-
Notifications
You must be signed in to change notification settings - Fork 4
/
FIDO.py
252 lines (195 loc) · 10.4 KB
/
FIDO.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
240
241
242
243
244
245
246
247
248
249
250
251
252
from __future__ import division, print_function
import torch
import torch.nn as nn
import numpy as np
from src.utils import BaseNet, MNIST_mean_std_norm
from src.probability import decompose_entropy_cat, decompose_std_gauss
from torch.optim.lr_scheduler import StepLR
from torch.optim import Adam
from src.utils import to_variable
# TODO: Might not be dividing loss by the number of samples, effectively adding unecesary variance
def gumbel_sigmoid(prob_map, temperature, eps=1e-20):
U = prob_map.new(prob_map.shape).uniform_(0, 1)
sigmoid_in = torch.log(prob_map + eps) - torch.log(1 - prob_map + eps) + torch.log(U + eps) - torch.log(1 - U + eps)
y = torch.sigmoid(sigmoid_in / temperature)
y_hard = torch.round(y)
return (y_hard - y).detach() + y
class bern_mask(nn.Module):
def __init__(self, shape, init_p=0.5, temp=0.1):
super(bern_mask, self).__init__()
self.mask_probs = nn.Parameter(torch.ones(shape) * init_p)
self.temp = temp
def forward(self, x):
hard_mask = gumbel_sigmoid(self.mask_probs, self.temp)
return x * hard_mask, (1 - hard_mask)
class mask_explainer(BaseNet):
def __init__(self, shape, mask_L1_weight, aleatoric_coeff, epistemic_coeff,
mask_samples=1, lr=0.05, decay_period=5, gamma=0.8, cuda=True):
super(mask_explainer, self).__init__()
self.mask_L1_weight = mask_L1_weight
self.aleatoric_coeff = aleatoric_coeff
self.epistemic_coeff = epistemic_coeff
self.mask_samples = mask_samples
self.model = bern_mask(shape, init_p=0.5, temp=0.1)
self.optimizer = Adam(self.model.parameters(), lr=lr)
self.scheduler = StepLR(self.optimizer, step_size=decay_period, gamma=gamma)
self.cuda = cuda
if self.cuda:
self.model = self.model.cuda()
def fit_cat(self, x, BNN, VAEAC, flatten_ims=True, test_dims=None, plot=False):
# note that x will need to be the same shape as specified at class initialisation
x, = to_variable(var=(x,), cuda=self.cuda)
self.set_mode_train(train=True)
BNN.set_mode_train(train=False)
VAEAC.set_mode_train(train=False)
self.optimizer.zero_grad()
loss_cum = 0
aleatoric_cum = 0
epistemic_cum = 0
for it in range(self.mask_samples):
masked_x, mask = self.model(x)
if flatten_ims:
flat_x = x.view(masked_x.shape[0], -1)
masked_x = masked_x.view(masked_x.shape[0], -1)
mask = mask.view(mask.shape[0], -1)
else:
flat_x = x
if test_dims is not None:
# x = torch.cat([x, x.new_zeros(x.shape[0], test_dims)], dim=1)
masked_x = torch.cat([masked_x, masked_x.new_zeros(masked_x.shape[0], test_dims)], dim=1)
mask = torch.cat([mask, mask.new_ones(mask.shape[0], test_dims)], dim=1)
# We dont want gradients from this
inpainted = VAEAC.inpaint(masked_x.data, mask.data, Nsample=1, z_mean=True).data.squeeze(0)
if test_dims is not None:
inpainted = inpainted[:, :-test_dims]
mask = mask[:, :-test_dims]
to_BNN = inpainted * mask + flat_x * (1 - mask)
to_BNN = MNIST_mean_std_norm(to_BNN)
probs = BNN.sample_predict(to_BNN, Nsamples=0, grad=True)
total_entropy, aleatoric_entropy, epistemic_entropy = decompose_entropy_cat(probs)
# We mean across batch
aleatoric_cum += aleatoric_entropy.mean().item()
epistemic_cum += epistemic_entropy.mean().item()
# we should average over MC samples but sum over batch and features
loss = (self.aleatoric_coeff * aleatoric_entropy +
self.epistemic_coeff * epistemic_entropy + self.mask_L1_weight * mask.sum(dim=1)).sum(dim=0) / self.mask_samples
loss.backward() # Gradient accumulation
loss_cum += loss.item() / aleatoric_entropy.shape[0]
# we average here to be invariant to number of samples
aleatoric_cum = aleatoric_cum / self.mask_samples
epistemic_cum = epistemic_cum / self.mask_samples
self.optimizer.step()
self.model.mask_probs.data = torch.clamp(self.model.mask_probs, min=0, max=1)
self.scheduler.step()
return loss_cum, aleatoric_cum, epistemic_cum
def fit_gauss(self, x, BNN, VAEAC, flatten_ims=True, test_dims=None, plot=False):
x, = to_variable(var=(x,), cuda=self.cuda)
self.set_mode_train(train=True)
BNN.set_mode_train(train=False)
VAEAC.set_mode_train(train=False)
self.optimizer.zero_grad()
loss_cum = 0
aleatoric_cum = 0
epistemic_cum = 0
for it in range(self.mask_samples):
masked_x, mask = self.model(x)
if flatten_ims:
flat_x = x.view(masked_x.shape[0], -1)
