Your method of computing accuracies seems not correct

[majinyu666]

First of all, thanks for sharing code. I haven't totally understood your way of computing accuracies in evaluate() in train.py. It seems you are using some kind of moving average of accuracies of test batches. I believe thats not correct. It should be computed as num_correct/num_total. Here's what I think, in detail:

    # evaluating
    correct = 0
    total = 0
    with torch.no_grad():
        for target_data, target_label in loader:
            target_data = Variable(target_data.to(device=args.device))
            target_label = Variable(target_label.to(device=args.device))

            out = model(target_data)
            predicted = out.argmax(1)
            correct += torch.sum(predicted==target_label).item()
            total += target_data.size(0)
    accuracy = correct / total

I compared this with your way of computing accuracies, it turns out that your way always computes higher accuracies. Could you please think about it?

I tried to implement deep CORAL by myself and only reach ~0.55 on A->W. Also, another github project got no more than that. That's far worse than the original paper (they got 66.4%). TAT