/
adversarial-impute.py
238 lines (214 loc) · 8.3 KB
/
adversarial-impute.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
from keras.models import Sequential, Graph
from keras.objectives import mse
from keras.layers.core import Dropout, Dense
import numpy as np
import theano
import seaborn
def create_models(
n_dims=25,
dropout_probability=0.1,
adversarial_weight=1,
reconstruction_weight=1,
generator_hidden_size_multipliers=[4, 0.5],
decoder_hidden_size_multiplier=[4, 2, 1, 0.5],
activation="relu"):
decoder = Graph()
decoder.add_input(name="input", input_shape=(n_dims,))
last_layer = "input"
for i, size in enumerate(decoder_hidden_size_multiplier):
hidden_name = "dense%d" % (i + 1)
decoder.add_node(
layer=Dense(size * n_dims, activation=activation),
name=hidden_name,
inputs=[last_layer, "input"])
dropout_name = hidden_name + "_dropout"
decoder.add_node(
layer=Dropout(dropout_probability),
name=dropout_name,
input=hidden_name)
last_layer = dropout_name
decoder.add_node(
layer=Dense(n_dims, activation="sigmoid"),
name="output",
inputs=[last_layer, "input"],
create_output=True)
decoder.compile(optimizer="rmsprop", loss={"output": "binary_crossentropy"})
def generator_loss(combined, imputed_vector):
"""
Ignores y_true and y_pred
"""
original_vector = combined[:, :n_dims]
missing_mask = combined[:, n_dims:]
input_variable = decoder.get_input()
decoder_compute_graph = decoder.get_output()
mask_prediction = theano.clone(
decoder_compute_graph,
{input_variable: imputed_vector},
share_inputs=True)
reconstruction_loss = mse(
y_true=original_vector * (1 - missing_mask),
y_pred=imputed_vector * (1 - missing_mask))
decoder_mask_loss = mse(missing_mask, missing_mask * mask_prediction)
return (
reconstruction_weight * reconstruction_loss
- adversarial_weight * decoder_mask_loss)
generator = Sequential()
generator.add(Dense(
(generator_hidden_size_multipliers[0] * n_dims),
input_dim=2 * n_dims,
activation=activation))
generator.add(Dropout(dropout_probability))
for layer_size_multiplier in generator_hidden_size_multipliers[1:] + [1]:
generator.add(Dense(
int(layer_size_multiplier * n_dims),
activation=activation))
generator.add(Dropout(dropout_probability))
generator.add(Dense(n_dims, activation='linear'))
generator.compile(optimizer="rmsprop", loss=generator_loss)
return generator, decoder
def create_data(
n_samples=1000,
n_dims=25,
offset=0,
fraction_missing=0.5):
t = np.linspace(-1, 1, n_dims, endpoint=False)
assert len(t) == n_dims
X_full = np.zeros((n_samples, n_dims))
for i in range(n_samples):
phase = np.random.randn() * np.pi / 2
frequency = 10 * np.random.rand()
X_full[i, :] = offset + np.sin(phase + t * frequency)
missing_mask = np.random.random(X_full.shape) < fraction_missing
X_incomplete = X_full.copy()
X_incomplete[missing_mask] = 0.0
return X_full, X_incomplete, missing_mask
def pretrain_decoder(
X_incomplete,
missing_mask,
decoder,
batch_size=128,
training_epochs=5):
X_incomplete = X_incomplete.copy()
n_samples, n_dims = X_incomplete.shape
X_incomplete[missing_mask] = np.nan
feature_means = np.nanmean(X_incomplete, axis=0)
assert len(feature_means) == n_dims
feature_stds = np.nanstd(X_incomplete, axis=0)
assert len(feature_stds) == n_dims
X_incomplete[missing_mask] = 0.0
indices = np.arange(n_samples)
for epoch in range(training_epochs):
np.random.shuffle(indices)
