Regarding ResNet structure, in http://www.pabloruizruiz10.com/resources/CNNs/ResNet-PyTorch.html
It is mentioned:
For instance, ResNet on the paper is mainly explained for ImageNet dataset… since CIFAR10 input images are (32x32) instead of (224x224), the structure of the ResNets need to be modified.
while I saw some Pytorch codes on GitHub for Resnet with CIFAR10, that are used the same structure as ResNet for ImageNet, I mean four-layers by “Conv64, Conv128, Conv256, Conv512”. Would you please guide me for Resnet CIFAR10, the structure of ResNet needs to be modified or not?
ResNet for cifar has less layers per block if i’m not wrong. I don’t remember about convolution size but maybe the first 7x7 convolution may be too big.
Thank you for your reply @JuanFMontesinos. I saw 3x3 convolution is used in CIFAR cases. This means by changing the filter size, the same structure can be used for CIFAR10?
Yep, the structure is the same (wrt to what i remember).
I think they were using 16 32 64 128 filters instead. It’s also remarkable the depth you use.
This means the structure is a little bit different by ResNet for ImageNet, and 3x3convolution with the number of filters 16 32 64 128 should be used for CIFAR10? I also noticed in the “Deep Residual Learning for Image Recognition” paper the numbers of filters are considered 16 32 64.
The main problem is that if you use imagenet filters it will probably overfit so hard. That’s why CIFAR uses less. Anyway everything is on the original papers you don’t really need to discuss so much.