Ask questions

[liheyoung]

Ask questions

[LiheYoung]

Thank you for raising these questions.

1. Among the two filtering conditions, conf_u_w_cutmixed2 >= cfg['conf_thresh'] is for selecting high-confidence pseudo labels, while ignore_mask_cutmixed2 != 255 is to avoid the loss computation in padded regions from image pre-processing. And we only divide the second conditiontorch.sum(ignore_mask_cutmixed2 != 255).item() in the third line, because we hope ((conf_u_w_cutmixed2 >= cfg['conf_thresh']) & (ignore_mask_cutmixed2 != 255)).sum() / (ignore_mask_cutmixed2 != 255).sum() can serve as an adaptive weight for our unsupervised loss. Concretely, at earlier training stages, the adaptive weight will be small, due to the abundant low-confidence pixels. Our model is mainly learned with labeled images. Then as training proceeds, the adaptive weight will gradually increase to learn more on unlabeled images. You may also refer to the Equation 2 in FixMatch.

2. I think that either 0.5 or 0.95 is just a hyperparameter, regardless of the number of classes. In your case of binary classification, I guess you may use a sigmoid function for classification. Then you can still maintain a 0.95 threshold by only considering the output > 0.95 (certainly class 1) and the output < 0.05 (certainly class 0). The 0.5 you mentioned probably denotes the classification threshold for sigmoid, rather than our confidence threshold.

3. I think the unsupervised loss weight can be adjusted according to your observations. It may depend on the proportion and hardness of unlabeled images.