def identity_bottleneck_1d(inputs, depth, depth_bottleneck, stride, v, rate=1, outputs_collections=None,
scope=None, use_bounded_activations=False, attention_module=None):
"""
Args:
inputs: A tensor of size [batch, height, width, channels].
depth、depth_bottleneck、stride三个参数是前面blocks类中的args
depth: 一个block中的某个unit中(第三个conv)输出的feature-map的个数
depth_bottleneck: 一个block中的某个unit中(前面两个conv)输出的feature-map个数
stride: 是short_cut路径对于para_inputs/pre_act(经过bn层的para_inputs)的subsample_2d的步长 -- (是否经过bn层主要看输入输出通道数是否一致)
以及unit中conv-2的步长
rate: An integer, rate for atrous convolution.
outputs_collections: 是收集end_points的collection
scope: 是这个unit的名称。
todo what is quantized inference??? what is bounded activations??? => must make it clear before any runs
use_bounded_activations: Whether or not to use bounded activations. Bounded
activations better lend themselves to quantized inference.
attention_module: SE-blocks or SESE-blocks
"""
with tf.variable_scope(scope, 'bottleneck_v1', [inputs]) as sc:
depth_in = slim.utils.last_dimension(inputs.get_shape(), min_rank=4)
# shortcut fine-tune
if depth == depth_in:
# pre se/sese in shortcut
if attention_module == 'se_block':
shortcut = se_block(inputs, name='se_block', ratio=2 if inputs.get_shape()[-1] <= 8 else 8)
elif attention_module == 'sese_block':
# todo ratio to be defined...
shortcut = sese_block(input_feature=inputs, name='sese_block', v=v, ratio=2 if inputs.get_shape()[-1] <= 8 else 8)
else:
shortcut = inputs
shortcut = subsample_1d(shortcut, stride, 'shortcut')
else:
# pre se/sese in shortcut
if attention_module == 'se_block':
shortcut = se_block(inputs, name='se_block', ratio=2 if inputs.get_shape()[-1] <= 8 else 8)
elif attention_module == 'sese_block':
# todo ratio to be defined...
shortcut = sese_block(input_feature=inputs, name='sese_block', v=v, ratio=2 if inputs.get_shape()[-1] <= 8 else 8)
else:
shortcut = inputs
shortcut = slim.conv2d(shortcut, depth, [1, 1], stride=stride, activation_fn=tf.nn.relu6 if use_bounded_activations else None,
scope='shortcut')
# convs
residual = slim.conv2d(inputs, depth_bottleneck, [1, 1], stride=1, scope='conv1')
residual = conv1d_same(residual, depth_bottleneck, 3, stride, rate=rate, scope='conv2')
residual = slim.conv2d(residual, depth, [1, 1], stride=1, activation_fn=None, scope='conv3')
# post-relu
if use_bounded_activations:
# Use clip_by_value to simulate bandpass activation.
# todo why using bandpass act?????
residual = tf.clip_by_value(residual, -6.0, 6.0)
output = tf.nn.relu6(shortcut + residual)
else:
output = tf.nn.relu(shortcut + residual)
return slim.utils.collect_named_outputs(outputs_collections, sc.name, output)
def resnet_v1_1d(inputs, blocks, num_classes=None, is_training=True, global_pool=True, output_stride=None,
include_root_block=True, spatial_squeeze=True, store_non_strided_activations=False,
reuse=None, scope=None, s=None):
"""
Args:
s = None (s is short for switch)!!!
output_stride: If None, then the output will be computed at the nominal
network stride. If output_stride is not None, it specifies the requested
ratio of input to output spatial resolution.
store_non_strided_activations: If True, we compute non-strided (undecimated)
activations at the last unit of each block and store them in the
`outputs_collections` before subsampling them. This gives us access to
higher resolution intermediate activations which are useful in some
dense prediction problems but increases 4x the computation and memory cost
at the last unit of each block.
reuse: whether or not the network and its variables should be reused. To be
able to reuse 'scope' must be given.
scope: Optional variable_scope.
"""
with tf.variable_scope(scope, 'resnet_v1', [inputs], reuse=reuse) as sc:
end_points_collection = sc.original_name_scope + '_end_points'
with slim.arg_scope([slim.conv2d, standard_bottleneck_1d if s == 0 else pre_bottleneck_1d if s == 1
else identity_bottleneck_1d, stack_blocks_dense_1d], outputs_collections=end_points_collection):
# todo if not training => enter noop scope!
with (slim.arg_scope([slim.batch_norm], is_training=is_training) if is_training is not None else NoOpScope()):
net = inputs
if include_root_block:
if output_stride is not None:
if output_stride % 4 != 0:
raise ValueError('The output_stride needs to be a multiple of 4.')
output_stride /= 4
net = conv1d_same(net, num_outputs=8, kernel_size=4, stride=2, scope='conv1')
net = slim.max_pool2d(net, kernel_size=[1, 3], stride=[1, 2], scope='pool1')
# todo stack 函数根据blocks中的args参数dict进行参数解析
net = stack_blocks_dense_1d(net, blocks, output_stride, store_non_strided_activations)
# convert to dict 'end_points'
end_points = slim.utils.convert_collection_to_dict(end_points_collection)
if global_pool: # gap
net = tf.reduce_mean(net, [1, 2], name='pool5', keepdims=True)
end_points['global_pool'] = net
if num_classes:
net = slim.conv2d(net, num_classes, [1, 1], activation_fn=None, normalizer_fn=None, scope='logits')
end_points[sc.name + '/logits'] = net
if spatial_squeeze:
# todo ... check this before run ->
net = tf.squeeze(net, [1, 2], name='SpatialSqueeze')
end_points[sc.name + '/spatial_squeeze'] = net
end_points['predictions'] = slim.softmax(net, scope='predictions')
return net, end_points
def pre_bottleneck_1d(inputs,
depth,
depth_bottleneck,
stride,
v,
rate=1,
outputs_collections=None,
scope=None,
attention_module=None):
"""
Args:
inputs: A tensor of size [batch, height, width, channels].
