def call(self, x, *args, **kwargs):
"""Call GAPConv1x1."""
size = ops.get_shape(x)[2:]
out = x
for model in self.children():
out = ops.mean(out)
out = model(out)
out = ops.interpolate(out, size)
return out
def channel_shuffle(x, groups):
"""Shuffle the channel of features.
:param x: feature maps
:type x: tensor
:param groups: group number of channels
:type groups: int
:return: shuffled feature map
:rtype: tensor
"""
batchsize, num_channels, height, width = ops.get_shape(x)
channels_per_group = num_channels // groups
x = ops.View([batchsize, groups, channels_per_group, height, width])(x)
x = ops.Transpose(1, 2)(x)
x = ops.View([batchsize, num_channels, height, width])(x)
return x
def call(self, x):
"""Forward x."""
out = x[self.collect_inds[0]]
for i in range(1, len(self.collect_inds)):
collect = x[self.collect_inds[i]]
if ops.get_shape(out)[2] > ops.get_shape(collect)[2]:
# upsample collect
collect = ops.interpolate(collect, size=ops.get_shape(
out)[2:], mode='bilinear', align_corners=True)
elif ops.get_shape(collect)[2] > ops.get_shape(out)[2]:
out = ops.interpolate(out, size=ops.get_shape(collect)[2:], mode='bilinear', align_corners=True)
if self.agg_concat:
out = ops.concat([out, collect])
else:
out += collect
out = ops.Relu()(out)
return out
def call(self, x1, x2):
"""Do an inference on AggregateCell.
:param x1: first input
:param x2: second input
:return: output
"""
if self.pre_transform:
x1 = self.branch_1(x1)
x2 = self.branch_2(x2)
if tuple(ops.get_shape(x1)[2:]) > tuple(ops.get_shape(x2)[2:]):
x2 = ops.interpolate(x2, size=ops.get_shape(
x1)[2:], mode='bilinear', align_corners=True)
elif tuple(ops.get_shape(x1)[2:]) < tuple(ops.get_shape(x2)[2:]):
x1 = ops.interpolate(x1, size=ops.get_shape(
x2)[2:], mode='bilinear', align_corners=True)
if self.concat:
return self.conv1x1(ops.concat([x1, x2]))
else:
return x1 + x2