def forward(self, conv_high, conv_low):
_, _, A, B, C = conv_high.size()
global_att = self.avg_pool(conv_high).to(dtype=torch.float).cuda()
global_att = self.sigmoid(self.conv1x1x1(global_att))
reweighted_conv_low = conv_low * global_att
reweighted_conv_low = Crop(reweighted_conv_low, [A, B, C])
reweighted_conv_low = self.reshape(reweighted_conv_low)
return conv_high + reweighted_conv_low
def forward(self, input):
x_dilated = self.dilated_conv(input)
x_normal = self.normal_conv(input)
_, _, a, b, c = x_dilated.size()
x_normal = Crop(x_normal, [a, b, c])
x_normal = x_normal
out = x_dilated + x_normal
return out
def forward(self, x):
identity = x
out = self.conv_relu(x)
out = self.conv(out)
_, _, a, b, c = out.size()
identity = Crop(identity, [a, b, c])
identity = self.reshape(identity)
out += identity
out = self.relu(out)
return out
def forward(self, x):
identity = x
out = self.conv_relu(x)
if self.out_channels_2 != 0:
out = self.conv(out)
_, _, a, b, c = out.size()
identity = Crop(identity, [a, b, c])
identity = self.conv1x1(identity)
out += identity
out = self.relu(out)
return out
def forward(self, x, args, triple=False):
cropped_x = Crop(x, [27, 27, 27])
#normal_path
x_normal = self.normal_path(cropped_x)
#dilated_path
x_dilated = self.dilated_path(x)
x_merge = torch.cat((x_normal, x_dilated), 1)
if triple:
att = AdaDropout(100, self.wrs_ratio, self.test_state)
x_FC = self.FC(att(x_merge))
else:
x_FC = self.FC(x_merge)
output = self.classification(x_FC)
return output