-
Notifications
You must be signed in to change notification settings - Fork 5
/
models.py
196 lines (173 loc) · 7.23 KB
/
models.py
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
from torch import nn
import torch
import torch.nn.functional as F
from guided_diffusion.unet import UNetModel
from einops import rearrange
import numpy as np
class AE(nn.Module):
def __init__(self, input_dim, z_dim, num_layers=2, channels=128):
super().__init__()
self.input_dim = input_dim
self.z_dim = z_dim
en_layers = []
de_layers = []
en_layers.append(nn.Linear(input_dim, channels))
for i in range(num_layers-1):
en_layers.append(nn.Linear(channels, channels))
en_layers.append(nn.ReLU())
en_layers.append(nn.Linear(channels, z_dim*2))
self.en_layers = nn.Sequential(*en_layers)
de_layers.append(nn.Linear(z_dim, channels))
for i in range(num_layers-1):
de_layers.append(nn.Linear(channels, channels))
de_layers.append(nn.ReLU())
de_layers.append(nn.Linear(channels, input_dim))
self.de_layers = nn.Sequential(*de_layers)
def forward(self, x):
mu_logvar = self.en_layers(x)
mu, logvar = mu_logvar[:, :self.z_dim], mu_logvar[:, self.z_dim:]
z = mu + torch.randn_like(mu)*torch.exp(logvar/2)
x_recon = self.de_layers(z)
return x_recon, z, mu, logvar
def encode(self, x):
mu_logvar = self.en_layers(x)
mu, logvar = mu_logvar[:, :self.z_dim], mu_logvar[:, self.z_dim:]
z = mu + torch.randn_like(mu)*torch.exp(logvar/2)
return z
class MLP(nn.Module):
def __init__(self, input_dim=2, output_dim = 2, hidden_num=100):
super().__init__()
self.fc1 = nn.Linear(input_dim+1, hidden_num, bias=True)
self.fc2 = nn.Linear(hidden_num, hidden_num, bias=True)
self.fc3 = nn.Linear(hidden_num, output_dim, bias=True)
self.act = lambda x: torch.tanh(x)
def forward(self, x_input, t):
inputs = torch.cat([x_input, t], dim=1)
x = self.fc1(inputs)
x = self.act(x)
x = self.fc2(x)
x = self.act(x)
x = self.fc3(x)
return x
class UNet(nn.Module):
def __init__(self, input_nc=1, output_nc = 1, ngf=64, norm_layer='bn', use_dropout=False):
super().__init__()
self.conv_en1 = nn.Conv2d(input_nc, ngf, kernel_size=3, stride=1, padding=1, bias=False)
self.norm1 = nn.BatchNorm2d(ngf)
self.conv_en2 = nn.Conv2d(ngf, ngf*2, kernel_size=3, stride=1, padding=1, bias=False)
self.norm2 = nn.BatchNorm2d(ngf*2)
self.conv_bottleneck = nn.Conv2d(ngf*2, ngf*2, kernel_size=3, stride=1, padding=1, bias=False)
self.conv_de1 = nn.Conv2d(ngf*2, ngf, kernel_size=3, stride=1, padding=1, bias=False)
self.norm3 = nn.BatchNorm2d(ngf)
self.conv_de2 = nn.Conv2d(ngf, output_nc, kernel_size=3, stride=1, padding=1, bias=False)
self.act = nn.ReLU()
self.use_dropout = use_dropout
self.dropout = nn.Dropout(0.5)
# Bilinear up/down sample
self.upsample = nn.Upsample(scale_factor=2, mode='bilinear', align_corners=True)
self.downsample = nn.AvgPool2d(2, stride=2)
def forward(self, x, t=None):
x_en1 = self.conv_en1(x)
x_en1 = self.act(x_en1)
x_en1 = self.norm1(x_en1)
if self.use_dropout:
x_en1 = self.dropout(x_en1)
x_en1 = self.downsample(x_en1)
x_en2 = self.conv_en2(x_en1)
x_en2 = self.act(x_en2)
x_en2 = self.norm2(x_en2)
if self.use_dropout:
