sample_models.py

from keras import backend as K
from keras.models import Model
from keras.layers import (BatchNormalization, Conv1D, Dense, Input, Dropout,
    TimeDistributed, Activation, Bidirectional, SimpleRNN, GRU, LSTM)

def simple_rnn_model(input_dim, output_dim=29):
    """ Build a recurrent network for speech 
    """
    # Main acoustic input
    input_data = Input(name='the_input', shape=(None, input_dim))
    # Add recurrent layer
    simp_rnn = GRU(output_dim, return_sequences=True, 
                 implementation=2, name='rnn')(input_data)
    # Add softmax activation layer
    y_pred = Activation('softmax', name='softmax')(simp_rnn)
    # Specify the model
    model = Model(inputs=input_data, outputs=y_pred)
    model.output_length = lambda x: x
    print(model.summary())
    return model

def rnn_model(input_dim, units, activation, output_dim=29):
    """ Build a recurrent network for speech 
    """
    # Main acoustic input
    input_data = Input(name='the_input', shape=(None, input_dim))
    # Add recurrent layer
    simp_rnn = GRU(units, activation=activation,
        return_sequences=True, implementation=2, name='rnn')(input_data)
    # Add batch normalization
    bn_rnn = BatchNormalization(name='bn_rnn')(simp_rnn)
    # Add a TimeDistributed(Dense(output_dim)) layer
    time_dense = TimeDistributed(Dense(output_dim), name='time_dense')(bn_rnn)
    # Add softmax activation layer
    y_pred = Activation('softmax', name='softmax')(time_dense)
    # Specify the model
    model = Model(inputs=input_data, outputs=y_pred)
    model.output_length = lambda x: x
    print(model.summary())
    return model


def cnn_rnn_model(input_dim, filters, kernel_size, conv_stride,
    conv_border_mode, units, output_dim=29):
    """ Build a recurrent + convolutional network for speech 
    """
    # Main acoustic input
    input_data = Input(name='the_input', shape=(None, input_dim))
    # Add convolutional layer
    conv_1d = Conv1D(filters, kernel_size, 
                     strides=conv_stride, 
                     padding=conv_border_mode,
                     activation='relu',
                     name='conv1d')(input_data)
    # Add batch normalization
    bn_cnn = BatchNormalization(name='bn_conv_1d')(conv_1d)
    # Add a recurrent layer
    simp_rnn = SimpleRNN(units, activation='relu',
        return_sequences=True, implementation=2, name='rnn')(bn_cnn)
    # Add batch normalization
    bn_rnn = BatchNormalization(name='bn_rnn')(simp_rnn)
    # Add a TimeDistributed(Dense(output_dim)) layer
    time_dense = TimeDistributed(Dense(output_dim), name='time_dense')(bn_rnn)
    # Add softmax activation layer
    y_pred = Activation('softmax', name='softmax')(time_dense)
    # Specify the model
    model = Model(inputs=input_data, outputs=y_pred)
    model.output_length = lambda x: cnn_output_length(
        x, kernel_size, conv_border_mode, conv_stride)
    print(model.summary())
    return model

def cnn_output_length(input_length, filter_size, border_mode, stride,
                       dilation=1):
    """ Compute the length of the output sequence after 1D convolution along
        time. Note that this function is in line with the function used in
        Convolution1D class from Keras.
    Params:
        input_length (int): Length of the input sequence.
        filter_size (int): Width of the convolution kernel.
        border_mode (str): Only support `same` or `valid`.
        stride (int): Stride size used in 1D convolution.
        dilation (int)
    """
    if input_length is None:
        return None
    assert border_mode in {'same', 'valid'}
    dilated_filter_size = filter_size + (filter_size - 1) * (dilation - 1)
    if border_mode == 'same':
        output_length = input_length
    elif border_mode == 'valid':
        output_length = input_length - dilated_filter_size + 1
    return (output_length + stride - 1) // stride

def deep_rnn_model(input_dim, units, recur_layers, output_dim=29):
    """ Build a deep recurrent network for speech 
    """
    # Main acoustic input
    input_data = Input(name='the_input', shape=(None, input_dim))
    # Add recurrent layers, each with batch normalization
    rnn = input_data
    for i in range(recur_layers):
        rnn = GRU(units, activation='relu', return_sequences=True,
                  implementation=2, name='rnn%d' % (i+1))(rnn)
        rnn = BatchNormalization(name='bn%d' % (i+1))(rnn)
    # Add a TimeDistributed(Dense(output_dim)) layer
    time_dense = TimeDistributed(Dense(output_dim), name='time_dense')(rnn)
    # Add softmax activation layer
    y_pred = Activation('softmax', name='softmax')(time_dense)
    # Specify the model
    model = Model(inputs=input_data, outputs=y_pred)
    model.output_length = lambda x: x
    print(model.summary())
    return model

def bidirectional_rnn_model(input_dim, units, output_dim=29):
    """ Build a bidirectional recurrent network for speech
    """
    # Main acoustic input
    input_data = Input(name='the_input', shape=(None, input_dim))
    # Add bidirectional recurrent layer
    bidir_rnn = Bidirectional(GRU(units, return_sequences=True), name='bidir_rnn')(input_data)
    # Add a TimeDistributed(Dense(output_dim)) layer
    time_dense = TimeDistributed(Dense(output_dim), name='time_dense')(bidir_rnn)
    # Add softmax activation layer
    y_pred = Activation('softmax', name='softmax')(time_dense)
    # Specify the model
    model = Model(inputs=input_data, outputs=y_pred)
    model.output_length = lambda x: x
    print(model.summary())
    return model

def final_model():
    """ Build a deep network for speech 
    """

    # parameters of the convolutional layer
    filters = 200
    kernel_size = 11
    stride = 2

    # number of units in recurrent layers
    rnn_units = (200, 200, 200)

    input_dim = 13 # use MFCC features
    output_dim = 29

    # Main acoustic input
    input_data = Input(name='the_input', shape=(None, input_dim))

    # Specify the layers in your network
    current_layer = input_data

    current_layer = Conv1D(filters, kernel_size=kernel_size, strides=stride, padding='causal',
                           activation='relu', name='conv_1')(current_layer)
    current_layer = Dropout(0.5)(current_layer)
    current_layer = BatchNormalization(name='bn_conv1')(current_layer)

    for i in range(len(rnn_units)):
        current_layer = GRU(rnn_units[i], activation='relu',
                            return_sequences=True, name='rnn%d'%(i+1))(current_layer)
        current_layer = BatchNormalization(name='bn_rnn%d'%(i+1))(current_layer)

    current_layer = Dropout(0.2)(current_layer)

    time_dense = TimeDistributed(Dense(output_dim), name='time_dense')(current_layer)
    # Add softmax activation layer
    y_pred = Activation('softmax', name='softmax')(time_dense)
    # Specify the model
    model = Model(inputs=input_data, outputs=y_pred)
    # Specify model.output_length
    model.output_length = lambda x: cnn_output_length(x, kernel_size, 'same', stride)

    print(model.summary())
    return model