n_batches = n_samples // batch_size
for batch_idx in range(n_batches):
batch_indices = indices[
batch_idx * batch_size:(batch_idx + 1) * batch_size]
X_batch = X_incomplete[batch_indices]
missing_mask_batch = missing_mask[batch_indices]
X_batch_noisy = X_batch.copy()
for feature_idx in range(n_dims):
noise = np.random.randn(batch_size) * feature_stds[feature_idx] + feature_means[feature_idx]
missing_rows = missing_mask_batch[:, feature_idx]
X_batch_noisy[missing_rows, feature_idx] = noise[missing_rows]
predicted_mask = decoder.predict({"input": X_batch_noisy})["output"]
print("Pre-training epoch %d, mini-batch %d, accuracy=%0.4f" % (
epoch + 1,
batch_idx + 1,
((predicted_mask > 0.5) == missing_mask_batch).mean(),))
decoder.train_on_batch({
"input": X_batch_noisy,
"output": missing_mask_batch
})
def train(
X_full,
X_incomplete,
mask,
generator,
decoder,
batch_size=128,
training_epochs=10,
alternating_updates=False,
plot_each_epoch=False,
decoder_pretrain_epochs=5):
combined = np.hstack([X_incomplete, mask])
n_samples = len(X_full)
indices = np.arange(n_samples)
for epoch in range(training_epochs):
np.random.shuffle(indices)
n_batches = n_samples // batch_size
for batch_idx in range(n_batches):
batch_indices = indices[
batch_idx * batch_size:(batch_idx + 1) * batch_size]
combined_batch = combined[batch_indices]
mask_batch = mask[batch_indices]
X_imputed = generator.predict(combined_batch)
X_full_batch = X_full[batch_indices]
reconstruction_mse = ((X_imputed - X_full_batch) ** 2).mean()
predicted_mask = decoder.predict({"input": X_imputed})["output"]
masking_mse = ((mask_batch - predicted_mask) ** 2).mean()
if np.isnan(reconstruction_mse):
raise ValueError("Generator Diverged!")
if np.isnan(masking_mse):
raise ValueError("Decoder Diverged!")
print((
"-- Epoch %d, batch %d, "
"Reconstruction MSE = %0.4f, "
"Decoder MSE = %0.4f, "
"Decoder accuracy = %0.4f "
"(mean mask prediction = %0.4f)") % (
epoch + 1,
batch_idx + 1,
reconstruction_mse,
masking_mse,
((predicted_mask > 0.5) == mask_batch).mean(),
predicted_mask.mean()))
print("Decoder mask predictions: %s" % (
list(zip(predicted_mask[0], mask_batch[0])),))
if not alternating_updates or batch_idx % 2 == 0:
decoder_input = X_imputed.copy()
decoder_input[~mask_batch] = X_full_batch[~mask_batch]
decoder.train_on_batch({
"input": decoder_input,
"output": mask_batch
})
if not alternating_updates or batch_idx % 2 == 1:
generator.train_on_batch(
X=combined_batch,
y=combined_batch)
if plot_each_epoch or epoch == training_epochs - 1:
seaborn.plt.plot(X_imputed[0, :], label="X_imputed")
seaborn.plt.plot(X_full_batch[0, :n_dims], label="X_full")
seaborn.plt.plot(mask_batch[0], label="mask")
seaborn.plt.legend()
seaborn.plt.show()
if __name__ == "__main__":
n_dims = 50
n_samples = 10 ** 5
training_epochs = 10
batch_size = 128
pretrain = False
X_full, X_incomplete, missing_mask = create_data(
n_dims=n_dims,
n_samples=n_samples,
fraction_missing=0.9)
generator, decoder = create_models(
n_dims=n_dims,
activation="relu",
adversarial_weight=0.1)
if pretrain:
pretrain_decoder(
X_incomplete=X_incomplete,
missing_mask=missing_mask,
decoder=decoder,
training_epochs=training_epochs,
batch_size=batch_size)
train(
X_full,
X_incomplete,
missing_mask,
generator=generator,
decoder=decoder,
training_epochs=10,
alternating_updates=False,
plot_each_epoch=False,
batch_size=batch_size)