Regarding transforms about trojanzoo

[ain-soph]

Close this issue if you have no further concern.

[ain-soph]

Actually I never test MNIST, let me know if any issue exists. The transform for MNIST is a simple ToTensor
And note that the normalization parameters are embedded into model as a layer rather than dataset transform. So that all image tensors range in [0, 1]

trojanzoo/trojanvision/datasets/imageset.py

Lines 56 to 61 in e38f90b

	elif self.data_shape in ([3, 16, 16], [3, 32, 32]):
	transform = get_transform_cifar(mode, auto_augment=self.auto_augment,
	cutout=self.cutout, cutout_length=self.cutout_length,
	data_shape=self.data_shape)
	else:
	transform = transforms.ToTensor()

trojanzoo/trojanvision/datasets/imageset.py

Lines 141 to 156 in e38f90b

	def get_transform_cifar(mode: str, auto_augment: bool = False,
	cutout: bool = False, cutout_length: int = None,
	data_shape: list[int] = [3, 32, 32]) -> transforms.Compose:
	if mode != 'train':
	return transforms.ToTensor()
	cutout_length = data_shape[-1] // 2 if cutout_length is None else cutout_length
	transform_list = [
	transforms.RandomCrop(data_shape[-2:], padding=data_shape[-1] // 8),
	transforms.RandomHorizontalFlip(),
	]
	if auto_augment:
	transform_list.append(transforms.AutoAugment(transforms.AutoAugmentPolicy.CIFAR10))
	transform_list.append(transforms.ToTensor())
	if cutout:
	transform_list.append(Cutout(cutout_length))
	return transforms.Compose(transform_list)

[ain-soph]

CUDA_VISIBLE_DEVICES=0 python examples/train.py --verbose 1 --color --epoch 600 --batch_size 96 --cutout --grad_clip 5.0 --lr 0.025 --lr_scheduler --save --dataset cifar10 --model resnet18_comp

And you will get ResNet18 (first convolutional layer compressed version) with 96.5% accuracy.

[agSidharth]

ya so I was testing these models manually and the transform that worked for me is

transform.Compose([transforms.ToTensor(),transforms.Normalize([0.49139968, 0.48215827, 0.44653124],[0.24703233, 0.24348505, 0.26158768]))

To give the expected accuracy.

[ain-soph]

I remembered that Without data augment such as random crop and cutout, the model accuracy won’t exceed 92%. But I could be wrong.

And if you use my model class and put normalization into the transform, you’d better set the model transform layer mean/std to be 0/1.