depth、depth_bottleneck、stride三个参数是前面blocks类中的args
depth: 一个block中的某个unit中(第三个conv)输出的feature-map的个数
depth_bottleneck: 一个block中的某个unit中(前面两个conv)输出的feature-map个数
stride: 是short_cut路径对于para_inputs/pre_act(经过bn层的para_inputs)的subsample_2d的步长 -- (是否经过bn层主要看输入输出通道数是否一致)
以及unit中conv-2的步长
rate: An integer, rate for atrous convolution.
outputs_collections: 是收集end_points的collection
scope: 是这个unit的名称
attention_module: SE-blocks or SESE-blocks
"""
with tf.variable_scope(scope, 'bottleneck_v2', [inputs]) as sc:
depth_in = slim.utils.last_dimension(inputs.get_shape(), min_rank=4)
# add se
if attention_module == 'se_block':
residual = se_block(inputs,
name='se_block',
ratio=2 if inputs.get_shape()[-1] <= 8 else 8)
# add sese
elif attention_module == 'sese_block':
# todo ratio to be defined...
residual = sese_block(
input_feature=inputs,
name='sese_block',
v=v,
ratio=2 if inputs.get_shape()[-1] <= 8 else 8)
# no other block implemented
else:
residual = inputs
# pre activate
preact = slim.batch_norm(inputs,
activation_fn=tf.nn.relu,
scope='preact')
# shortcut fine-tune
if depth == depth_in:
shortcut = subsample_1d(inputs, stride, 'shortcut')
else:
shortcut = slim.conv2d(preact,
depth, [1, 1],
stride=stride,
normalizer_fn=None,
activation_fn=None,
scope='shortcut')
# convs
residual = slim.conv2d(residual,
depth_bottleneck, [1, 1],
stride=1,
scope='conv1')
residual = conv1d_same(residual,
depth_bottleneck,
3,
stride,
rate=rate,
scope='conv2')
residual = slim.conv2d(residual,
depth, [1, 1],
stride=1,
normalizer_fn=None,
activation_fn=None,
scope='conv3')
# junction
output = shortcut + residual
return slim.utils.collect_named_outputs(outputs_collections, sc.name,
output)
def resnet_v2_1d(inputs,
blocks,
num_classes=None,
is_training=True,
global_pool=True,
output_stride=None,
include_root_block=True,
spatial_squeeze=True,
reuse=None,
scope=None,
s=None):
"""
implementation for resnet_v2 1d | more detail see tf/slim/.../nets/resnet_v2_discarded.py
"""
with tf.variable_scope(scope, 'resnet_v2', [inputs], reuse=reuse) as sc:
end_points_collection = sc.original_name_scope + '_end_points'
# todo slim.conv2d?????????
with slim.arg_scope([
slim.conv2d, standard_bottleneck_1d if s == 0 else
pre_bottleneck_1d if s == 1 else identity_bottleneck_1d,
stack_blocks_dense_1d
],
outputs_collections=end_points_collection):
with slim.arg_scope(
[slim.batch_norm],
is_training=is_training): # 单独为batch_norm设置train的参数状态
net = inputs
if include_root_block:
if output_stride is not None:
if output_stride % 4 != 0:
raise ValueError(
'The output_stride needs to be a multiple of 4.'
)
output_stride /= 4
# We do not include batch normalization or activation functions in
# conv1 because the first ResNet unit will perform these. Cf.
# Appendix of [2].
with slim.arg_scope([slim.conv2d],
activation_fn=None,
normalizer_fn=None):
net = conv1d_same(net,
num_outputs=8,
kernel_size=4,
stride=2,
scope='conv1')
net = slim.max_pool2d(net, [1, 3],
stride=[1, 2],
scope='pool1')
net = stack_blocks_dense_1d(net, blocks, output_stride)
# This is needed because the pre-activation variant does not have batch
# normalization or activation functions in the residual unit output. See
# Appendix of [2].
net = slim.batch_norm(net,
activation_fn=tf.nn.relu,
scope='postnorm')
# Convert end_points_collection into a dictionary of end_points.
end_points = slim.utils.convert_collection_to_dict(
end_points_collection)
if global_pool:
# Global average pooling.
net = tf.reduce_mean(
net, [1, 2], name='pool5',
keepdims=True) # (x,1,xx,xxx) -> (x,xxx)
end_points['global_pool'] = net
if num_classes:
net = slim.conv2d(net,
num_classes, [1, 1],
activation_fn=None,
normalizer_fn=None,
scope='logits')
end_points[sc.name + '/logits'] = net
# todo make it clear before run
if spatial_squeeze:
net = tf.squeeze(net, [1, 2], name='SpatialSqueeze')
end_points[sc.name + '/spatial_squeeze'] = net
end_points['predictions'] = slim.softmax(
net, scope='predictions')
return net, end_points
