-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathlayer.py
240 lines (165 loc) · 8.12 KB
/
layer.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
from __future__ import division
from itertools import combinations
from keras import regularizers
from keras.layers import Dense, Dropout, Lambda, Input, concatenate
from keras import backend as K
def add_dropout(layer, params, name='dropout'):
return Dropout(params['dropout'], name=name)(layer)
def add_fc(layer, params, n_units=False, regularization=False,
name='main-hidden'):
if n_units is False:
n_units = params['fc_hidden_u']
kernel_reg, activity_reg = get_regularization(regularization)
return Dense(n_units, activation=params['activation'], name=name,
kernel_regularizer=kernel_reg,
activity_regularizer=activity_reg)(layer)
def add_softmax(layer, params, name='main-output'):
return Dense(params['n_classes'], activation='softmax', name=name)(layer)
def baseline(params):
input_layer = Input(shape=(params['n_features'],), name='inputs')
if params['feature_set'] == 'nocontext':
# Filter out context (course info + peers columns).
base_layer = NoContext(input_layer, params)
elif params['feature_set'] == 'withpeers':
# Filter out course info.
base_layer = NoCourseInfo(input_layer, params)
elif params['feature_set'] == 'all':
base_layer = input_layer
return input_layer, base_layer
def subselect_cols(x, cols):
# For multiprocessing.
import tensorflow as tf
tf.logging.set_verbosity(tf.logging.ERROR)
# Return the selected columns from the set of input features.
return tf.gather(x, cols, axis=1)
def get_regularization(reg):
if reg is False:
kernel_reg = None
activity_reg = None
else:
kernel_reg = regularizers.l1(0.01)
activity_reg = None
return kernel_reg, activity_reg
def Combinations(inputs, params, reg=False):
"""Input features C(n, r) combinations."""
kernel_reg, activity_reg = get_regularization(reg)
# Generate comb(n_features, r) combinations.
combs = combinations(range(params['n_features']), 2)
layers = []
for i, combination in enumerate(combs):
# This combination features as input.
cols = K.constant(combination, dtype='int32')
layer_name = 'combination-' + str(i)
combination_input = Lambda(subselect_cols, arguments={
'cols': cols}, name=layer_name)(inputs)
# 1 single output unit.
layers.append(Dense(1, activation=params['activation'],
name='W-' + layer_name,
kernel_regularizer=kernel_reg,
activity_regularizer=activity_reg)
(combination_input))
return concatenate(layers, axis=1)
def NoContext(inputs, params):
"""Filter out all context features (course info and course peers)."""
cols = K.constant(params['cols']['activity'], dtype='int32')
layer_name = 'student-activity'
return Lambda(subselect_cols,
arguments={'cols': cols}, name=layer_name)(inputs)
def NoCourseInfo(inputs, params):
"""Filter out all context features."""
col_indexes = (params['cols']['activity'] + params['cols']['peers'])
cols = K.constant(col_indexes, dtype='int32')
layer_name = 'student-activity-and-peers'
return Lambda(subselect_cols,
arguments={'cols': cols}, name=layer_name)(inputs)
def ContextualiseActivity(inputs, params, reg=True):
"""Each no-context feature combined with a list of context features."""
kernel_reg, activity_reg = get_regularization(reg)
layers = []
for col_index in params['cols']['activity']:
# This no-context feature + all context features as input.
cols = K.constant([col_index] + params['cols']['ctx'], dtype='int32')
layer_name = 'in-context-' + str(col_index)
contextualised_input = Lambda(subselect_cols, arguments={
'cols': cols}, name=layer_name)(inputs)
# 'multiplier' units as num outputs.
layers.append(Dense(1, activation=params['activation'],
name='W-' + layer_name,
kernel_regularizer=kernel_reg,
activity_regularizer=activity_reg)
(contextualised_input))
return concatenate(layers, axis=1)
def ContextualiseActivityAndOriginalActivity(inputs, params, reg=False):
"""Each no-context feature combined with a list of context features + inputs.
"""
kernel_reg, activity_reg = get_regularization(reg)
# Inputs as they are.
layers = [inputs]
for col_index in params['cols']['activity']:
# This no-context feature + all context features as input.
cols = K.constant([col_index] + params['cols']['ctx'], dtype='int32')
layer_name = 'in-context-' + str(col_index)
contextualised_input = Lambda(subselect_cols, arguments={
'cols': cols}, name=layer_name)(inputs)
# 'multiplier' units as num outputs.
layers.append(Dense(1, activation=params['activation'],
name='W-' + layer_name,
kernel_regularizer=kernel_reg,
activity_regularizer=activity_reg)
(contextualised_input))
# Add all original no-context inputs.
activity_cols = K.constant(params['cols']['activity'], dtype='int32')
layer_name = 'no-context'
layers.append(Lambda(subselect_cols, arguments={
'cols': activity_cols}, name=layer_name)(inputs))
return concatenate(layers, axis=1)
def SplitActivityAndContext(inputs, params, n_ctx_units=False, reg=False,
context_includes_peers=True):
"""Split input features in context and no-context."""
layers = []
if context_includes_peers is True:
context_cols = params['cols']['ctx']
no_context_cols = params['cols']['activity']
ctx_layer_activation = params['activation']
else:
context_cols = params['cols']['courseinfo']
no_context_cols = params['cols']['activity'] + params['cols']['peers']
ctx_layer_activation = 'relu'
# Context features learn separately.
cols = K.constant(context_cols, dtype='int32')
layer_name = 'context'
ctx_input = Lambda(subselect_cols, arguments={
'cols': cols}, name=layer_name)(inputs)
if n_ctx_units is False:
# By default the number of units equal to the number of context cols.
n_ctx_units = len(params['cols']['ctx'])
kernel_reg, activity_reg = get_regularization(reg)
ctx_layer = Dense(n_ctx_units, activation=ctx_layer_activation,
name='W-' + layer_name,
kernel_regularizer=kernel_reg,
activity_regularizer=activity_reg)(ctx_input)
layers.append(ctx_layer)
# No context features learn separately.
cols = K.constant(no_context_cols, dtype='int32')
layer_name = 'no-context'
no_ctx_input = Lambda(subselect_cols, arguments={
'cols': cols}, name=layer_name)(inputs)
# Number of units equal to the number of no context features.
layers.append(Dense(len(no_context_cols),
activation=params['activation'],
name='W-' + layer_name)(no_ctx_input))
return concatenate(layers, axis=1)
def SplitAllInputs(inputs, params):
"""Multiple inputs separated based on a hardcoded set of cols."""
layers = []
for layer_name, layer_data in params['separate_cols'].items():
cols = K.constant(layer_data['cols'], dtype='int32')
layer_input = Lambda(subselect_cols, arguments={
'cols': cols}, name=layer_name)(inputs)
# Number of units equal to the number of columns.
layers.append(Dense(layer_data['units'],
activation=layer_data['activation'],
kernel_regularizer=layer_data['kernelreg'],
activity_regularizer=layer_data['kernelreg'],
name='W-' + layer_name)(layer_input))
return concatenate(layers, axis=1)