ResNets for CIFAR that match the paper, not ImageNet leftovers

What it does

Provides PyTorch implementations of ResNet-20, 32, 44, 56, 110, and 1202 specifically for CIFAR-10/100, matching the layer counts and parameter budgets described in the original 2015 paper. It also distributes pretrained checkpoints whose test errors meet or beat the paper’s reported figures.

The interesting bit

Most PyTorch ResNet ports target ImageNet, so dropping them onto 32×32 CIFAR images quietly alters the network depth and parameter count. This repo is essentially a pedantic but necessary correction, ensuring that when you compare against He et al., you are actually comparing the same model.

Key highlights

• Pretrained models for six depths, from 0.27M-parameter ResNet20 up to 19.4M-parameter ResNet1202.
• Reported test errors are lower than the paper’s across the board (e.g., ResNet110 hits 6.32% vs. the paper’s 6.43%).
• Training logic is intentionally straightforward, without extra validation splits or model selection tricks.
• ResNet1202 training requires a GPU with 16GB of memory.

Caveats

• The original paper selects the best model using a 45k/5k train/validation split; this implementation skips that validation step, so direct head-to-head comparisons should account for the difference.
• Training ResNet1202 demands 16GB of VRAM, which may limit hardware options.

Verdict

Grab this if you need a faithful CIFAR baseline for ablation studies or paper reproductions. Skip it if you just need any ResNet that works on small images and do not care about matching the canonical architecture.