def block_m3(inputs, filters, training, strides, data_format):
resnet_log_helper.log_begin_block(input_tensor=inputs,
block_type=mlperf_log.BOTTLENECK_BLOCK)
avg_pool = tf.layers.average_pooling2d(inputs=inputs,
pool_size=strides,
strides=strides,
padding='SAME',
data_format=data_format)
shortcut = conv2d_fixed_padding(inputs=avg_pool,
filters=1664,
kernel_size=1,
strides=1,
data_format=data_format)
resnet_log_helper.log_projection(input_tensor=inputs,
output_tensor=shortcut)
shortcut = batch_norm(inputs=shortcut,
training=training,
data_format=data_format)
inputs = conv2d_fixed_padding(inputs=inputs,
filters=256,
kernel_size=1,
strides=1,
data_format=data_format)
inputs = batch_norm(inputs, training, data_format)
mlperf_log.resnet_print(key=mlperf_log.MODEL_HP_RELU)
inputs = tf.nn.relu(inputs)
inputs = conv2d_fixed_padding(inputs=inputs,
filters=256,
kernel_size=3,
strides=strides,
data_format=data_format)
inputs = batch_norm(inputs, training, data_format)
mlperf_log.resnet_print(key=mlperf_log.MODEL_HP_RELU)
inputs = tf.nn.relu(inputs)
inputs = conv2d_fixed_padding(inputs=inputs,
filters=1664,
kernel_size=1,
strides=1,
data_format=data_format)
inputs = batch_norm(inputs, training, data_format)
mlperf_log.resnet_print(key=mlperf_log.MODEL_HP_SHORTCUT_ADD)
inputs += shortcut
mlperf_log.resnet_print(key=mlperf_log.MODEL_HP_RELU)
inputs = tf.nn.relu(inputs)
resnet_log_helper.log_end_block(output_tensor=inputs)
return inputs
def block_m4(inputs, filters, training, strides, data_format):
resnet_log_helper.log_begin_block(input_tensor=inputs,
block_type=mlperf_log.BOTTLENECK_BLOCK)
shortcut = inputs
inputs = conv2d_fixed_padding(inputs=inputs,
filters=384,
kernel_size=1,
strides=1,
data_format=data_format)
inputs = batch_norm(inputs, training, data_format)
mlperf_log.resnet_print(key=mlperf_log.MODEL_HP_RELU)
inputs = tf.nn.relu(inputs)
inputs = conv2d_fixed_padding(
inputs=inputs,
filters=384,
kernel_size=3,
strides=strides, # 384
data_format=data_format)
inputs = batch_norm(inputs, training, data_format)
mlperf_log.resnet_print(key=mlperf_log.MODEL_HP_RELU)
inputs = tf.nn.relu(inputs)
inputs = conv2d_fixed_padding(inputs=inputs,
filters=1664,
kernel_size=1,
strides=1,
data_format=data_format)
inputs = batch_norm(inputs, training, data_format)
mlperf_log.resnet_print(key=mlperf_log.MODEL_HP_SHORTCUT_ADD)
inputs += shortcut
mlperf_log.resnet_print(key=mlperf_log.MODEL_HP_RELU)
inputs = tf.nn.relu(inputs)
resnet_log_helper.log_end_block(output_tensor=inputs)
return inputs
def _bottleneck_block_v1(inputs, filters, training, projection_shortcut,
strides, data_format):
"""A single block for ResNet v1, with a bottleneck.
Similar to _building_block_v1(), except using the "bottleneck" blocks
described in:
Convolution then batch normalization then ReLU as described by:
Deep Residual Learning for Image Recognition
https://arxiv.org/pdf/1512.03385.pdf
by Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun, Dec 2015.
Args:
inputs: A tensor of size [batch, channels, height_in, width_in] or
[batch, height_in, width_in, channels] depending on data_format.
filters: The number of filters for the convolutions.
training: A Boolean for whether the model is in training or inference
mode. Needed for batch normalization.
projection_shortcut: The function to use for projection shortcuts
(typically a 1x1 convolution when downsampling the input).
strides: The block's stride. If greater than 1, this block will ultimately
downsample the input.
data_format: The input format ('channels_last' or 'channels_first').
Returns:
The output tensor of the block; shape should match inputs.
