models.py 10.6 KB
Newer Older
1
2
3
# @author lucasmiranda42

from tensorflow.keras import Input, Model, Sequential
4
from tensorflow.keras.constraints import UnitNorm
5
from tensorflow.keras.initializers import he_uniform, Orthogonal
6
7
from tensorflow.keras.layers import BatchNormalization, Bidirectional, Dense
from tensorflow.keras.layers import Dropout, Lambda, LSTM
8
9
10
from tensorflow.keras.layers import RepeatVector, TimeDistributed
from tensorflow.keras.losses import Huber
from tensorflow.keras.optimizers import Adam
11
from source.model_utils import *
12
13
14
15
import tensorflow as tf


class SEQ_2_SEQ_AE:
16
17
18
    def __init__(
        self,
        input_shape,
lucas_miranda's avatar
lucas_miranda committed
19
20
21
22
23
24
25
        CONV_filters=256,
        LSTM_units_1=256,
        LSTM_units_2=64,
        DENSE_2=64,
        DROPOUT_RATE=0.25,
        ENCODING=32,
        learn_rate=1e-3,
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
    ):
        self.input_shape = input_shape
        self.CONV_filters = CONV_filters
        self.LSTM_units_1 = LSTM_units_1
        self.LSTM_units_2 = LSTM_units_2
        self.DENSE_1 = LSTM_units_2
        self.DENSE_2 = DENSE_2
        self.DROPOUT_RATE = DROPOUT_RATE
        self.ENCODING = ENCODING
        self.learn_rate = learn_rate

    def build(self):
        # Encoder Layers
        Model_E0 = tf.keras.layers.Conv1D(
            filters=self.CONV_filters,
            kernel_size=5,
            strides=1,
            padding="causal",
            activation="relu",
45
            kernel_initializer=he_uniform(),
46
        )
47
        Model_E1 = Bidirectional(
48
            LSTM(
49
50
51
                self.LSTM_units_1,
                activation="tanh",
                return_sequences=True,
52
                kernel_constraint=UnitNorm(axis=0),
53
54
            )
        )
55
        Model_E2 = Bidirectional(
56
            LSTM(
57
58
59
                self.LSTM_units_2,
                activation="tanh",
                return_sequences=False,
60
                kernel_constraint=UnitNorm(axis=0),
61
62
            )
        )
63
        Model_E3 = Dense(
64
65
66
67
            self.DENSE_1,
            activation="relu",
            kernel_constraint=UnitNorm(axis=0),
            kernel_initializer=he_uniform(),
68
69
        )
        Model_E4 = Dense(
70
71
72
73
            self.DENSE_2,
            activation="relu",
            kernel_constraint=UnitNorm(axis=0),
            kernel_initializer=he_uniform(),
74
        )
75
76
77
        Model_E5 = Dense(
            self.ENCODING,
            activation="relu",
78
            kernel_constraint=UnitNorm(axis=1),
79
            activity_regularizer=UncorrelatedFeaturesConstraint(3, weightage=1.0),
80
            kernel_initializer=Orthogonal(),
81
82
83
        )

        # Decoder layers
84
        Model_D0 = DenseTranspose(
85
            Model_E5, activation="relu", output_dim=self.ENCODING,
86
        )
87
88
        Model_D1 = DenseTranspose(Model_E4, activation="relu", output_dim=self.DENSE_2,)
        Model_D2 = DenseTranspose(Model_E3, activation="relu", output_dim=self.DENSE_1,)
89
        Model_D3 = RepeatVector(self.input_shape[1])
90
        Model_D4 = Bidirectional(
91
            LSTM(
92
93
94
                self.LSTM_units_1,
                activation="tanh",
                return_sequences=True,
95
                kernel_constraint=UnitNorm(axis=1),
96
97
            )
        )
98
        Model_D5 = Bidirectional(
99
            LSTM(
100
101
102
                self.LSTM_units_1,
                activation="sigmoid",
                return_sequences=True,
103
                kernel_constraint=UnitNorm(axis=1),
104
105
106
107
            )
        )

        # Define and instanciate encoder
lucas_miranda's avatar
lucas_miranda committed
108
        encoder = Sequential(name="SEQ_2_SEQ_Encoder")
109
        encoder.add(Input(shape=self.input_shape[1:]))
110
        encoder.add(Model_E0)
111
        encoder.add(BatchNormalization())
112
        encoder.add(Model_E1)
113
        encoder.add(BatchNormalization())
114
        encoder.add(Model_E2)
115
        encoder.add(BatchNormalization())
116
        encoder.add(Model_E3)
117
        encoder.add(BatchNormalization())
118
119
        encoder.add(Dropout(self.DROPOUT_RATE))
        encoder.add(Model_E4)
120
        encoder.add(BatchNormalization())
121
122
123
        encoder.add(Model_E5)

