Question about loss function

[Lecxxx]

Thank you for your outstanding work, but now I have a few questions about the Loss function in the code. In line 180 of model. py, you annotated the following code： cha1, row1, col1 = self.mask_a[3].shape[1], self.mask_a[3].shape[2], self.mask_a[3].shape[3] mask_a_gt = torch.rand(opts.batch_size, cha1, row1, col1).cuda() loss_att_a = self.criterionL2(self.mask_a[5], mask_a_gt) * 10

This makes the attention loss only contain half of the description in your paper, which conflicts with the description in the paper. Is this a problem?

[OaDsis]

Yes, that is a mistake, you can cancel the annotation.

[Lecxxx]

@OaDsis Thanks for your reply. I cancel the annotation and modified the Loss function to the following form： loss_G = loss_G_GAN_A + loss_G_GAN_B + \ loss_G_L1_A + loss_G_L1_B + \ loss_perceptual + \ loss_cons + \ loss_att_b + loss_att_a

The results during the training process are in the following form： total_it: 1 (ep 0, it 1), lr 0.000100, disA 0.454318, disB 0.500568, ganA 0.973853, ganB 0.973228, recA 4.425363, recB 4.133862, percp 0.880033, cons_loss 2.424589, attB 0.001361, attA 3.451179, total 17.263468 total_it: 2 (ep 0, it 2), lr 0.000100, disA 0.416719, disB 0.463749, ganA 0.912895, ganB 0.900647, recA 4.648975, recB 3.706006, percp 0.746614, cons_loss 2.885847, attB 0.000231, attA 3.364104, total 17.165318 total_it: 3 (ep 0, it 3), lr 0.000100, disA 0.399364, disB 0.425298, ganA 0.838068, ganB 0.805129, recA 3.031458, recB 3.395942, percp 0.725828, cons_loss 1.318987, attB 0.000112, attA 3.283001, total 13.398525 total_it: 4 (ep 0, it 4), lr 0.000100, disA 0.349654, disB 0.377379, ganA 0.745493, ganB 0.688415, recA 4.654613, recB 2.549643, percp 0.787967, cons_loss 3.059796, attB 0.001204, attA 3.235075, total 15.722206 total_it: 5 (ep 0, it 5), lr 0.000100, disA 0.285982, disB 0.324985, ganA 0.632568, ganB 0.546403, recA 4.403074, recB 2.788066, percp 0.822683, cons_loss 2.784659, attB 0.004070, attA 3.141724, total 15.123247 total_it: 6 (ep 0, it 6), lr 0.000100, disA 0.256373, disB 0.261119, ganA 0.482774, ganB 0.388865, recA 3.252872, recB 3.830938, percp 1.007862, cons_loss 1.571942, attB 0.007443, attA 3.076305, total 13.619002 ... total_it: 393 (ep 1, it 193), lr 0.000100, disA 0.230090, disB 0.243723, ganA 0.313256, ganB 0.452222, recA 1.619573, recB 0.958254, percp 0.542632, cons_loss 0.319973, attB 0.498841, attA 1.572791, total 6.277541 total_it: 394 (ep 1, it 194), lr 0.000100, disA 0.183904, disB 0.337638, ganA 0.351476, ganB 0.603120, recA 1.062369, recB 0.992846, percp 0.322584, cons_loss 0.190934, attB 0.510730, attA 1.636341, total 5.670399 total_it: 395 (ep 1, it 195), lr 0.000100, disA 0.208328, disB 0.235580, ganA 0.536630, ganB 0.382105, recA 1.527501, recB 1.149172, percp 0.353436, cons_loss 0.281888, attB 0.513301, attA 1.520630, total 6.264664 total_it: 396 (ep 1, it 196), lr 0.000100, disA 0.241589, disB 0.231161, ganA 0.113991, ganB 0.529447, recA 1.181571, recB 1.609867, percp 0.367512, cons_loss 0.177205, attB 0.500499, attA 1.643606, total 6.123698 total_it: 397 (ep 1, it 197), lr 0.000100, disA 0.168638, disB 0.325995, ganA 0.252802, ganB 0.252872, recA 2.025960, recB 1.257719, percp 0.423531, cons_loss 0.238254, attB 0.529761, attA 1.571633, total 6.552532 total_it: 398 (ep 1, it 198), lr 0.000100, disA 0.184195, disB 0.348183, ganA 0.307939, ganB 0.246130, recA 1.996846, recB 1.250796, percp 0.378959, cons_loss 0.439748, attB 0.550344, attA 1.528978, total 6.699740 total_it: 399 (ep 1, it 199), lr 0.000100, disA 0.291252, disB 0.241256, ganA 0.271846, ganB 0.480793, recA 1.353552, recB 1.062438, percp 0.246189, cons_loss 0.137802, attB 0.517819, attA 1.617824, total 5.688264 Is the value of attA loss normal? Can you compare it with your value during training?

[Lecxxx]

@OaDsis Hello！The design of the Loss function in the notes does not correspond to the original text, as shown below：

[skyknights]

@OaDsis Hello！The design of the Loss function in the notes does not correspond to the original text, as shown below：

Hello, is the current project's loss function correct? Can it be consistent with the one in the paper?

[OaDsis]

Yes, you can!