-
Notifications
You must be signed in to change notification settings - Fork 3
/
model_keras.py
243 lines (189 loc) · 8.75 KB
/
model_keras.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
from __future__ import absolute_import
from __future__ import print_function
import numpy as np
import pandas as pd
from keras.models import Sequential
from keras.layers.core import Dense, Dropout, Activation
from keras.layers.normalization import BatchNormalization
from keras.layers.advanced_activations import PReLU
from keras.utils import np_utils, generic_utils
from keras.constraints import nonneg
from keras.regularizers import l1, l1l2
from keras.optimizers import SGD
from sklearn.preprocessing import LabelEncoder
from kappa import quadratic_weighted_kappa
from sklearn.metrics import accuracy_score, confusion_matrix
import os
import constants
from scaler import variance, scale, fit2distribution
import h5py
from utils import dump_prediction
def preprocess_labels(labels, encoder=None, categorical=True):
if not encoder:
encoder = LabelEncoder()
encoder.fit(labels)
y = encoder.transform(labels).astype(np.int32)
if categorical:
y = np_utils.to_categorical(y)
return y, encoder
def pick_activations(model, X, net_type):
input_dim = model.get_config()[0]['input_dim']
output_dim = model.get_config()[0]['output_dim']
dropout = model.get_config()[2]['p']
m = Sequential()
m.add(Dense(input_dim, output_dim, weights=model.layers[0].get_weights()))
m.add(Activation('relu'))
m.add(Dropout(dropout))
if net_type == 'softmax':
loss='categorical_crossentropy'
elif net_type == 'regression':
loss='mean_squared_error'
m.compile(loss=loss, optimizer='adam')
activations = m.predict(X, verbose=0)
return activations
def validate(n_epochs, n_models, n_steps=5, activations=False):
with h5py.File(constants.train_features_scaled_strat_file, "r") as fi:
labels_train = fi.get("y_train")[:60000]
X_train = fi.get("X_train")[:60000]
y_train, _ = preprocess_labels(labels_train, categorical=(net_type=='softmax'))
labels_test = fi.get("y_test")[()]
X_test = fi.get("X_test")[()]
y_test, _ = preprocess_labels(labels_test, categorical=(net_type=='softmax'))
y_train = y_train/5.0/2+0.5
y_test = y_test/5.0/2+0.5
if net_type == 'softmax':
n_classes = y_train.shape[1]
elif net_type == 'regression':
n_classes = 1
print(n_classes, 'classes')
n_dims = X_train.shape[1]
print(n_dims, 'dims')
cum_blend = 0
models = range(1, n_models+1)
for i in models:
print("\n-------------- Model %d --------------\n" % i)
model = model_factory(n_classes, n_dims, net_type)
for n in range(0, n_epochs, n_steps):
model.fit(X_train, y_train, nb_epoch=n_steps, batch_size=128,
show_accuracy=False, verbose=2)#, validation_data=(X_test, y_test))
# validate individual net
if net_type == 'softmax':
y_pred = model.predict_classes(X_test, verbose=0)
elif net_type == 'regression':
y_pred = model.predict(X_test, verbose=0)
y_pred = np.floor((y_pred-0.5)*2*5.0).flatten()
y_pred[y_pred<0] = 0
y_pred[y_pred>4] = 4
print('Epoch: %d. Accuracy: %0.2f%%. Kappa: %0.2f' %
(n+n_steps,
100 * accuracy_score(labels_test, y_pred),
quadratic_weighted_kappa(labels_test, y_pred)))
# validate ensemble
if net_type == 'softmax':
cum_blend += model.predict_proba(X_test, verbose=0)
y_pred = np.argmax(cum_blend, axis=1)
elif net_type == 'regression':
cum_blend += model.predict(X_test, verbose=0)
y_pred = np.floor((cum_blend/i-0.5)*2*5.0).flatten()
y_pred[y_pred<0] = 0
y_pred[y_pred>4] = 4
print('\nBlend %d. Accuracy: %0.2f%%. Kappa: %0.2f' %
(i, 100 * accuracy_score(labels_test, y_pred),
quadratic_weighted_kappa(labels_test, y_pred)))
print('Confusion matrix:\n', confusion_matrix(labels_test, y_pred))
