def inception_resnet_block(x,
scale,
filters_branch0,
filters_branch1,
kernel_size,
activation='relu',
bn_scale=False):
"""Adds a Inception-ResNet block.
This function builds only one types of Inception-ResNet block
reduced version of keras/applications/inception_resnet_v2.py
Arguments:
x: input tensor.
scale: scaling factor to scale the residuals (i.e., the output of
passing `x` through an inception module) before adding them
to the shortcut branch.
Let `r` be the output from the residual branch,
the output of this block will be `x + scale * r`.
filters_branch0: number of filters in branch0
filters_branch1: number of filters in branch1
kernel_size: kernel size of conv in branch1
activation: activation function to use at the end of the block
bn_scale: use scale in batch normalization layer
Returns:
Output tensor for the block.
"""
# only one type of branching is supported
branch_0 = utils.conv2d_bn(x, filters_branch0, 1, scale=bn_scale)
branch_1 = utils.conv2d_bn(x, filters_branch0, 1, scale=bn_scale)
branch_1 = utils.conv2d_bn(branch_1,
filters_branch1, [kernel_size, 1],
scale=bn_scale)
branch_1 = utils.conv2d_bn(branch_1,
filters_branch1, [kernel_size, 1],
scale=bn_scale)
branches = [branch_0, branch_1]
mixed = tf.keras.layers.Concatenate()(branches)
up = utils.conv2d_bn(mixed,
tf.keras.backend.int_shape(x)[-1],
1,
activation=None,
scale=bn_scale,
use_bias=True)
x = tf.keras.layers.Lambda(
lambda inputs, scale: inputs[0] + inputs[1] * scale,
output_shape=tf.keras.backend.int_shape(x)[1:],
arguments={'scale': scale})([x, up])
if activation is not None:
x = tf.keras.layers.Activation(activation)(x)
return x
def model(flags):
"""Inception model.
It is based on paper:
Rethinking the Inception Architecture for Computer Vision
http://arxiv.org/abs/1512.00567
Args:
flags: data/model parameters
Returns:
Keras model for training
"""
input_audio = tf.keras.layers.Input(shape=modes.get_input_data_shape(
flags, modes.Modes.TRAINING),
batch_size=flags.batch_size)
net = input_audio
if flags.preprocess == 'raw':
# it is a self contained model, user need to feed raw audio only
net = speech_features.SpeechFeatures(
speech_features.SpeechFeatures.get_params(flags))(net)
# [batch, time, feature]
net = tf.keras.backend.expand_dims(net, axis=-1)
# [batch, time, feature, 1]
for filters in utils.parse(flags.cnn_filters0):
net = tf.keras.layers.SeparableConv2D(filters, (3, 3),
padding='valid',
use_bias=False)(net)
net = tf.keras.layers.BatchNormalization(scale=flags.bn_scale)(net)
net = tf.keras.layers.Activation('relu')(net)
net = tf.keras.layers.MaxPooling2D((3, 3), strides=(2, 2))(net)
# [batch, time, feature, filters]
filters = utils.parse(flags.cnn_filters0)[-1]
net = utils.conv2d_bn(net,
filters, (3, 1),
padding='valid',
scale=flags.bn_scale)
net = utils.conv2d_bn(net,
filters, (1, 3),
padding='valid',
scale=flags.bn_scale)
for stride, filters1, filters2 in zip(utils.parse(flags.cnn_strides),
utils.parse(flags.cnn_filters1),
utils.parse(flags.cnn_filters2)):
if stride > 1:
net = tf.keras.layers.MaxPooling2D((3, 3), strides=stride)(net)
branch1 = utils.conv2d_bn(net, filters2, (1, 1), scale=flags.bn_scale)
branch2 = utils.conv2d_bn(net, filters1, (1, 1), scale=flags.bn_scale)
branch2 = utils.conv2d_bn(branch2,
filters1, (3, 1),
scale=flags.bn_scale)
branch2 = utils.conv2d_bn(branch2,
filters2, (1, 3),
scale=flags.bn_scale)
branch3 = utils.conv2d_bn(net, filters1, (1, 1), scale=flags.bn_scale)
branch3 = utils.conv2d_bn(branch3,
filters1, (3, 1),
scale=flags.bn_scale)
branch3 = utils.conv2d_bn(branch3,
filters1, (1, 3),
scale=flags.bn_scale)
branch3 = utils.conv2d_bn(branch3,
filters1, (3, 1),
scale=flags.bn_scale)
branch3 = utils.conv2d_bn(branch3,
filters2, (1, 3),
scale=flags.bn_scale)
branch4 = tf.keras.layers.AveragePooling2D((3, 3),
strides=(1, 1),
padding='same')(net)
branch4 = utils.conv2d_bn(branch4,
filters2, (1, 1),
scale=flags.bn_scale)
net = tf.keras.layers.concatenate([branch1, branch2, branch3, branch4])
# [batch, time, feature, filters*4]
net = tf.keras.layers.GlobalAveragePooling2D()(net)
# [batch, filters*4]
net = tf.keras.layers.Dropout(flags.dropout)(net)
net = tf.keras.layers.Dense(flags.label_count)(net)
return tf.keras.Model(input_audio, net)
def model(flags):
"""Inception model.
