Testing with the best model

[LeeDoYup]

I found that cifar.py code does not test the performance with the best model.

augmix/cifar.py

Line 431 in 7c84885

test_c_acc = test_c(net, test_data, base_c_path)

In addition, why the validation set of CIFAR is not used in the model selection ??

[cygerts]

I think they choose the model from 100th epoch.
From the paper: "The All Convolutional Network and Wide ResNet train for 100 epochs,
and the DenseNet and ResNeXt require 200 epochs for convergence"

[LeeDoYup]

I think it's much unfair, because the code evenly save the best model.

augmix/cifar.py

Lines 413 to 414 in 7c84885

	if is_best:
	shutil.copyfile(save_path, os.path.join(args.save, 'model_best.pth.tar'))

We can just use the best model in evaluation :) !!

[LeeDoYup]

I also checked, the selection of the best model doesn't change augmix's performances on test_c.

[normster]

Hi @LeeDoYup, there isn't much of a difference between the best and latest models so we just evaluate on the latest model checkpoint.