In this repo, I implemented ResNet for CIFAR10 as described in the original paper.
I load and normalize CIFAR10 training and test datasets using torchvision. The training dataset contains 50K images and the test dataset contains 10K images.
I compare ResNet20 and ResNet32 with 20-layer and 44-layer plain networks on CIFAR10. As you can see in the plain network with increasing layers the performance dropped while in ResNet with increasing the layer, the network outperforms.
| Plain network: test error on CIFAR10, the deeper network has higher error | ResNet: train error on CIFAR10, the deeper network has less error |
|---|---|
 |
 |