/
net.py
239 lines (207 loc) · 8.68 KB
/
net.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
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
230
231
232
233
234
235
236
237
238
239
'''
Pytorch implementation of ResNet models.
Reference:
[1] He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR, 2016.
'''
from typing import OrderedDict
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.distributions as dist
import os
class BasicBlock(nn.Module):
expansion = 1
def __init__(self, in_planes, planes, stride=1):
super(BasicBlock, self).__init__()
self.conv1 = nn.Conv2d(in_planes, planes, kernel_size=3, stride=stride, padding=1, bias=False)
self.bn1 = nn.BatchNorm2d(planes)
self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, stride=1, padding=1, bias=False)
self.bn2 = nn.BatchNorm2d(planes)
self.shortcut = nn.Sequential()
if stride != 1 or in_planes != self.expansion*planes:
self.shortcut = nn.Sequential(
nn.Conv2d(in_planes, self.expansion*planes, kernel_size=1, stride=stride, bias=False),
nn.BatchNorm2d(self.expansion*planes)
)
def forward(self, x):
out = F.relu(self.bn1(self.conv1(x)))
out = self.bn2(self.conv2(out))
out += self.shortcut(x)
out = F.relu(out)
return out
class Bottleneck(nn.Module):
expansion = 4
def __init__(self, in_planes, planes, stride=1):
super(Bottleneck, self).__init__()
self.conv1 = nn.Conv2d(in_planes, planes, kernel_size=1, bias=False)
self.bn1 = nn.BatchNorm2d(planes)
self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, stride=stride, padding=1, bias=False)
self.bn2 = nn.BatchNorm2d(planes)
self.conv3 = nn.Conv2d(planes, self.expansion*planes, kernel_size=1, bias=False)
self.bn3 = nn.BatchNorm2d(self.expansion*planes)
self.shortcut = nn.Sequential()
if stride != 1 or in_planes != self.expansion*planes:
self.shortcut = nn.Sequential(
nn.Conv2d(in_planes, self.expansion*planes, kernel_size=1, stride=stride, bias=False),
nn.BatchNorm2d(self.expansion*planes)
)
def forward(self, x):
out = F.relu(self.bn1(self.conv1(x)))
out = F.relu(self.bn2(self.conv2(out)))
out = self.bn3(self.conv3(out))
out += self.shortcut(x)
out = F.relu(out)
return out
class ResNet(nn.Module):
def __init__(self, block, num_blocks, num_classes=10, temp=1.0):
super(ResNet, self).__init__()
self.in_planes = 64
self.conv1 = nn.Conv2d(3, 64, kernel_size=3, stride=1, padding=1, bias=False)
self.bn1 = nn.BatchNorm2d(64)
self.layer1 = self._make_layer(block, 64, num_blocks[0], stride=1)
self.layer2 = self._make_layer(block, 128, num_blocks[1], stride=2)
self.layer3 = self._make_layer(block, 256, num_blocks[2], stride=2)
self.layer4 = self._make_layer(block, 512, num_blocks[3], stride=2)
self.fc = nn.Linear(512*block.expansion, num_classes)
self.temp = temp
def _make_layer(self, block, planes, num_blocks, stride):
strides = [stride] + [1]*(num_blocks-1)
layers = []
for stride in strides:
layers.append(block(self.in_planes, planes, stride))
self.in_planes = planes * block.expansion
return nn.Sequential(*layers)
def forward(self, x):
out = F.relu(self.bn1(self.conv1(x)))
out = self.layer1(out)
out = self.layer2(out)
out = self.layer3(out)
out = self.layer4(out)
out = F.avg_pool2d(out, 4)
out = out.view(out.size(0), -1)
out = self.fc(out) / self.temp
return out
BASE_URL = "https://www.robots.ox.ac.uk/~viveka/focal_calibration/CIFAR10/"
def resnet50(temp=1.0, download=False, **kwargs):
model = ResNet(Bottleneck, [3, 4, 6, 3], temp=temp, **kwargs)
if download:
# download weights
di = torch.hub.load_state_dict_from_url(
os.path.join('https://www.robots.ox.ac.uk/~viveka/focal_calibration/CIFAR10/',
'resnet50_cross_entropy_350.model'),
progress=True,
map_location=torch.device('cpu'))
# remove module from keys
new_di = OrderedDict((keys[len("module."):], v) for keys, v in di.items())
