reshape back to B, T, Do
|
|
219
|
+
nobs_features = nobs_features.reshape(batch_size, horizon, -1)
|
|
220
|
+
cond_data = torch.cat([nactions, nobs_features], dim=-1)
|
|
221
|
+
trajectory = cond_data.detach()
|
|
222
|
+
|
|
223
|
+
# generate impainting mask
|
|
224
|
+
condition_mask = self.mask_generator(trajectory.shape)
|
|
225
|
+
|
|
226
|
+
# Sample noise that we'll add to the images
|
|
227
|
+
noise = torch.randn(trajectory.shape, device=trajectory.device)
|
|
228
|
+
# input perturbation by adding additonal noise to alleviate exposure bias
|
|
229
|
+
# reference: https://github.com/forever208/DDPM-IP
|
|
230
|
+
noise_new = noise + self.input_pertub * torch.randn(trajectory.shape, device=trajectory.device)
|
|
231
|
+
|
|
232
|
+
# Sample a random timestep for each image
|
|
233
|
+
timesteps = torch.randint(
|
|
234
|
+
0, self.noise_scheduler.config.num_train_timesteps,
|
|
235
|
+
(trajectory.shape[0],), device=trajectory.device
|
|
236
|
+
).long()
|
|
237
|
+
# Add noise to the clean images according to the noise magnitude at each timestep
|
|
238
|
+
# (this is the forward diffusion process)
|
|
239
|
+
noisy_trajectory = self.noise_scheduler.add_noise(
|
|
240
|
+
trajectory, noise_new, timesteps)
|
|
241
|
+
|
|
242
|
+
# compute loss mask
|
|
243
|
+
loss_mask = ~condition_mask
|
|
244
|
+
|
|
245
|
+
# apply conditioning
|
|
246
|
+
noisy_trajectory[condition_mask] = cond_data[condition_mask]
|
|
247
|
+
|
|
248
|
+
# Predict the noise residual
|
|
249
|
+
pred = self.model(noisy_trajectory, timesteps,
|
|
250
|
+
local_cond=local_cond, global_cond=global_cond)
|
|
251
|
+
|
|
252
|
+
pred_type = self.noise_scheduler.config.prediction_type
|
|
253
|
+
if pred_type == 'epsilon':
|
|
254
|
+
target = noise
|
|
255
|
+
elif pred_type == 'sample':
|
|
256
|
+
target = trajectory
|
|
257
|
+
else:
|
|
258
|
+
raise ValueError(f"Unsupported prediction type {pred_type}")
|
|
259
|
+
|
|
260
|
+
loss = F.mse_loss(pred, target, reduction='none')
|
|
261
|
+
loss = loss * loss_mask.type(loss.dtype)
|
|
262
|
+
loss = reduce(loss, 'b ... -> b (...)', 'mean')
|
|
263
|
+
loss = loss.mean()
|
|
264
|
+
return loss
|