hongyi-zhang/mixup

Can't reproduce the results with sample run provided for CIFAR-10

feat7 opened this issue · 0 comments

feat7 commented

I ran the command as provided in readme:

python easy_mixup.py --sess my_session_1 --seed 11111

Since, I am using pytorch >= 1.0
I had to fix the loss calculation:

train_loss += loss.data[0] to train_loss += loss.data.item()
and test_loss += loss.data[0] to test_loss += loss.data.item()

The maximum accuracy I got was 93.90% which translates to an error of 6.1

What am I doing wrong here?