masked_x = masked_x.view(masked_x.shape[0], -1)
mask = mask.view(mask.shape[0], -1)
else:
flat_x = x
if test_dims is not None:
# x = torch.cat([x, x.new_zeros(x.shape[0], test_dims)], dim=1)
masked_x = torch.cat([masked_x, masked_x.new_zeros(masked_x.shape[0], test_dims)], dim=1)
mask = torch.cat([mask, mask.new_ones(mask.shape[0], test_dims)], dim=1)
# We dont want gradients from this
# Switched to non gauss output
inpainted = VAEAC.inpaint(masked_x.data, mask.data, Nsample=1, z_mean=True).data.squeeze(0)
if test_dims is not None:
inpainted = inpainted[:, :-test_dims]
mask = mask[:, :-test_dims]
to_BNN = inpainted * mask + flat_x * (1 - mask)
mu, std = BNN.sample_predict(to_BNN, Nsamples=0, grad=True)
total_std, aleatoric_std, epistemic_std = decompose_std_gauss(mu, std)
# we average here to be invariant to batch size
aleatoric_cum += aleatoric_std.mean().item()
epistemic_cum += epistemic_std.mean().item()
loss = (self.aleatoric_coeff * aleatoric_std +
self.epistemic_coeff * epistemic_std + self.mask_L1_weight * mask.sum(dim=1)).sum(dim=0) / self.mask_samples
loss.backward() # Gradient accumulation
loss_cum += loss.item() / aleatoric_std.shape[0]
# we average here to be invariant to number of samples
aleatoric_cum = aleatoric_cum / self.mask_samples
epistemic_cum = epistemic_cum / self.mask_samples
# loss_cum.backward()
self.optimizer.step()
self.model.mask_probs.data = torch.clamp(self.model.mask_probs, min=0, max=1)
self.scheduler.step()
return loss_cum, aleatoric_cum, epistemic_cum
def get_mask(self):
self.set_mode_train(train=False)
"""Note that this returns 1s for input features which are masked"""
return 1 - self.model.mask_probs.data.round()
def get_mask_probs(self):
self.set_mode_train(train=False)
"""Note that this returns 1s for input features which are masked"""
return 1 - self.model.mask_probs.data
def mask_input(self, x):
self.set_mode_train(train=False)
x, = to_variable(var=(x,), cuda=self.cuda)
self.set_mode_train(train=False)
return x * self.model.mask_probs.data.round()
def mask_inpaint(self, x, VAEAC, flatten_ims=True, test_dims=None, cat=False):
x, = to_variable(var=(x,), cuda=self.cuda)
self.set_mode_train(train=False)
VAEAC.set_mode_train(train=False)
masked_x = x * self.model.mask_probs.data.round().data
mask = 1 - self.model.mask_probs.data.round().data
if flatten_ims:
flat_x = x.view(masked_x.shape[0], -1)
masked_x = masked_x.view(masked_x.shape[0], -1)
mask = mask.view(mask.shape[0], -1)
else:
flat_x = x
if test_dims is not None:
masked_x = torch.cat([masked_x, masked_x.new_zeros(masked_x.shape[0], test_dims)], dim=1)
mask = torch.cat([mask, mask.new_ones(mask.shape[0], test_dims)], dim=1)
# We dont want gradients from this
if cat:
inpainted = VAEAC.inpaint(masked_x.data, mask.data, Nsample=1, z_mean=True).data.squeeze(0)
else:
inpainted = VAEAC.inpaint(masked_x.data, mask.data, Nsample=1, z_mean=True).data.squeeze(0)
if test_dims is not None:
inpainted = inpainted[:, :-test_dims]
mask = mask[:, :-test_dims]
out = inpainted * mask + flat_x * (1 - mask)
return out.data, mask.data
@staticmethod
def train_mask(x, BNN, VAEAC, aleatoric_coeff, epistemic_coeff, L1w=1, N_epochs=30,
mask_samples=20, mask_samples2=10, cat=True, flatten_ims=True, test_dims=None):
torch.cuda.empty_cache()
x_pixels = x.view(x.shape[0], -1).shape[1]
explainer = mask_explainer(shape=x.shape, mask_L1_weight=L1w/x_pixels, aleatoric_coeff=aleatoric_coeff,
epistemic_coeff=epistemic_coeff, mask_samples=mask_samples, lr=0.05, decay_period=5,
gamma=0.8, cuda=True)
loss_vec = []
aleatoric_vec = []
epistemic_vec = []
for i in range(N_epochs):
if i > 10: # We train with more samples at the beginning as there is a lot more variability
explainer.mask_samples = mask_samples2
if cat:
loss, aleatoric_ent, epistemic_ent = explainer.fit_cat(x, BNN, VAEAC, flatten_ims=flatten_ims,
test_dims=test_dims, plot=False)
else:
loss, aleatoric_ent, epistemic_ent = explainer.fit_gauss(x, BNN, VAEAC, flatten_ims=flatten_ims,
test_dims=test_dims, plot=False)
loss_vec.append(loss)
aleatoric_vec.append(aleatoric_ent)
epistemic_vec.append(epistemic_ent)
print('it: %d, loss: %3.3f, aleatoric: %3.3f, epistemic: %3.3f' % (i, loss, aleatoric_ent, epistemic_ent))
loss_vec = np.array(loss_vec)
aleatoric_vec = np.array(aleatoric_vec)
epistemic_vec = np.array(epistemic_vec)
return explainer, loss_vec, aleatoric_vec, epistemic_vec