def conv2d_fixed_padding(inputs,
filters,
kernel_size,
strides,
pruning_method='baseline',
data_format='channels_first',
weight_decay=0.,
name=None):
"""Strided 2-D convolution with explicit padding.
The padding is consistent and is based only on `kernel_size`, not on the
dimensions of `inputs` (as opposed to using `tf.layers.conv2d` alone).
Args:
inputs: Input tensor, float32 or bfloat16 of size [batch, channels, height,
width].
filters: Int specifying number of filters for the first two convolutions.
kernel_size: Int designating size of kernel to be used in the convolution.
strides: Int specifying the stride. If stride >1, the input is downsampled.
pruning_method: String that specifies the pruning method used to identify
which weights to remove.
data_format: String that specifies either "channels_first" for [batch,
channels, height,width] or "channels_last" for [batch, height, width,
channels].
weight_decay: Weight for the l2 regularization loss.
name: String that specifies name for model layer.
Returns:
The output activation tensor of size [batch, filters, height_out, width_out]
Raises:
ValueError: If the data_format provided is not a valid string.
"""
if strides > 1:
inputs = resnet_model.fixed_padding(inputs,
kernel_size,
data_format=data_format)
padding = 'VALID'
else:
padding = 'SAME'
kernel_initializer = tf.variance_scaling_initializer()
kernel_regularizer = contrib_layers.l2_regularizer(weight_decay)
return sparse_conv2d(x=inputs,
units=filters,
activation=None,
kernel_size=[kernel_size, kernel_size],
use_bias=False,
kernel_initializer=kernel_initializer,
kernel_regularizer=kernel_regularizer,
bias_initializer=None,
biases_regularizer=None,
sparsity_technique=pruning_method,
normalizer_fn=None,
strides=[strides, strides],
padding=padding,
data_format=data_format,
name=name)
def depthwise_conv2d_fixed_padding(inputs,
kernel_size,
stride,
data_format='channels_first',
name=None):
"""Depthwise Strided 2-D convolution with explicit padding.
The padding is consistent and is based only on `kernel_size`, not on the
dimensions of `inputs` (as opposed to using `tf.layers.conv2d` alone).
Args:
inputs: Input tensor, float32 or bfloat16 of size [batch, channels, height,
width].
kernel_size: Int designating size of kernel to be used in the convolution.
stride: Int specifying the stride. If stride >1, the input is downsampled.
data_format: String that specifies either "channels_first" for [batch,
channels, height,width] or "channels_last" for [batch, height, width,
channels].
name: String that specifies name for model layer.
Returns:
The output activation tensor of size [batch, filters, height_out, width_out]
Raises:
ValueError: If the data_format provided is not a valid string.
"""
if stride > 1:
inputs = resnet_model.fixed_padding(inputs,
kernel_size,
data_format=data_format)
padding = 'SAME' if stride == 1 else 'VALID'
if data_format == 'channels_last':
data_format_channels = 'NHWC'
elif data_format == 'channels_first':
data_format_channels = 'NCHW'
else:
raise ValueError('Not a valid channel string:', data_format)
return contrib_layers.separable_conv2d(inputs=inputs,
num_outputs=None,
kernel_size=kernel_size,
stride=stride,
padding=padding,
data_format=data_format_channels,
activation_fn=None,
weights_regularizer=None,
biases_initializer=None,
biases_regularizer=None,
scope=name)
def model(inputs, is_training):
"""Creation of the model graph."""
with tf.variable_scope(name, 'resnet_model'):
inputs = resnet_model.fixed_padding(inputs,
kernel_size=3,
data_format=data_format)
padding = 'VALID'
kernel_initializer = tf.variance_scaling_initializer()
kernel_regularizer = contrib_layers.l2_regularizer(weight_decay)
inputs = tf.layers.conv2d(inputs=inputs,
filters=_make_divisible(32 * width),
kernel_size=3,
strides=2,
padding=padding,
use_bias=False,
kernel_initializer=kernel_initializer,
kernel_regularizer=kernel_regularizer,
data_format=data_format,
name='initial_conv')
inputs = tf.identity(inputs, 'initial_conv')
inputs = resnet_model.batch_norm_relu(inputs,
is_training,
data_format=data_format)
mb_block = functools.partial(mbv1_block_,
is_training=is_training,
width=width,
pruning_method=pruning_method,
data_format=data_format,
weight_decay=weight_decay)
inputs = mb_block(inputs, filters=64, stride=1, block_id=0)
inputs = mb_block(inputs, filters=128, stride=2, block_id=1)
inputs = mb_block(inputs, filters=128, stride=1, block_id=2)
inputs = mb_block(inputs, filters=256, stride=2, block_id=3)
inputs = mb_block(inputs, filters=256, stride=1, block_id=4)
inputs = mb_block(inputs, filters=512, stride=2, block_id=5)
inputs = mb_block(inputs, filters=512, stride=1, block_id=6)
inputs = mb_block(inputs, filters=512, stride=1, block_id=7)
inputs = mb_block(inputs, filters=512, stride=1, block_id=8)
inputs = mb_block(inputs, filters=512, stride=1, block_id=9)
inputs = mb_block(inputs, filters=512, stride=1, block_id=10)
inputs = mb_block(inputs, filters=1024, stride=2, block_id=11)
inputs = mb_block(inputs, filters=1024, stride=1, block_id=12)
last_block_filters = _make_divisible(int(1024 * width), 8)
if data_format == 'channels_last':
pool_size = (inputs.shape[1], inputs.shape[2])
elif data_format == 'channels_first':
pool_size = (inputs.shape[2], inputs.shape[3])
inputs = tf.layers.average_pooling2d(inputs=inputs,
pool_size=pool_size,
strides=1,
padding='VALID',
data_format=data_format,
name='final_avg_pool')
inputs = tf.identity(inputs, 'final_avg_pool')
inputs = tf.reshape(inputs, [-1, last_block_filters])
kernel_initializer = tf.variance_scaling_initializer()
kernel_regularizer = contrib_layers.l2_regularizer(weight_decay)
if prune_last_layer:
inputs = sparse_fully_connected(
x=inputs,
units=num_classes,
sparsity_technique=pruning_method
if prune_last_layer else 'baseline',
kernel_initializer=kernel_initializer,
kernel_regularizer=kernel_regularizer,
name='final_dense')
else:
inputs = tf.layers.dense(inputs=inputs,
units=num_classes,
activation=None,
use_bias=True,
kernel_initializer=kernel_initializer,
kernel_regularizer=kernel_regularizer,
name='final_dense')
inputs = tf.identity(inputs, 'final_dense')
return inputs