fitted = fit2distribution(labels_test, cum_blend)
print('\nFitted. Accuracy: %0.2f%%. Kappa: %0.2f' %
(100 * accuracy_score(labels_test, fitted),
quadratic_weighted_kappa(labels_test, fitted)))
print('Confusion matrix:\n', confusion_matrix(labels_test, fitted))
if activations:
F_train = pick_activations(model, X_train, net_type)
F_test = pick_activations(model, X_test, net_type)
fout = os.path.join(constants.features_NN_dir,
features_NN_prefix + format(i,'02d') +'.hd5')
with h5py.File(fout, "w") as fo:
fo.create_dataset("X_train", data=F_train)
fo.create_dataset("y_train", data=labels_train)
fo.create_dataset("X_test", data=F_test)
fo.create_dataset("y_test", data=labels_test)
with h5py.File(fout, "r") as fi:
X = fi.get("X_train")
y = fi.get("y_train")
XX = fi.get("X_test")
yy = fi.get("y_test")
print(X.shape, y.shape, XX.shape, yy.shape)
def full_train(n_epochs, n_models, activations=False):
with h5py.File(constants.train_features_scaled_file, "r") as fi:
labels_train = fi.get("y")[()]
X_train = fi.get("X")[()]
y_train, _ = preprocess_labels(labels_train, categorical=(net_type=='softmax'))
y_train = y_train/5.0/2+0.5
if net_type == 'softmax':
n_classes = y_train.shape[1]
elif net_type == 'regression':
n_classes = 1
print(n_classes, 'classes')
n_dims = X_train.shape[1]
print(n_dims, 'dims')
cum_blend = 0
models = range(1, n_models+1)
for i in models:
print("\n-------------- Model %d --------------\n" % i)
model = model_factory(n_classes, n_dims, net_type)
model.fit(X_train, y_train, nb_epoch=n_epochs, batch_size=128,
show_accuracy=False, verbose=2)
# make ensemble prediction
with h5py.File(constants.test_features_scaled_file, "r") as ft:
X_test = ft.get("X")
if net_type == 'softmax':
cum_blend += model.predict_proba(X_test, verbose=0)
elif net_type == 'regression':
cum_blend += model.predict(X_test, verbose=0)
dump_prediction(cum_blend/i, constants.test_packed_list,
constants.pred_dump)
def model_factory(n_classes, n_dims, net_type):
print('Building model. Net type is %s.' % net_type)
layer1_sz = 1024
layer2_sz = 1024
model = Sequential()
model.add(Dense(n_dims, layer1_sz, init='glorot_uniform'))
model.add(Activation('relu'))
# model.add(Dropout(0.5))
model.add(Dense(layer1_sz, layer2_sz, init='glorot_uniform'))
model.add(Activation('relu'))
# model.add(Dropout(0.5))
model.add(Dense(layer2_sz, n_classes, init='glorot_uniform'))
if net_type == 'softmax':
model.add(Activation('softmax'))
elif net_type == 'regression':
pass
if net_type == 'softmax':
loss='categorical_crossentropy'
elif net_type == 'regression':
loss='mean_squared_error'
model.compile(loss=loss, optimizer='adam')
return model
#np.random.seed(1337) # for reproducibility
net_type = 'regression' # softmax|regression
mode = 'validate' # full_train|validate
activations = False
features_NN_prefix = net_type + '.'
# Scale
if not (os.path.exists(constants.train_features_scaled_file) and
os.path.exists(constants.test_features_scaled_file)):
var, mean = variance(constants.train_features_file)
sd = np.sqrt(var)
if not os.path.exists(constants.train_features_scaled_file):
print('Scaling ' + constants.train_features_scaled_file)
scale(constants.train_features_file, constants.train_features_scaled_file, sd, mean)
if not os.path.exists(constants.test_features_scaled_file):
print('Scaling ' + constants.test_features_scaled_file)
scale(constants.test_features_file, constants.test_features_scaled_file, sd, mean)
n_epochs = 200
n_models = 10
if mode == 'validate':
validate(n_epochs, n_models)
elif mode == 'full_train':
full_train(n_epochs, n_models)
else:
print('Unknown mode. Exiting.')