def grouped_conv2d_Discriminator(input,
num_outputs,
groups,
is_train,
info=False,
k=3,
s=1,
activation_fn=lrelu,
norm='batch',
name='groupedConv2d'):
with tf.variable_scope(name):
_ = grouped_convolution(input, num_outputs, [k, k], groups, stride=s)
_ = norm_and_act(_, is_train, norm=norm, activation_fn=activation_fn)
if info: print_info(name, _.get_shape().as_list(), activation_fn)
return _
def grouped_conv2d_GsoP(input,
num_outputs,
groups,
is_train=True,
info=False,
a=1,
b=1,
s=1,
name='groupedConv2d'):
with tf.variable_scope(name):
_ = grouped_convolution(input,
num_outputs, [a, b],
groups,
stride=s,
padding='VALID')
# _ = norm_and_act(_, is_train, norm=norm, activation_fn=activation_fn)
if info: print_info(name, _.get_shape().as_list(), activation_fn=None)
return _
def grouped_conv2d_Discriminator_one(input,
num_outputs,
groups,
is_train,
info=False,
a=1,
b=2,
s=1,
activation_fn=lrelu,
norm='batch',
name='groupedConv2dKdifferent'):
with tf.variable_scope(name):
_ = grouped_convolution(input,
num_outputs, [a, b],
groups,
stride=s,
padding='VALID')
_ = norm_and_act(_, is_train, norm=norm, activation_fn=activation_fn)
if info: print_info(name, _.get_shape().as_list(), activation_fn)
return _
def squeezenext_unit(inputs, filters, stride, height_first_order, groups,
seperate_relus):
"""
Squeezenext unit according to:
https://arxiv.org/pdf/1803.10615.pdf
:param inputs:
Input tensor
:param filters:
Number of filters at output of this unit
:param stride:
Input stride
:param height_first_order:
Whether to first perform seperable convolution in the vertical direcation or horizontal direction
:param groups:
Number of groups for some of the convolutions (which ones are different from the paper but equal to:
https://github.com/amirgholami/SqueezeNext/blob/master/1.0-G-SqNxt-23/train_val.prototxt)
:return:
Output tensor, not(height_first_order)
"""
input_channels = inputs.get_shape().as_list()[-1]
shortcut = inputs
out_activation = tf.nn.relu if seperate_relus else None
# shorcut convolution only to be executed if input channels is different from output channels or
# stride is greater than 1.
if input_channels != filters or stride != 1:
shortcut = slim.conv2d(shortcut,
filters, [1, 1],
stride=stride,
activation_fn=out_activation)
# input 1x1 reduction convolutions
block = tfe.grouped_convolution(inputs,
filters / 2, [1, 1],
groups,
stride=stride)
block = slim.conv2d(block, block.get_shape().as_list()[-1] / 2, [1, 1])
# seperable convolutions
if height_first_order:
input_channels_seperated = block.get_shape().as_list()[-1]
block = tfe.grouped_convolution(block, input_channels_seperated * 2,
[3, 1], groups)
block = tfe.grouped_convolution(block,
block.get_shape().as_list()[-1],
[1, 3], groups)
else:
input_channels_seperated = block.get_shape().as_list()[-1]
block = tfe.grouped_convolution(block, input_channels_seperated * 2,
[1, 3], groups)
block = tfe.grouped_convolution(block,
block.get_shape().as_list()[-1],
[3, 1], groups)
# switch order next unit
height_first_order = not height_first_order
# output convolutions
block = slim.conv2d(block,
block.get_shape().as_list()[-1] * 2, [1, 1],
activation_fn=out_activation)
assert block.get_shape().as_list(
)[-1] == filters, "Block output channels not equal to number of specified filters"
return tf.nn.relu(block + shortcut), height_first_order