x_en2 = self.dropout(x_en2)
x_en2 = self.downsample(x_en2)
x_bottleneck = self.conv_bottleneck(x_en2)
x_bottleneck = self.act(x_bottleneck)
x_de1 = self.upsample(x_bottleneck)
x_de1 = self.conv_de1(x_de1)
x_de1 = self.act(x_de1)
x_de1 = self.norm3(x_de1)
if self.use_dropout:
x_de1 = self.dropout(x_de1)
x_de1 = x_de1 + x_en1
x_de2 = self.upsample(x_de1)
x_de2 = self.conv_de2(x_de2)
x_de2 = x_de2
return x_de2
class UNetAE(nn.Module):
def __init__(self, res, input_nc=1, output_nc = 2, ngf=64, norm_layer=nn.BatchNorm2d, large = False, use_dropout=False, encoder_only = False):
super().__init__()
self.input_nc = input_nc
if not large:
self.encoder = UNet(input_nc, output_nc, ngf, norm_layer, use_dropout)
if not encoder_only:
self.decoder = UNet(input_nc, input_nc, ngf, norm_layer, use_dropout)
else:
self.encoder = UNetModel(res, input_nc, 32, output_nc, no_time = True)
if not encoder_only:
self.decoder = UNetModel(res, input_nc, 32, input_nc, no_time = True)
def forward(self, x):
z, mu, logvar = self.encode(x)
x_recon = self.decode(z)
return x_recon, z, mu, logvar
def encode(self, x):
mu_logvar = self.encoder(x)
mu, logvar = mu_logvar[:, :self.input_nc], mu_logvar[:, self.input_nc:]
z = mu + torch.randn_like(mu)*torch.exp(logvar/2)
return z, mu, logvar
def decode(self, z):
x_recon = self.decoder(z)
return x_recon
class UNetEncoder(nn.Module):
def __init__(self, encoder, input_nc=3):
super().__init__()
if encoder is None:
raise NotImplementedError
self.input_nc = input_nc
self.encoder = encoder
def forward(self, x, t = None, noise = None):
if t is None:
t = torch.ones((x.shape[0]), device=x.device)
if noise is None:
noise = torch.randn_like(x)
z, mu, logvar = self.encode(x, t, noise)
return z, mu, logvar
def encode(self, x, t, noise):
mu_logvar = self.encoder(x, t)
mu, logvar = mu_logvar[:, :self.input_nc], mu_logvar[:, self.input_nc:]
z = mu + noise*torch.exp(logvar/2)
return z, mu, logvar
class FourierMLP(nn.Module):
def __init__(self, input_dim=2, output_dim = 2, num_layers=2, channels=128):
super().__init__()
self.data_shape = [input_dim]
self.output_dim = output_dim
self.register_buffer(
"timestep_coeff", torch.linspace(start=0.1, end=100, steps=channels)[None]
)
self.timestep_phase = nn.Parameter(torch.randn(channels)[None])
self.input_embed = nn.Linear(int(np.prod(input_dim)), channels)
self.timestep_embed = nn.Sequential(
nn.Linear(2 * channels, channels),
nn.GELU(),
nn.Linear(channels, channels),
)
self.layers = nn.Sequential(
nn.GELU(),
*[
nn.Sequential(nn.Linear(channels, channels), nn.GELU())
for _ in range(num_layers)
],
nn.Linear(channels, int(np.prod(output_dim))),
)
def forward(self, inputs, cond):
sin_embed_cond = torch.sin(
(self.timestep_coeff * cond.float()) + self.timestep_phase
)
cos_embed_cond = torch.cos(
(self.timestep_coeff * cond.float()) + self.timestep_phase
)
embed_cond = self.timestep_embed(
rearrange([sin_embed_cond, cos_embed_cond], "d b w -> b (d w)")
)
embed_ins = self.input_embed(inputs.view(inputs.shape[0], -1))
out = self.layers(embed_ins + embed_cond)
return out.view(inputs.shape[0], self.output_dim)