model.load_state_dict(new_di)
return model
class Encoder(nn.Module):
def __init__(self, z_dim, hidden_dim=[512, 512], in_dim=28**2):
super().__init__()
self.im_dim = in_dim
layers = []
indim = in_dim
for dim in hidden_dim[:-1]:
layers.append(nn.Linear(indim, dim))
layers.append(nn.ELU())
indim = dim
layers.append(nn.Linear(indim, hidden_dim[-1]))
self.fc1 = nn.Sequential(*layers)
self.fc21 = nn.Linear(hidden_dim[-1], z_dim)
self.fc22 = nn.Linear(hidden_dim[-1], z_dim)
def forward(self, x):
hidden = F.softplus(self.fc1(x))
z_loc = self.fc21(hidden)
z_scale = torch.clamp(F.softplus(self.fc22(hidden)), min=1e-3)
return z_loc, z_scale
class Decoder(nn.Module):
def __init__(self, z_dim, hidden_dim=[512, 512], in_dim=28**2, n_nets=1):
super().__init__()
# setup the two linear transformations used
layers = []
self.n_nets = n_nets
indim = z_dim
for dim in hidden_dim:
layers.append(nn.Linear(indim, dim))
layers.append(nn.ELU())
indim = dim
layers.append(nn.Linear(indim, in_dim))
self.model = nn.Sequential(*layers)
def forward(self, z):
return self.model(z)
class Classifier(nn.Module):
def __init__(self, in_dim, num_classes, hidden_dim=[512, 512]):
super().__init__()
self.in_dim = in_dim
self.num_classes = num_classes
layers = []
indim = in_dim
for dim in hidden_dim:
layers.append(nn.Linear(indim, dim))
layers.append(nn.ELU())
indim = dim
layers.append(nn.Linear(indim, num_classes))
self.net = nn.Sequential(*layers)
def forward(self, x):
return self.net(x)
class CondPrior(nn.Module):
def __init__(self, z_dim, n_classes) -> None:
super().__init__()
self.z_dim = z_dim
self.n_classes = n_classes
self.loc = nn.Linear(n_classes, z_dim, bias=False)
self.scale = nn.Linear(n_classes, z_dim, bias=False)
def forward(self, y):
one_hot = F.one_hot(y, num_classes=self.n_classes).float()
loc = self.loc(one_hot)
scale = F.softplus(self.scale(one_hot)).clamp(min=1e-3)
return loc, scale
def compute_kl(locs_q, scale_q, locs_p=None, scale_p=None):
"""
Computes the KL(q||p)
"""
if locs_p is None:
locs_p = torch.zeros_like(locs_q)
if scale_p is None:
scale_p = torch.ones_like(scale_q)
kl = 0.5 * (2 * scale_p.log() - 2 * scale_q.log() + \
(locs_q - locs_p).pow(2) / scale_p.pow(2) + \
scale_q.pow(2) / scale_p.pow(2) - torch.ones_like(locs_q)).sum(dim=-1)
return kl
def log_likelihood(recon, xs):
return dist.Laplace(recon, torch.ones_like(recon)).log_prob(xs).sum(dim=-1)
class CondVAE(torch.nn.Module):
def __init__(self, z_dim, num_classes, device,
in_dim, n_nets=1):
super(CondVAE, self).__init__()
self.z_dim = z_dim
self.in_dim = in_dim
self.device = device
self.num_classes = num_classes
self.n_nets = n_nets
self.encoder = Encoder(self.z_dim, hidden_dim=[1024, 512, 512], in_dim=self.in_dim)
self.decoder = Decoder(self.z_dim, hidden_dim=[1024, 512, 512], in_dim=self.in_dim)
self.t_pred = Classifier(self.num_classes, 1, hidden_dim=[128, 128])
self.cond_prior = CondPrior(self.z_dim, self.num_classes)
self.to(device)
def elbo(self, x, y):
bs = x.shape[0]
post_params = self.encoder(x)
z = dist.Normal(*post_params).rsample()
kl = compute_kl(*post_params, *self.cond_prior(y))
recon = self.decoder(z)
log_pxz = log_likelihood(recon, x)
loss = - (log_pxz - kl)
return loss.mean()
def t_ce(self, x, pred_logits, y):
t = self.sample_t(x)
return F.cross_entropy(pred_logits / t.view(-1, 1), y)
def sample_t(self, x):
y = torch.linspace(0, self.num_classes-1, self.num_classes).unsqueeze(0).long().to(x.device)
y = y.expand(x.shape[0], -1)
bs = x.shape[0]
post_params = self.encoder(x)
z = dist.Normal(*post_params).rsample().unsqueeze(1).expand(-1, self.num_classes, -1)
p = dist.Normal(*self.cond_prior(y)).log_prob(z).sum(dim=-1)
t_preds = self.t_pred(p)
return F.softplus(t_preds).clamp(min=1e-3)