class LSTMRec(Model):
def __init__(self, vocabulary_size, emb_output_dim, lstm_units,
lstm_activation, lstm_recurrent_activation, lstm_dropout,
lstm_recurrent_dropout, dense_activation):
super(LSTMRec, self).__init__()
self.emb = Embedding(input_dim=vocabulary_size + 1,
output_dim=emb_output_dim,
mask_zero=True)
self.lstm = LSTM(units=lstm_units,
activation=lstm_activation,
recurrent_activation=lstm_recurrent_activation,
dropout=lstm_dropout,
recurrent_dropout=lstm_recurrent_dropout)
self.dense = Dense(units=vocabulary_size, activation=dense_activation)
# https://stackoverflow.com/questions/39854390/nan-in-summary-histogram
# https://stackoverflow.com/questions/53080289/keras-nan-in-summary-histogram-lstm
# self.emb.add_loss(10.0)
# self.lstm.add_loss(10.0)
# self.dense.add_loss(10.0)
def call(self, x):
mask = self.emb.compute_mask(x)
x = self.emb(x)
x = self.lstm(x, mask=mask)
x = self.dense(x)
return x
