def bottleneck_resnet_block(grid_structured,
hp,
kernel_size=3,
out_multiple=4):
data_format = "channels_last"
is_training = hp.mode == tf.estimator.ModeKeys.TRAIN
hidden_size_base = hp.hidden_size
filters_out = out_multiple * hidden_size_base
def projection_shortcut(inputs):
inputs = resnet.conv2d_fixed_padding(inputs,
filters_out,
kernel_size=1,
data_format=data_format,
strides=1,
is_training=is_training)
return resnet.batch_norm_relu(inputs,
is_training,
relu=False,
data_format=data_format)
residual = projection_shortcut(grid_structured)
inputs = resnet.conv2d_fixed_padding(inputs=grid_structured,
filters=hidden_size_base,
kernel_size=1,
strides=1,
data_format=data_format,
is_training=is_training)
inputs = resnet.batch_norm_relu(inputs,
is_training,
data_format=data_format)
inputs = resnet.conv2d_fixed_padding(inputs=inputs,
filters=hidden_size_base,
kernel_size=kernel_size,
strides=1,
data_format=data_format,
is_training=is_training)
inputs = resnet.batch_norm_relu(inputs,
is_training,
data_format=data_format)
inputs = resnet.conv2d_fixed_padding(inputs=inputs,
filters=filters_out,
kernel_size=1,
strides=1,
data_format=data_format,
is_training=is_training)
inputs = resnet.batch_norm_relu(inputs,
is_training,
relu=False,
init_zero=False,
data_format=data_format)
return tf.nn.relu(inputs + residual)
def projection_shortcut(inputs):
inputs = resnet.conv2d_fixed_padding(inputs,
filters_out,
kernel_size=1,
data_format=data_format,
strides=1,
is_training=is_training)
return resnet.batch_norm_relu(inputs,
is_training,
relu=False,
data_format=data_format)
def projection_shortcut(inputs):
"""Project identity branch."""
inputs = resnet.conv2d_fixed_padding(inputs=inputs,
filters=filters_out,
kernel_size=1,
strides=strides,
data_format=data_format,
use_td=use_td,
targeting_rate=targeting_rate,
keep_prob=keep_prob,
is_training=is_training)
return resnet.batch_norm_relu(inputs,
is_training,
relu=False,
data_format=data_format)