OCSPruner
Pruned image-classification checkpoints for One-Cycle Structured Pruning via Stability-Driven Subnetwork Search (WACV 2026).
Checkpoints
FLOPs reduction labels come from the original checkpoint names.
| Dataset | Architecture | FLOPs reduction | Download |
|---|---|---|---|
| CIFAR-10 | ResNet110 | 33% | cifar-10-resnet110-flops-33.pth |
| CIFAR-10 | ResNet110 | 47% | cifar-10-resnet110-flops-47.pth |
| CIFAR-10 | ResNet56 | 39% | cifar-10-resnet56-flops-39.pth |
| CIFAR-10 | ResNet56 | 47% | cifar-10-resnet56-flops-47.pth |
| CIFAR-10 | VGG16 | 20% | cifar-10-vgg16-flops-20.pth |
| CIFAR-10 | VGG16 | 26% | cifar-10-vgg16-flops-26.pth |
| CIFAR-100 | VGG19 | 11% | cifar-100-vgg19-flops-11.pth |
| ImageNet | ResNet18 | 45% | imagenet-resnet18-flops-45.pth |
| ImageNet | ResNet50 | 30% | imagenet-resnet50-flops-30.pth |
| ImageNet | ResNet50 | 43% | imagenet-resnet50-flops-43.pth |
| ImageNet | ResNet50 | 48% | imagenet-resnet50-flops-48.pth |
| ImageNet | ResNet50 | 67% | imagenet-resnet50-flops-67.pth |
Download and evaluate
Clone the source repository and run the commands from its root. The project documents Python 3.11.3, PyTorch 2.0.1, and torchvision 0.15.2 as its experimental environment. Install the Hub client with pip install huggingface_hub.
Download a checkpoint from Python.
from huggingface_hub import hf_hub_download
path = hf_hub_download(
repo_id="ghimiredhikura/OCSPruner",
filename="CIFAR-10-100/cifar-10-resnet56-flops-47.pth",
)
print(path)
Or use the CLI to download and evaluate the CIFAR-10 ResNet56 checkpoint.
hf download ghimiredhikura/OCSPruner CIFAR-10-100/cifar-10-resnet56-flops-47.pth --local-dir checkpoints
python main_cifar.py --mode eval --model resnet56 --dataset cifar10 \
--model-path checkpoints/CIFAR-10-100/cifar-10-resnet56-flops-47.pth \
--output-dir eval_result_cifar
For ImageNet, provide the ILSVRC 2012 directory as described in the project README.
hf download ghimiredhikura/OCSPruner ImageNet/imagenet-resnet50-flops-67.pth --local-dir checkpoints
python main_imagenet.py --eval --model resnet50 --data-path /path/to/imagenet \
--model-path checkpoints/ImageNet/imagenet-resnet50-flops-67.pth \
--output-dir eval_result_imagenet
These original PyTorch .pth files contain serialized model objects. Use the project code to load the CIFAR checkpoints, which reference its custom model classes. On PyTorch 2.6 or later, the evaluation scripts may require TORCH_FORCE_NO_WEIGHTS_ONLY_LOAD=1 for trusted checkpoints when using their existing torch.load calls.
Citation
@inproceedings{ghimire2026ocspruner,
title={One-Cycle Structured Pruning via Stability-Driven Subnetwork Search},
author={Ghimire, Deepak and Kil, Dayoung and Jeong, Seonghwan and Park, Jaesik and Kim, Seong-heum},
booktitle={Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
year={2026}
}