CIFAR-pretrained-models

Accuracy in the Validation Set

The validation is performed with the original view of the image(size=32x32).

Note: the FLOPs and the number of parameters only counts the conv and linear layer.

Model	Acc@1(C10)	Acc@5(C10)	Acc@1(C100)	Acc@5(C100)	#param.	FLOPs
Resnet20[1]	91.65	99.68	66.61	89.95	0.27M	40.81M
Resnet32[1]	92.81	99.72	68.74	90.23	0.46M	69.12M
Resnet44[1]	93.24	99.75	69.49	90.39	0.66M	97.44M
Resnet56[1]	93.69	99.68	70.79	91.10	0.85M	125.75M

Pretrained Models

All the pretrained models are avaliable in the release.

Implementation Details

The models are trained and exported with PyTorch(1.1.0) and torchvision(0.2.2).

The training data augumentation follow [1],

torchvision.transforms.Compose([
    torchvision.transforms.RandomCrop(size=32, padding=4),
    torchvision.transforms.RandomHorizontalFlip(),
    torchvision.transforms.ToTensor(),
    torchvision.transforms.Normalize(
        mean=[0.4914, 0.4822, 0.4465], # mean=[0.5071, 0.4865, 0.4409] for cifar100
        std=[0.2023, 0.1994, 0.2010], # std=[0.2009, 0.1984, 0.2023] for cifar100
    ),
])

All the models are trained with a mini batch size of 256 and the following optimizer,

torch.optim.SGD(lr=0.1, momentum=0.9, dampening=0, weight_decay=1e-4, nesterov=True)

the following scheduler,

ctx.scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(T_max=200,eta_min=0.001)

the total training epochs is 200.

Reference

Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun. Deep residual learning for image recognition. In CVPR, 2016.

Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun. Identity Mappings in Deep Residual Networks. In ECCV, 2016.

Acknowledgement

Thanks for the computer vision community and github open source community.

TLESORT / pytorch-cifar-models