It is based on paper:
Rethinking the Inception Architecture for Computer Vision
http://arxiv.org/abs/1512.00567
Args:
flags: data/model parameters
Returns:
Keras model for training
"""
input_audio = tf.keras.layers.Input(shape=modes.get_input_data_shape(
flags, modes.Modes.TRAINING),
batch_size=flags.batch_size)
net = input_audio
if flags.preprocess == 'raw':
# it is a self contained model, user need to feed raw audio only
net = speech_features.SpeechFeatures(
speech_features.SpeechFeatures.get_params(flags))(net)
# [batch, time, feature]
net = tf.keras.backend.expand_dims(net, axis=2)
# [batch, time, 1, feature]
for stride, filters, kernel_size in zip(
utils.parse(flags.cnn1_strides), utils.parse(flags.cnn1_filters),
utils.parse(flags.cnn1_kernel_sizes)):
net = utils.conv2d_bn(net,
filters, (kernel_size, 1),
padding='valid',
scale=flags.bn_scale)
if stride > 1:
net = tf.keras.layers.MaxPooling2D((3, 1),
strides=(stride, 1))(net)
for stride, filters1, filters2, kernel_size in zip(
utils.parse(flags.cnn2_strides), utils.parse(flags.cnn2_filters1),
utils.parse(flags.cnn2_filters2),
utils.parse(flags.cnn2_kernel_sizes)):
branch1 = utils.conv2d_bn(net, filters1, (1, 1), scale=flags.bn_scale)
branch2 = utils.conv2d_bn(net, filters1, (1, 1), scale=flags.bn_scale)
branch2 = utils.conv2d_bn(branch2,
filters1, (kernel_size, 1),
scale=flags.bn_scale)
branch3 = utils.conv2d_bn(net, filters1, (1, 1), scale=flags.bn_scale)
branch3 = utils.conv2d_bn(branch3,
filters1, (kernel_size, 1),
scale=flags.bn_scale)
branch3 = utils.conv2d_bn(branch3,
filters1, (kernel_size, 1),
scale=flags.bn_scale)
net = tf.keras.layers.concatenate([branch1, branch2, branch3])
# [batch, time, 1, filters*4]
net = utils.conv2d_bn(net, filters2, (1, 1), scale=flags.bn_scale)
# [batch, time, 1, filters2]
if stride > 1:
net = tf.keras.layers.MaxPooling2D((3, 1),
strides=(stride, 1))(net)
net = tf.keras.layers.GlobalAveragePooling2D()(net)
# [batch, filters*4]
net = tf.keras.layers.Dropout(flags.dropout)(net)
net = tf.keras.layers.Dense(flags.label_count)(net)
if flags.return_softmax:
net = tf.keras.layers.Activation('softmax')(net)
return tf.keras.Model(input_audio, net)
def model(flags):
"""Inception resnet model.
It is based on paper:
Inception-v4, Inception-ResNet and the Impact of
Residual Connections on Learning https://arxiv.org/abs/1602.07261
Args:
flags: data/model parameters
Returns:
Keras model for training
"""
input_audio = tf.keras.layers.Input(
shape=modes.get_input_data_shape(flags, modes.Modes.TRAINING),
batch_size=flags.batch_size)
net = input_audio
if flags.preprocess == 'raw':
# it is a self contained model, user need to feed raw audio only
net = speech_features.SpeechFeatures(
speech_features.SpeechFeatures.get_params(flags))(
net)
# [batch, time, feature]
net = tf.keras.backend.expand_dims(net, axis=2)
# [batch, time, 1, feature]
for filters, kernel_size, stride in zip(
utils.parse(flags.cnn1_filters), utils.parse(flags.cnn1_kernel_sizes),
utils.parse(flags.cnn1_strides)):
net = utils.conv2d_bn(
net, filters, (kernel_size, 1), scale=flags.bn_scale, padding='valid')
if stride > 1:
net = tf.keras.layers.MaxPooling2D((3, 1), strides=(stride, 1))(net)
# [batch, time, 1, filters]
for stride, scale, filters_branch0, filters_branch1, filters_branch2, kernel_size in zip(
utils.parse(flags.cnn2_strides), utils.parse(flags.cnn2_scales),
utils.parse(flags.cnn2_filters_branch0),
utils.parse(flags.cnn2_filters_branch1),
utils.parse(flags.cnn2_filters_branch2),
utils.parse(flags.cnn2_kernel_sizes)):
net = inception_resnet_block(
net,
scale,
filters_branch0,
filters_branch1,
kernel_size,
bn_scale=flags.bn_scale)
net = utils.conv2d_bn(
net, filters_branch2, (1, 1), scale=flags.bn_scale, padding='valid')
if stride > 1:
net = tf.keras.layers.MaxPooling2D((3, 1),
strides=(stride, 1),
padding='valid')(
net)
# [batch, time, 1, filters]
net = tf.keras.layers.GlobalAveragePooling2D()(net)
# [batch, filters]
net = tf.keras.layers.Dropout(flags.dropout)(net)
net = tf.keras.layers.Dense(flags.label_count)(net)
if flags.return_softmax:
net = tf.keras.layers.Activation('softmax')(net)
return tf.keras.Model(input_audio, net)