        # Define and instanciate decoder
lucas_miranda's avatar
lucas_miranda committed
124
        decoder = Sequential(name="SEQ_2_SEQ_Decoder")
125
        decoder.add(Model_D0)
126
        encoder.add(BatchNormalization())
127
        decoder.add(Model_D1)
128
        encoder.add(BatchNormalization())
129
        decoder.add(Model_D2)
130
        encoder.add(BatchNormalization())
131
        decoder.add(Model_D3)
132
        decoder.add(BatchNormalization())
133
        decoder.add(Model_D4)
134
        encoder.add(BatchNormalization())
135
136
137
        decoder.add(Model_D5)
        decoder.add(TimeDistributed(Dense(self.input_shape[2])))

lucas_miranda's avatar
lucas_miranda committed
138
        model = Sequential([encoder, decoder], name="SEQ_2_SEQ_AE")
139
140
141

        model.compile(
            loss=Huber(reduction="sum", delta=100.0),
142
            optimizer=Adam(lr=self.learn_rate, clipvalue=0.5,),
143
144
145
            metrics=["mae"],
        )

lucas_miranda's avatar
lucas_miranda committed
146
        return encoder, decoder, model
147
148
149


class SEQ_2_SEQ_VAE:
150
151
152
    def __init__(
        self,
        input_shape,
lucas_miranda's avatar
lucas_miranda committed
153
154
155
156
157
158
159
        CONV_filters=256,
        LSTM_units_1=256,
        LSTM_units_2=64,
        DENSE_2=64,
        DROPOUT_RATE=0.25,
        ENCODING=32,
        learn_rate=1e-3,
160
        loss="ELBO+MMD",
161
162
163
164
165
166
167
168
169
170
    ):
        self.input_shape = input_shape
        self.CONV_filters = CONV_filters
        self.LSTM_units_1 = LSTM_units_1
        self.LSTM_units_2 = LSTM_units_2
        self.DENSE_1 = LSTM_units_2
        self.DENSE_2 = DENSE_2
        self.DROPOUT_RATE = DROPOUT_RATE
        self.ENCODING = ENCODING
        self.learn_rate = learn_rate
171
        self.loss = loss
172
173
174
175

    def build(self):
        # Encoder Layers
        Model_E0 = tf.keras.layers.Conv1D(
176
            filters=self.CONV_filters,
177
178
179
180
            kernel_size=5,
            strides=1,
            padding="causal",
            activation="relu",
181
            kernel_initializer=he_uniform(),
182
        )
183
        Model_E1 = Bidirectional(
184
            LSTM(
185
                self.LSTM_units_1,
186
187
                activation="tanh",
                return_sequences=True,
188
                kernel_constraint=UnitNorm(axis=0),
189
190
            )
        )
191
        Model_E2 = Bidirectional(
192
            LSTM(
193
                self.LSTM_units_2,
194
195
                activation="tanh",
                return_sequences=False,
196
                kernel_constraint=UnitNorm(axis=0),
197
198
            )
        )
199
        Model_E3 = Dense(
200
201
202
203
            self.DENSE_1,
            activation="relu",
            kernel_constraint=UnitNorm(axis=0),
            kernel_initializer=he_uniform(),
204
205
        )
        Model_E4 = Dense(
206
207
208
209
            self.DENSE_2,
            activation="relu",
            kernel_constraint=UnitNorm(axis=0),
            kernel_initializer=he_uniform(),
210
        )
211
        Model_E5 = Dense(
212
            self.ENCODING,
213
            activation="relu",
214
            kernel_constraint=UnitNorm(axis=1),
215
            activity_regularizer=UncorrelatedFeaturesConstraint(3, weightage=1.0),
216
            kernel_initializer=Orthogonal(),
217
218
219
        )

        # Decoder layers
lucas_miranda's avatar
lucas_miranda committed
220

221
        Model_D0 = DenseTranspose(
222
            Model_E5, activation="relu", output_dim=self.ENCODING,
223
        )
224
225
        Model_D1 = DenseTranspose(Model_E4, activation="relu", output_dim=self.DENSE_2,)
        Model_D2 = DenseTranspose(Model_E3, activation="relu", output_dim=self.DENSE_1,)
lucas_miranda's avatar
lucas_miranda committed
226
        Model_D3 = RepeatVector(self.input_shape[1])
227
        Model_D4 = Bidirectional(
228
            LSTM(
229
                self.LSTM_units_1,
230
231
                activation="tanh",
                return_sequences=True,
232
                kernel_constraint=UnitNorm(axis=1),
233
234
            )
        )
235
        Model_D5 = Bidirectional(
236
            LSTM(
237
                self.LSTM_units_1,
238
239
                activation="sigmoid",
                return_sequences=True,
240
                kernel_constraint=UnitNorm(axis=1),
241
242
243
244
245
            )
        )

