When training, I meet the problem of "NaN or Inf found in input tensor". Is it a mistake? about uniot-for-unida

[changwxx]

Can you give a more detailed description about the problem you encountered? (e.g. which part of the code would output like that? Which transfer task did you set?)

[JunkunPeng17]

I met the same situation as xzm2002 when I did the transfer task on Product2Art. I also did the other transfer tasks, and they worked well.

[changwxx]

I would like to thank @JunkunPeng17 for providing details for this bug!
I found sometimes the loss_CCD would be NaN for OfficeHome transfer tasks, since CCD did not find any top-confident common target samples and high_conf_label_id is a 0-size tensor.
I have fixed this bug in main.py

UniOT-for-UniDA/main.py

Lines 170 to 176 in 9ba3bad

	# fix the bug raised in https://github.com/changwxx/UniOT-for-UniDA/issues/1
	# Due to mini-batch sampling, current mini-batch samples might be all target-private.
	# (especially when target-private samples dominate target domain, e.g. OfficeHome)
	if high_conf_label_id.size(0) > 0:
	loss_CCD = criterion(after_lincls_t[high_conf_label_id,:], high_conf_label[high_conf_label_id])
	else:
	loss_CCD = 0

However, this bug would not lead to performance deterioration, and hence the results reported in the paper are still reliable.