forked from huggingface/diffusers
-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathunet_rl.py
More file actions
229 lines (180 loc) 路 6.72 KB
/
Copy pathunet_rl.py
File metadata and controls
229 lines (180 loc) 路 6.72 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
# model adapted from diffuser https://github.com/jannerm/diffuser/blob/main/diffuser/models/temporal.py
import torch
import torch.nn as nn
from diffusers.models.resnet import Downsample1D, ResidualTemporalBlock, Upsample1D
from ..configuration_utils import ConfigMixin, register_to_config
from ..modeling_utils import ModelMixin
from .embeddings import get_timestep_embedding
class SinusoidalPosEmb(nn.Module):
def __init__(self, dim):
super().__init__()
self.dim = dim
def forward(self, x):
return get_timestep_embedding(x, self.dim)
class RearrangeDim(nn.Module):
def __init__(self):
super().__init__()
def forward(self, tensor):
if len(tensor.shape) == 2:
return tensor[:, :, None]
if len(tensor.shape) == 3:
return tensor[:, :, None, :]
elif len(tensor.shape) == 4:
return tensor[:, :, 0, :]
else:
raise ValueError(f"`len(tensor)`: {len(tensor)} has to be 2, 3 or 4.")
class Conv1dBlock(nn.Module):
"""
Conv1d --> GroupNorm --> Mish
"""
def __init__(self, inp_channels, out_channels, kernel_size, n_groups=8):
super().__init__()
self.block = nn.Sequential(
nn.Conv1d(inp_channels, out_channels, kernel_size, padding=kernel_size // 2),
RearrangeDim(),
# Rearrange("batch channels horizon -> batch channels 1 horizon"),
nn.GroupNorm(n_groups, out_channels),
RearrangeDim(),
# Rearrange("batch channels 1 horizon -> batch channels horizon"),
nn.Mish(),
)
def forward(self, x):
return self.block(x)
class TemporalUNet(ModelMixin, ConfigMixin): # (nn.Module):
@register_to_config
def __init__(
self,
training_horizon=128,
transition_dim=14,
cond_dim=3,
predict_epsilon=False,
clip_denoised=True,
dim=32,
dim_mults=(1, 4, 8),
):
super().__init__()
self.transition_dim = transition_dim
self.cond_dim = cond_dim
self.predict_epsilon = predict_epsilon
self.clip_denoised = clip_denoised
dims = [transition_dim, *map(lambda m: dim * m, dim_mults)]
in_out = list(zip(dims[:-1], dims[1:]))
time_dim = dim
self.time_mlp = nn.Sequential(
SinusoidalPosEmb(dim),
nn.Linear(dim, dim * 4),
nn.Mish(),
nn.Linear(dim * 4, dim),
)
self.downs = nn.ModuleList([])
self.ups = nn.ModuleList([])
num_resolutions = len(in_out)
for ind, (dim_in, dim_out) in enumerate(in_out):
is_last = ind >= (num_resolutions - 1)
self.downs.append(
nn.ModuleList(
[
ResidualTemporalBlock(dim_in, dim_out, embed_dim=time_dim, horizon=training_horizon),
ResidualTemporalBlock(dim_out, dim_out, embed_dim=time_dim, horizon=training_horizon),
Downsample1D(dim_out, use_conv=True) if not is_last else nn.Identity(),
]
)
)
if not is_last:
training_horizon = training_horizon // 2
mid_dim = dims[-1]
self.mid_block1 = ResidualTemporalBlock(mid_dim, mid_dim, embed_dim=time_dim, horizon=training_horizon)
self.mid_block2 = ResidualTemporalBlock(mid_dim, mid_dim, embed_dim=time_dim, horizon=training_horizon)
for ind, (dim_in, dim_out) in enumerate(reversed(in_out[1:])):
is_last = ind >= (num_resolutions - 1)
self.ups.append(
nn.ModuleList(
[
ResidualTemporalBlock(dim_out * 2, dim_in, embed_dim=time_dim, horizon=training_horizon),
ResidualTemporalBlock(dim_in, dim_in, embed_dim=time_dim, horizon=training_horizon),
Upsample1D(dim_in, use_conv_transpose=True) if not is_last else nn.Identity(),
]
)
)
if not is_last:
training_horizon = training_horizon * 2
self.final_conv = nn.Sequential(
Conv1dBlock(dim, dim, kernel_size=5),
nn.Conv1d(dim, transition_dim, 1),
)
def forward(self, sample, timestep):
"""
x : [ batch x horizon x transition ]
"""
x = sample
x = x.permute(0, 2, 1)
t = self.time_mlp(timestep)
h = []
for resnet, resnet2, downsample in self.downs:
x = resnet(x, t)
x = resnet2(x, t)
h.append(x)
x = downsample(x)
x = self.mid_block1(x, t)
x = self.mid_block2(x, t)
for resnet, resnet2, upsample in self.ups:
x = torch.cat((x, h.pop()), dim=1)
x = resnet(x, t)
x = resnet2(x, t)
x = upsample(x)
x = self.final_conv(x)
x = x.permute(0, 2, 1)
return x
class TemporalValue(nn.Module):
def __init__(
self,
horizon,
transition_dim,
cond_dim,
dim=32,
time_dim=None,
out_dim=1,
dim_mults=(1, 2, 4, 8),
):
super().__init__()
dims = [transition_dim, *map(lambda m: dim * m, dim_mults)]
in_out = list(zip(dims[:-1], dims[1:]))
time_dim = time_dim or dim
self.time_mlp = nn.Sequential(
SinusoidalPosEmb(dim),
nn.Linear(dim, dim * 4),
nn.Mish(),
nn.Linear(dim * 4, dim),
)
self.blocks = nn.ModuleList([])
print(in_out)
for dim_in, dim_out in in_out:
self.blocks.append(
nn.ModuleList(
[
ResidualTemporalBlock(dim_in, dim_out, kernel_size=5, embed_dim=time_dim, horizon=horizon),
ResidualTemporalBlock(dim_out, dim_out, kernel_size=5, embed_dim=time_dim, horizon=horizon),
Downsample1d(dim_out),
]
)
)
horizon = horizon // 2
fc_dim = dims[-1] * max(horizon, 1)
self.final_block = nn.Sequential(
nn.Linear(fc_dim + time_dim, fc_dim // 2),
nn.Mish(),
nn.Linear(fc_dim // 2, out_dim),
)
def forward(self, x, cond, time, *args):
"""
x : [ batch x horizon x transition ]
"""
x = x.permute(0, 2, 1)
t = self.time_mlp(time)
for resnet, resnet2, downsample in self.blocks:
x = resnet(x, t)
x = resnet2(x, t)
x = downsample(x)
x = x.view(len(x), -1)
out = self.final_block(torch.cat([x, t], dim=-1))
return out