def projection_block_3D(input_tensor,
filters,
kernel_size=(3, 3, 3),
stage=0,
block=0,
se_enabled=False,
se_ratio=16):
numFilters1, numFilters2 = filters
if K.image_data_format() == 'channels_last':
bn_axis = -1
else:
bn_axis = 1
conv_name_base = 'res' + str(stage) + '_' + str(block) + '_branch'
bn_name_base = 'bn' + str(stage) + '_' + str(block) + '_branch'
# downsampling directly by convolution with stride 2
x = Conv3D(filters=numFilters1,
kernel_size=kernel_size,
strides=(2, 2, 2),
padding='same',
kernel_initializer='he_normal',
name=conv_name_base + '2a')(input_tensor)
x = BatchNormalization(axis=bn_axis, name=bn_name_base + '2a')(x)
x = LeakyReLU(alpha=0.01)(x)
x = Conv3D(filters=numFilters2,
kernel_size=kernel_size,
strides=(1, 1, 1),
padding='same',
kernel_initializer='he_normal',
name=conv_name_base + '2b')(x)
# squeeze and excitation block
if se_enabled:
x = squeeze_excitation_block_3D(x, ratio=se_ratio)
x = BatchNormalization(axis=bn_axis, name=bn_name_base + '2b')(x)
# projection shortcut convolution
x_shortcut = Conv3D(filters=numFilters2,
kernel_size=(2, 2, 2),
strides=(2, 2, 2),
padding='same',
kernel_initializer='he_normal',
name=conv_name_base + '1')(input_tensor)
x_shortcut = BatchNormalization(axis=bn_axis,
name=bn_name_base + '1')(x_shortcut)
# addition of shortcut
x = Add()([x, x_shortcut])
x = LeakyReLU(alpha=0.01)(x)
return x
def zero_padding_block(input_tensor,
filters,
stage,
block,
se_enabled=False,
se_ratio=16):
numFilters1, numFilters2 = filters
if K.image_data_format() == 'channels_last':
bn_axis = -1
else:
bn_axis = 1
conv_name_base = 'res' + str(stage) + '_' + str(block) + '_branch'
bn_name_base = 'bn' + str(stage) + '_' + str(block) + '_branch'
# downsampling directly by convolution with stride 2
x = Conv2D(numFilters1, (3, 3),
strides=(2, 2),
kernel_initializer='he_normal',
name=conv_name_base + '2a')(input_tensor)
x = BatchNormalization(axis=bn_axis, name=bn_name_base + '2a')(x)
x = Activation('relu')(x)
x = Conv2D(numFilters2, (3, 3),
kernel_initializer='he_normal',
name=conv_name_base + '2b')(x)
# squeeze and excitation block
if se_enabled:
x = squeeze_excitation_block(x, ratio=se_ratio)
x = BatchNormalization(axis=bn_axis, name=bn_name_base + '2b')(x)
# zero padding and downsampling with 1x1 conv shortcut connection
x_shortcut = Conv2D(1, (1, 1),
strides=(2, 2),
kernel_initializer='he_normal',
name=conv_name_base + '1')(input_tensor)
x_shortcut2 = MaxPooling2D(pool_size=(1, 1),
strides=(2, 2),
border_mode='same')(input_tensor)
x_shortcut = Lambda(zeropad, output_shape=zeropad_output_shape)(x_shortcut)
x_shortcut = BatchNormalization(axis=bn_axis,
name=bn_name_base + '1')(x_shortcut)
# addition of shortcut
x = Add()([x, x_shortcut])
x = Activation('relu')(x)
return x
def projection_bottleneck_block(input_tensor,
filters,
stage,
block,
se_enabled=False,
se_ratio=16):
numFilters1, numFilters2, numFilters3 = filters
if K.image_data_format() == 'channels_last':
bn_axis = -1
else:
bn_axis = 1
conv_name_base = 'res' + str(stage) + '_' + str(block) + '_branch'
bn_name_base = 'bn' + str(stage) + '_' + str(block) + '_branch'
x = Conv2D(numFilters1, (1, 1),
strides=(2, 2),
kernel_initializer='he_normal',
name=conv_name_base + '2a')(input_tensor)
x = BatchNormalization(axis=bn_axis, name=bn_name_base + '2a')(x)
x = Activation('relu')(x)
x = Conv2D(numFilters2, (3, 3),
padding='same',
kernel_initializer='he_normal',
name=conv_name_base + '2b')(x)
x = BatchNormalization(axis=bn_axis, name=bn_name_base + '2b')(x)
x = Activation('relu')(x)
x = Conv2D(numFilters3, (1, 1),
kernel_initializer='he_normal',
name=conv_name_base + '2c')(x)
# squeeze and excitation block
if se_enabled:
x = squeeze_excitation_block(x, ratio=se_ratio)
x = BatchNormalization(axis=bn_axis, name=bn_name_base + '2c')(x)
# projection shortcut
x_shortcut = Conv2D(numFilters3, (1, 1),
