The 2017 ImageNet winner, rebuilt for PyTorch

What it does

This repo implements SENet—the architecture that won the ILSVRC 2017 classification competition—in PyTorch. It covers SE-ResNet variants (18, 34, 50, 101, 152, 20, 32) and SE-Inception-v3, and includes training scripts for CIFAR-10 and ImageNet. A pretrained SE-ResNet50 checkpoint is available, and several models can be loaded via torch.hub.

The interesting bit

The README frames SENet as a surgical upgrade to standard backbones rather than a new architecture family. The reported tables show SE-ResNet20 beating plain ResNet20 by one percentage point on CIFAR-10, and SE-ResNet50 improving ImageNet top-1 by roughly a point over the torchvision ResNet-50 baseline. The implication is that SE blocks are a low-friction way to nudge familiar models upward.

Key highlights

• SE-ResNet20 hits 93% on CIFAR-10, compared to 92% for plain ResNet20.
• SE-ResNet50 reaches 77.06% ImageNet top-1 accuracy when trained with the distributed script on 8 GPUs; the author provides a downloadable weight file.
• SE-ResNet20, 56, 50, and 101 are accessible via torch.hub, including a pretrained SE-ResNet50.
• The ImageNet script supports distributed training across multiple GPUs.
• Training scripts depend on the external homura library.

Caveats

• The author warns that initial learning rates and batch sizes differ from the original paper due to limited computational resources, so exact reproduction is not guaranteed.
• The author explicitly states they cannot maintain the repository actively.

Verdict

A solid reference implementation if you want pretrained SENet weights or torch.hub convenience in PyTorch. Look elsewhere if you need active upkeep or a faithful reproduction of the original training recipe.