leofansq / cpcl Goto Github PK

@leofansq Why ignore index=255 is used in the lost function? I have observed that there are 21 classes in prediction where groundtruth has values 0 to 20 and 255? Please explain this.

Great job! I've recently been trying to experiment on your CPCL algorithm using other datasets, so I'm wondering what format the mask of your dataset is in? That way I can modify my dataset accordingly to your format. Thanks!

I would like to know how the files "pseudoseg_labeled_1-x.txt" and "pseudoseg_unlabeled_1-x.txt" in this directory are divided? How are they divided?

Or in other words, why do the contents of these files (image names) originate from the directory "train_aug"? For example, why does "pseudoseg_labeled_1-2.txt" have 733 image names when 733 is not half the number of images (10,582) in the "train_aug" directory? Can you give me the code you used to divide the dataset, or explain it? Recently I would like to test your proposed CPCL method with other datasets.

Thank you very much!

@leofansq I am not able to access pre-trained weights from the Baidu link. Please guide regarding it. i am not able to create account on the Baidu portal as I dont have Chinese phone number.

I tried to reproduce your project, but due to hardware limitations such as GPU video memory, I got a very bad result (mIoU of 55%), which is far from the result in your paper (76%). I was experimenting with 1/8 partitioning. I would like to know what are the possible reasons for this?
Due to the limited GPU memory, I changed the batchsize to 4 and the learning rate to 0.0025, both of which are half of the original configuration, and changed the parameter "pin_memory" from "True" to "False". No other changes have been made. My GPU is a 1080Ti.

Thank you!

When I am trying to reproduce this project,I encountered this problem.What can I do to solve this problem?Thanks in advance!

loss_w_1 = torch.max(F.softmax(logits_cons_tea_1, dim=-1))
loss_w_2 = torch.max(F.softmax(logits_cons_tea_2, dim=-1))

in train.py ,line 181-182, the dim seem to be wrong?