from keras.layers import Input, Dense, Masking, Dropout, LSTM, Bidirectional, Activation
from keras.layers.merge import dot
from keras.models import Model
from keras import backend as K
from utility import globalvars
def create_softmax_la_network(input_shape, nb_lstm_cells=128, nb_classes=7):
'''
input_shape: (time_steps, features,)
'''
with K.name_scope('BLSTMLayer'):
# Bi-directional Long Short-Term Memory for learning the temporal aggregation
input_feature = Input(shape=input_shape)
x = Masking(mask_value=globalvars.masking_value)(input_feature)
x = Dense(globalvars.nb_hidden_units, activation='relu')(x)
x = Dropout(0.5)(x)
x = Dense(globalvars.nb_hidden_units, activation='relu')(x)
x = Dropout(0.5)(x)
y = Bidirectional(LSTM(nb_lstm_cells, return_sequences=True, dropout=0.5))(x)
with K.name_scope('AttentionLayer'):
# Logistic regression for learning the attention parameters with a standalone feature as input
input_attention = Input(shape=(nb_lstm_cells * 2,))
u = Dense(nb_lstm_cells * 2, activation='softmax')(input_attention)
# To compute the final weights for the frames which sum to unity
alpha = dot([u, y], axes=-1) # inner prod.
alpha = Activation('softmax')(alpha)
with K.name_scope('WeightedPooling'):
# Weighted pooling to get the utterance-level representation
z = dot([alpha, y], axes=1)
# Get posterior probability for each emotional class
output = Dense(nb_classes, activation='softmax')(z)
return Model(inputs=[input_attention, input_feature], outputs=output)