        # Define and instanciate encoder
        x = Input(shape=self.input_shape[1:])
246
        encoder = Model_E0(x)
247
        encoder = BatchNormalization()(encoder)
248
        encoder = Model_E1(encoder)
249
        encoder = BatchNormalization()(encoder)
250
        encoder = Model_E2(encoder)
251
        encoder = BatchNormalization()(encoder)
252
        encoder = Model_E3(encoder)
253
        encoder = BatchNormalization()(encoder)
254
        encoder = Dropout(self.DROPOUT_RATE)(encoder)
255
        encoder = Model_E4(encoder)
256
        encoder = BatchNormalization()(encoder)
257
258
        encoder = Model_E5(encoder)

259
260
        z_mean = Dense(self.ENCODING)(encoder)
        z_log_sigma = Dense(self.ENCODING)(encoder)
261
262
263
264
265
266
267
268
269

        if "ELBO" in self.loss:
            z_mean, z_log_sigma = KLDivergenceLayer()([z_mean, z_log_sigma])

        z = Lambda(sampling)([z_mean, z_log_sigma])

        if "MMD" in self.loss:
            z = MMDiscrepancyLayer()(z)

lucas_miranda's avatar
lucas_miranda committed
270
271
        # Define and instanciate generator
        generator = Model_D0(z)
272
        generator = BatchNormalization()(generator)
lucas_miranda's avatar
lucas_miranda committed
273
        generator = Model_D1(generator)
274
        generator = BatchNormalization()(generator)
lucas_miranda's avatar
lucas_miranda committed
275
        generator = Model_D2(generator)
276
        generator = BatchNormalization()(generator)
lucas_miranda's avatar
lucas_miranda committed
277
        generator = Model_D3(generator)
278
        generator = BatchNormalization()(generator)
lucas_miranda's avatar
lucas_miranda committed
279
        generator = Model_D4(generator)
280
        generator = BatchNormalization()(generator)
lucas_miranda's avatar
lucas_miranda committed
281
282
        generator = Model_D5(generator)
        x_decoded_mean = TimeDistributed(Dense(self.input_shape[2]))(generator)
283

284
        # end-to-end autoencoder
lucas_miranda's avatar
lucas_miranda committed
285
        encoder = Model(x, z_mean, name="SEQ_2_SEQ_VEncoder")
286
        vae = Model(x, x_decoded_mean, name="SEQ_2_SEQ_VAE")
lucas_miranda's avatar
lucas_miranda committed
287

288
289
290
        # Build generator as a separate entity
        g = Input(shape=self.ENCODING)
        _generator = Model_D0(g)
291
        _generator = BatchNormalization()(_generator)
292
        _generator = Model_D1(_generator)
293
        _generator = BatchNormalization()(_generator)
294
        _generator = Model_D2(_generator)
295
        _generator = BatchNormalization()(_generator)
296
        _generator = Model_D3(_generator)
297
        _generator = BatchNormalization()(_generator)
298
        _generator = Model_D4(_generator)
299
        _generator = BatchNormalization()(_generator)
300
301
302
        _generator = Model_D5(_generator)
        _x_decoded_mean = TimeDistributed(Dense(self.input_shape[2]))(_generator)
        generator = Model(g, _x_decoded_mean, name="SEQ_2_SEQ_VGenerator")
303

304
305
306
        def huber_loss(x_, x_decoded_mean_):
            huber = Huber(reduction="sum", delta=100.0)
            return self.input_shape[1:] * huber(x_, x_decoded_mean_)
307
308
309

        vae.compile(
            loss=huber_loss,
lucas_miranda's avatar
lucas_miranda committed
310
            optimizer=Adam(lr=self.learn_rate,),
311
312
313
314
            metrics=["mae"],
            experimental_run_tf_function=False,
        )

315
        return encoder, generator, vae
316
317


318
class SEQ_2_SEQ_VAME:
319
320
321
    pass


322
class SEQ_2_SEQ_MMVAE:
323
    pass
lucas_miranda's avatar
lucas_miranda committed
324

325

326
# TODO next:
lucas_miranda's avatar
lucas_miranda committed
327
#      - VAE loss function (though this should be analysed later on taking the encodings into account)
328
#      - Smaller input sliding window (10-15 frames)