swapped_prediction computation about uno

[DonkeyShot21]

I have just merged a pull request that fixes this bug. Closing. Thanks @DeepTecher

[DonkeyShot21]

Yes, you are right, it is a small typo, I changed the code multiple times and I forgot to remove the loop. Anyway, the swapped assignment is performed in the same way, since I only index the view axis inside the swapped_assignment method, and the cross-entropy loss compares tensors element-wise. Also, the losses are then averaged, so nothing should change.

[DeepTecher]

yeah...
But it still has a problem on cross_entropy_loss function.
if I guess right, the dims of preds is [num_views, bach_size, num_label+num_unlabel], our F.log_softmax should be

preds = F.log_softmax(preds / 0.1, dim=2)

on the last dim to do log_softmax.
we do dim=1 will work on batch_size. Is it right?

[DonkeyShot21]

Yes, you are right. For some reason, this still works. Let me look into it.

[DeepTecher]

Ok.
if you have a new conclusion, please let me know
many thanks

[DonkeyShot21]

Hi, I fixed it and ran CIFAR100-20. I got similar results for CIFAR100-20 on the test set, while performance is slightly worse on the training set. I am now trying to do some hyperparameter tuning. I'll upload the fix as soon as possible.

[DeepTecher]

nice~ 👍

[DonkeyShot21]

Hi, I have some good news. It seems that normalizing on the correct dimension improves performance quite significantly. I needed to tune the parameters a bit, but I just had one run hit 55% on the training set (unlab/train/acc) and 56% on the test set (unlab/test/acc) for CIFAR100-50. I am testing if the same parameters work on the other settings.

I also went back to my logs and found that ImageNet experiments were probably run without the bug, while all other datasets were affected. I will upload the new version as soon as I finish running experiments.

[DeepTecher]

okay,
I cannot wait for the newest result.