"""
resnet_log_helper.log_begin_block(
input_tensor=inputs, block_type=mlperf_log.BOTTLENECK_BLOCK)
shortcut = inputs
if projection_shortcut is not None:
shortcut = projection_shortcut(inputs)
resnet_log_helper.log_projection(input_tensor=inputs,
output_tensor=shortcut)
shortcut = batch_norm(inputs=shortcut, training=training,
data_format=data_format)
inputs = conv2d_fixed_padding(
inputs=inputs, filters=filters, kernel_size=1, strides=1,
data_format=data_format)
inputs = batch_norm(inputs, training, data_format)
mlperf_log.resnet_print(key=mlperf_log.MODEL_HP_RELU)
inputs = tf.nn.relu(inputs)
inputs = conv2d_fixed_padding(
inputs=inputs, filters=filters, kernel_size=3, strides=strides,
data_format=data_format)
inputs = batch_norm(inputs, training, data_format)
mlperf_log.resnet_print(key=mlperf_log.MODEL_HP_RELU)
inputs = tf.nn.relu(inputs)
inputs = conv2d_fixed_padding(
inputs=inputs, filters=4 * filters, kernel_size=1, strides=1,
data_format=data_format)
inputs = batch_norm(inputs, training, data_format)
mlperf_log.resnet_print(key=mlperf_log.MODEL_HP_SHORTCUT_ADD)
inputs += shortcut
mlperf_log.resnet_print(key=mlperf_log.MODEL_HP_RELU)
inputs = tf.nn.relu(inputs)
resnet_log_helper.log_end_block(output_tensor=inputs)
return inputs
def bottleneck_block(inputs, filters, is_training, strides,
use_projection=False, data_format='channels_first'):
"""Bottleneck block variant for residual networks with BN after convolutions.
Args:
inputs: `Tensor` of size `[batch, channels, height, width]`.
filters: `int` number of filters for the first two convolutions. Note that
the third and final convolution will use 4 times as many filters.
is_training: `bool` for whether the model is in training.
strides: `int` block stride. If greater than 1, this block will ultimately
downsample the input.
use_projection: `bool` for whether this block should use a projection
shortcut (versus the default identity shortcut). This is usually `True`
for the first block of a block group, which may change the number of
filters and the resolution.
data_format: `str` either "channels_first" for `[batch, channels, height,
width]` or "channels_last for `[batch, height, width, channels]`.
Returns:
The output `Tensor` of the block.
"""
if is_training and FLAGS.mlperf_logging:
mlperf_log.resnet_print(
key=mlperf_log.MODEL_HP_BLOCK_TYPE, value=mlperf_log.BOTTLENECK_BLOCK)
resnet_log_helper.log_begin_block(
input_tensor=inputs, block_type=mlperf_log.BOTTLENECK_BLOCK)
shortcut = inputs
if use_projection:
# Projection shortcut only in first block within a group. Bottleneck blocks
# end with 4 times the number of filters.
filters_out = 4 * filters
shortcut = conv2d_fixed_padding(
inputs=inputs,
filters=filters_out,
kernel_size=1,
strides=strides,
is_training=is_training,
data_format=data_format)
shortcut = batch_norm_relu(shortcut, is_training, relu=False,
data_format=data_format)
if is_training and FLAGS.mlperf_logging:
resnet_log_helper.log_projection(
input_tensor=inputs, output_tensor=shortcut)
inputs = conv2d_fixed_padding(
inputs=inputs,
filters=filters,
kernel_size=1,
strides=1,
is_training=is_training,
data_format=data_format)
inputs = batch_norm_relu(inputs, is_training, data_format=data_format)
inputs = conv2d_fixed_padding(
inputs=inputs,
filters=filters,
kernel_size=3,
strides=strides,
is_training=is_training,
data_format=data_format)
inputs = batch_norm_relu(inputs, is_training, data_format=data_format)
inputs = conv2d_fixed_padding(
inputs=inputs,
filters=4 * filters,
kernel_size=1,
strides=1,
is_training=is_training,
data_format=data_format)
inputs = batch_norm_relu(inputs, is_training, relu=False, init_zero=True,
data_format=data_format)
output = tf.nn.relu(inputs + shortcut)
if is_training and FLAGS.mlperf_logging:
mlperf_log.resnet_print(key=mlperf_log.MODEL_HP_SHORTCUT_ADD)
mlperf_log.resnet_print(key=mlperf_log.MODEL_HP_RELU)
resnet_log_helper.log_end_block(output_tensor=output)
return output
def log_begin_block(self, input_tensor, block_type):
if self.model == 'resnet50':
resnet_log_helper.log_begin_block(input_tensor, block_type)
