seanXYZ
commented
5 years ago If you wanna use resnet 18, make such modifications like below:
-
class FeatureFusionModule(torch.nn.Module): def init(self, num_classes): super().init()
self.in_channels = input_1.channels + input_2.channels
#self.in_channels = 3328 # 3328 = 256(from context path) + 1024(from spatial path) + 2048(from spatial path) self.in_channels = 1024 # 1024 = 256(from context path) + 256(from spatial path) + 512(from spatial path) -
# build attention refinement module for resnet 101 #self.attention_refinement_module1 = AttentionRefinementModule(1024, 1024) #self.attention_refinement_module2 = AttentionRefinementModule(2048, 2048) # build attention refinement module for resnet 18 self.attention_refinement_module1 = AttentionRefinementModule(256, 256) self.attention_refinement_module2 = AttentionRefinementModule(512, 512) -
# supervision block # self.supervision1 = nn.Conv2d(in_channels=1024, out_channels=num_classes, kernel_size=1) # self.supervision2 = nn.Conv2d(in_channels=2048, out_channels=num_classes, kernel_size=1) self.supervision1 = nn.Conv2d(in_channels=256, out_channels=num_classes, kernel_size=1) self.supervision2 = nn.Conv2d(in_channels=512, out_channels=num_classes, kernel_size=1)
Fisrt I met the same error with https://github.com/ooooverflow/BiSeNet/issues/1 using pytorch==0.4.0, and I met other error using pytorch==1.0.0 When I change the version into 0.4.1, it solved. But when I start training, it can work well with resnet101. When I change the build_contextpath into "resnet18", I met the same error with https://github.com/ooooverflow/BiSeNet/issues/3