Questions about output segmentation

[Legendsevl]

Thanks for your creative work! But I want to know how to make the L0,L1,L2,L3 to be the segmention?and L0,L1,L2,L3 are same?
Best wishes!

[josedolz]

Hi, these segmentations come from the attentive feature maps at different levels. You basically need to add a convolution to change the number of channels of your feature maps (from n_channels to number of classes -in this case 5-), and then resample to the original resolution.

This is the piece of code that does this:

Multi-Scale-Attention/my_stacked_danet.py

Lines 1019 to 1032 in 2b3ae2d

	predict4_2 = self.predict4_2(refine4)
	predict3_2 = self.predict3_2(refine3)
	predict2_2 = self.predict2_2(refine2)
	predict1_2 = self.predict1_2(refine1)
	predict1 = F.upsample(predict1, size=x.size()[2:], mode='bilinear')
	predict2 = F.upsample(predict2, size=x.size()[2:], mode='bilinear')
	predict3 = F.upsample(predict3, size=x.size()[2:], mode='bilinear')
	predict4 = F.upsample(predict4, size=x.size()[2:], mode='bilinear')
	predict1_2 = F.upsample(predict1_2, size=x.size()[2:], mode='bilinear')
	predict2_2 = F.upsample(predict2_2, size=x.size()[2:], mode='bilinear')
	predict3_2 = F.upsample(predict3_2, size=x.size()[2:], mode='bilinear')
	predict4_2 = F.upsample(predict4_2, size=x.size()[2:], mode='bilinear')

[Legendsevl]

ok,thanks a lot.