def EBlock(self,
last,
channels,
resblocks=1,
bottleneck=True,
kernel=[1, 3],
stride=[1, 2],
format=DATA_FORMAT,
activation=ACTIVATION,
normalizer=None,
regularizer=None,
collections=None):
initializer = tf.initializers.variance_scaling(1.0, 'fan_in', 'normal',
self.random_seed,
self.dtype)
skip = last
in_channels = skip.get_shape()[-3 if format == 'NCHW' else -1]
if activation: last = activation(last)
if bottleneck and in_channels > channels:
last = slim.conv2d(last,
channels, [1, 1], [1, 1],
'SAME',
format,
1,
activation,
None,
weights_initializer=initializer,
weights_regularizer=regularizer,
variables_collections=collections)
last = slim.conv2d(last,
channels,
kernel,
stride,
'SAME',
format,
1,
None,
None,
weights_initializer=initializer,
weights_regularizer=regularizer,
variables_collections=collections)
for i in range(resblocks):
with tf.variable_scope('ResBlock_{}'.format(i)):
last = self.ResBlock(last,
channels,
format=format,
activation=activation,
normalizer=normalizer,
regularizer=regularizer,
collections=collections)
with tf.variable_scope('DenseConnection'):
last = layers.SEUnit(last, channels, format, collections)
if stride != 1 or stride != [1, 1]:
pool_stride = [1, 1] + stride if format == 'NCHW' else [
1
] + stride + [1]
skip = tf.nn.avg_pool(skip, pool_stride, pool_stride, 'SAME',
format)
last = tf.concat([skip, last], -3 if format == 'NCHW' else -1)
return last
def ResBlock(self,
last,
channels,
kernel=[1, 3],
stride=[1, 1],
biases=True,
format=DATA_FORMAT,
dilate=1,
activation=ACTIVATION,
normalizer=None,
regularizer=None,
collections=None):
biases = tf.initializers.zeros(self.dtype) if biases else None
initializer = tf.initializers.variance_scaling(1.0, 'fan_in', 'normal',
self.random_seed,
self.dtype)
skip = last
# pre-activation
if normalizer: last = normalizer(last)
if activation: last = activation(last)
# convolution
last = slim.conv2d(last,
channels,
kernel,
stride,
'SAME',
format,
dilate,
activation,
normalizer,
None,
initializer,
regularizer,
biases,
variables_collections=collections)
last = slim.conv2d(last,
channels,
kernel,
stride,
'SAME',
format,
dilate,
None,
None,
None,
initializer,
regularizer,
biases,
variables_collections=collections)
# residual connection
last = layers.SEUnit(last, channels, format, collections)
last += skip
return last
def EBlock(self,
last,
channels,
resblocks=1,
kernel=[4, 4],
stride=[2, 2],
format=DATA_FORMAT,
activation=ACTIVATION,
normalizer=None,
regularizer=None,
collections=None):
initializer = tf.initializers.variance_scaling(1.0, 'fan_in', 'normal',
self.random_seed,
self.dtype)
skip = last
# pre-activation
if activation: last = activation(last)
# convolution
last = slim.conv2d(last,
channels,
kernel,
stride,
'SAME',
format,
1,
None,
None,
weights_initializer=initializer,
weights_regularizer=regularizer,
variables_collections=collections)
# residual blocks
for i in range(resblocks):
with tf.variable_scope('ResBlock_{}'.format(i)):
last = self.ResBlock(last,
channels,
format=format,
activation=activation,
normalizer=normalizer,
regularizer=regularizer,
collections=collections)
# dense connection
with tf.variable_scope('DenseConnection'):
last = layers.SEUnit(last, channels, format, collections)
if stride != 1 or stride != [1, 1]:
pool_stride = [1, 1] + stride if format == 'NCHW' else [
1
] + stride + [1]
skip = tf.nn.avg_pool(skip, pool_stride, pool_stride, 'SAME',
format)
last = tf.concat([skip, last], -3 if format == 'NCHW' else -1)
return last