strides=(2, 2),
kernel_initializer='he_normal',
name=conv_name_base + '1')(input_tensor)
x_shortcut = BatchNormalization(axis=bn_axis,
name=bn_name_base + '1')(x_shortcut)
x = Add()([x, x_shortcut])
x = Activation('relu')(x)
return x
def identity_block_3D(input_tensor,
filters,
kernel_size=(3, 3, 3),
stage=0,
block=0,
se_enabled=False,
se_ratio=16):
numFilters1, numFilters2 = filters
if K.image_data_format() == 'channels_last':
bn_axis = -1
else:
bn_axis = 1
conv_name_base = 'res' + str(stage) + '_' + str(block) + '_branch'
bn_name_base = 'bn' + str(stage) + '_' + str(block) + '_branch'
x = Conv3D(filters=numFilters1,
kernel_size=kernel_size,
strides=(1, 1, 1),
padding='same',
kernel_initializer='he_normal',
name=conv_name_base + '2a')(input_tensor)
x = BatchNormalization(axis=bn_axis, name=bn_name_base + '2a')(x)
x = LeakyReLU(alpha=0.01)(x)
x = Conv3D(filters=numFilters2,
kernel_size=kernel_size,
strides=(1, 1, 1),
padding='same',
kernel_initializer='he_normal',
name=conv_name_base + '2b')(x)
# squeeze and excitation block
if se_enabled:
x = squeeze_excitation_block_3D(x, ratio=se_ratio)
x = BatchNormalization(axis=bn_axis, name=bn_name_base + '2b')(x)
x = Add()([x, input_tensor])
x = LeakyReLU(alpha=0.01)(x)
return x
def __init__(self,
n_word_vocab=50001,
n_role_vocab=7,
n_factors_emb=256,
n_factors_cls=512,
n_hidden=256,
word_vocabulary={},
role_vocabulary={},
unk_word_id=50000,
unk_role_id=7,
missing_word_id=50001,
using_dropout=False,
dropout_rate=0.3,
optimizer='adagrad',
loss='sparse_categorical_crossentropy',
metrics=['accuracy']):
super(NNRF_ResROFA,
self).__init__(n_word_vocab, n_role_vocab, n_factors_emb,
n_hidden, word_vocabulary, role_vocabulary,
unk_word_id, unk_role_id, missing_word_id,
using_dropout, dropout_rate, optimizer, loss,
metrics)
# minus 1 here because one of the role is target role
self.input_length = n_role_vocab - 1
# each input is a fixed window of frame set, each word correspond to one role
input_words = Input(
shape=(self.input_length, ), dtype=tf.uint32,
name='input_words') # Switched dtype to tf specific (team1-change)
input_roles = Input(
shape=(self.input_length, ), dtype=tf.uint32,
name='input_roles') # Switched dtype to tf specific (team1-change)
target_role = Input(
shape=(1, ), dtype=tf.uint32,
name='target_role') # Switched dtype to tf specific (team1-change)
# role based embedding layer
embedding_layer = role_based_word_embedding(
input_words, input_roles, n_word_vocab, n_role_vocab,
glorot_uniform(), missing_word_id, self.input_length,
n_factors_emb, True, using_dropout, dropout_rate)
# fully connected layer, output shape is (batch_size, input_length, n_hidden)
lin_proj = Dense(n_factors_emb,
activation='linear',
use_bias=False,
input_shape=(n_factors_emb, ),
name='lin_proj')(embedding_layer)
non_lin = PReLU(alpha_initializer='ones', name='non_lin')(lin_proj)
# fully connected layer, output shape is (batch_size, input_length, n_hidden)
lin_proj2 = Dense(n_factors_emb,
activation='linear',
use_bias=False,
input_shape=(n_factors_emb, ),
name='lin_proj2')(non_lin)
residual_0 = Add(name='residual_0')([embedding_layer, lin_proj2])
# mean on input_length direction;
# obtaining context embedding layer, shape is (batch_size, n_hidden)
context_embedding = Lambda(lambda x: K.mean(x, axis=1),
name='context_embedding',
output_shape=(n_factors_emb, ))(residual_0)
# hidden layer
hidden_layer2 = target_word_hidden(context_embedding,
target_role,
n_word_vocab,
n_role_vocab,
glorot_uniform(),
n_factors_cls,
n_hidden,
using_dropout=using_dropout,
dropout_rate=dropout_rate)
# softmax output layer
output_layer = Dense(n_word_vocab,
activation='softmax',
input_shape=(n_factors_cls, ),
name='softmax_word_output')(hidden_layer2)
self.model = Model(inputs=[input_words, input_roles, target_role],
outputs=[output_layer])
self.model.compile(optimizer, loss, metrics)