using Python 3.7.6 and PyTorch 1.4.0 loss is always NaN

[skariel]

Haven't tested on Python 2

[HobbitLong]

My environment:

Python 3.6.9 (default, Nov  7 2019, 10:44:02)
>>> torch.__version__
'1.2.0'

[HobbitLong]

I will close for now, feel free to reopen

[BestJuly]

I also met the same problem in one server (Python 3.7.4, pytorch 1.2.0) and did not have this problem in another server (Python 3.7.4, pytorch 1.3.0). I think this may be caused by different settings of environment (but why?).
A strange thing is that for the first iteration, without calculating loss, some extracted features are nan. I have not figured out why and how to solve. And now I am running on the server which is fine. One solution may be setting the same experimental environment.

[HobbitLong]

@BestJuly ,

I added some fix 4 days ago to improve the stability. Can you try again now?

[BestJuly]

@HobbitLong Thank you for your kindness. The problem vanished several days ago even though I do not know why. When I using the newest version of your code, it also runs well.

By the way, I think for the newest version, .cuda() is missing in main_supcon.py.

[HobbitLong]

@BestJuly,

Ah, nice catch! Just pushed a fix. Thanks for spotting this!

[shijianjian]

To me, the problem lies in:

exp_logits = torch.exp(logits) * logits_mask
log_prob = logits - torch.log(exp_logits.sum(1, keepdim=True))

my exp_logits runs into 0 occasionally which causing undefined log results. Thus, I did a simple adding with 1e-20.

[shaunakjoshi12]

To me, the problem lies in:

exp_logits = torch.exp(logits) * logits_mask
log_prob = logits - torch.log(exp_logits.sum(1, keepdim=True))

my exp_logits runs into 0 occasionally which causing undefined log results. Thus, I did a simple adding with 1e-20.

Thanks man you saved my day!

[HenryPengZou]

One method to avoid exp_logits becoming 0 or having loss = Nan is to normalize your feature vectors before passing it the loss function. Hope it helps you guys!