return_sequences=False,
dropout=0.25,
recurrent_dropout=0.25,
activation='tanh')
else:
if use_TT == 0:
rnn_layer = LSTM(output_dim=np.array(tt_output_shape).prod(),
return_sequences=False,
dropout=0.25,
recurrent_dropout=0.25,
activation='tanh')
else:
rnn_layer = TT_LSTM(tt_input_shape=tt_input_shape,
tt_output_shape=tt_output_shape,
tt_ranks=tt_ranks,
return_sequences=False,
dropout=0.25,
recurrent_dropout=0.25,
activation='tanh')
h = rnn_layer(masked_input)
output = Dense(output_dim=11,
activation='softmax',
kernel_regularizer=l2(alpha))(h)
model = Model(input, output)
model.compile(optimizer=Adam(1e-4),
loss='categorical_crossentropy',
metrics=['accuracy'])
# Start training -------------------------------------------------------------------------------------------------------
start = datetime.datetime.now()
for l in range(1001):
if use_TT:
# TT settings
tt_input_shape = [7, 7, 16]
tt_output_shape = [4, 4, 4]
tt_ranks = [1, 4, 4, 1]
rnn_size = 64
X = Input(shape=X_tr.shape[1::])
X_mask = Masking(mask_value=0.0, input_shape=X_tr.shape[1::])(X)
if use_TT:
Z = TT_LSTM(tt_input_shape=tt_input_shape,
tt_output_shape=tt_output_shape,
tt_ranks=tt_ranks,
return_sequences=False,
recurrent_dropout=.5)(X_mask)
Out = Dense(units=1, activation='sigmoid', kernel_regularizer=l2(1e-2))(Z)
else:
Z = LSTM(units=rnn_size, return_sequences=False,
recurrent_dropout=.5)(X_mask) # dropout=.5,
Out = Dense(units=1, activation='sigmoid', kernel_regularizer=l2(1e-2))(Z)
rnn_model = Model(X, Out)
rnn_model.compile(optimizer=Adam(1e-3),
loss='binary_crossentropy',
metrics=['accuracy'])
# Train the model and save the results ######################################################
rnn_model.fit(X_tr,