classification with freeze_ and unfreeze_

[yongjin-shin]

Hi,
First of all, I very much appreciate your wonderful work!
Currently, I am testing CIFAR10 classification task with your example code.
I would like to ask whether 10% of the test performance is an usual case when I call freeze_() in the test time.

# Test Time
    with torch.no_grad():
        for data in test_loader:
            images, labels = data
            classifier.freeze_()
            outputs = classifier(images.to(device))
            _, predicted = torch.max(outputs.data, 1)
            total += labels.size(0)
            correct += (predicted == labels.to(device)).sum().item()
    print(f'Freeze Epoch {epoch} | {str(100 * correct / total)}% | Elpased: {time.time() - tic:.1f}s')
    classifier.unfreeze_()

I only add classifier.freeze_() before getting outputs.
I thought the accuracy should be similar with unfreeeze_(), however, it seems not.
When I activated freezemode, I got 10%, but unfreeze mode reach 45% at the first epoch.
Since I only activate it at the test time, the training loss keeps going down.

Best Regards,
YJ

[piEsposito]

This is not usual and should be fixed by now.

[jinqijinqi]

The problem is still there.