Torchvision model zoo provides number of implementations of various state-of-the-art architectures, however, most of them are defined and implemented for ImageNet. Usage of these backbones with small input sizes (as in CIFAR) is not trivial.
This repo provides training and evaluation scripts for CIFAR100 with ResNet backbones. 60% accuracy can be obtained with default training parameters.
Check the notebook for a demo of the submodules used in this repo.
- Clone:
git clone https://github.com/fcakyon/cifar100-resnet.git- Prepare conda environment:
conda env create -f environment.ymlconda activate cifarresnet- Train:
python main.py --arch resnet32 --save-dir checkpoints/- Evaluate:
python main.py --evaluate checkpoints/resnet32_final.th