Possible bug in meanteacher algorithm

[5v3D3]

In meanteacher train_step (line 55), logits_x_ulb_s gets assigned the logits of outs_x_ulb_w.
So the consistency loss is calculated only with logits_x_ulb_w, which results in 0 consistency loss in every iteration.
outs_x_ulb_s is never used.

Semi-supervised-learning/semilearn/algorithms/meanteacher/meanteacher.py

Lines 54 to 56 in 9a24e00

	outs_x_ulb_s = self.model(x_ulb_s)
	logits_x_ulb_s = outs_x_ulb_w['logits']
	feats_x_ulb_s = outs_x_ulb_w['feat']

this should be:

outs_x_ulb_s = self.model(x_ulb_s) 
logits_x_ulb_s = outs_x_ulb_s['logits'] 
feats_x_ulb_s = outs_x_ulb_s['feat']

[Hhhhhhao]

This might be caused by a typo. Have you run the modified code? Does the result vary from the reported result?

[5v3D3]

Sorry but i haven't run the modified code on any of the benchmarks, since right now I am only working locally with my own dataset. With the modified code I get an unsup_loss that is greater than 0 though. So I can only suspect that it would perform better.
But I had a look at your results and meanteacher performance looked a bit odd as it is quite similar to supervised (Both in classic_cv and usb_cv), while most other methods perform better . Which would make sense, if only the supervised loss is used in optimization.

[Hhhhhhao]

Sorry but i haven't run the modified code on any of the benchmarks, since right now I am only working locally with my own dataset. With the modified code I get an unsup_loss that is greater than 0 though. So I can only suspect that it would perform better. But I had a look at your results and meanteacher performance looked a bit odd as it is quite similar to supervised (Both in classic_cv and usb_cv), while most other methods perform better . Which would make sense, if only the supervised loss is used in optimization.

Will check this in next update.

[Hhhhhhao]

Fixed in PR #135