LAASP
Representative pruned PyTorch checkpoints for Loss-Aware Automatic Selection of Structured Pruning Criteria for Deep Neural Network Acceleration by Deepak Ghimire, Kilho Lee, and Seong-heum Kim. Published in Image and Vision Computing, volume 136, article 104745 (2023).
- arXiv preprint
- Journal article
- Source code, results, and evaluation scripts
- GitHub checkpoint release
Checkpoints
Each row provides one representative checkpoint. FLOPs reduction labels follow the GitHub results table, where CIFAR-10 accuracy values summarize multiple runs.
| Dataset | Architecture | FLOPs reduction | Download |
|---|---|---|---|
| CIFAR-10 | VGG16 | 34.6% | cifar10-vgg16-flops-rr-34.6.pth.tar |
| CIFAR-10 | VGG16 | 60.5% | cifar10-vgg16-flops-rr-60.5.pth.tar |
| CIFAR-10 | ResNet32 | 53.3% | cifar10-resnet32-flops-rr-53.3.pth.tar |
| CIFAR-10 | ResNet56 | 52.6% | cifar10-resnet56-flops-rr-52.6.pth.tar |
| CIFAR-10 | ResNet110 | 52.5% | cifar10-resnet110-flops-rr-52.5.pth.tar |
| CIFAR-10 | ResNet110 | 58.5% | cifar10-resnet110-flops-rr-58.5.pth.tar |
| ImageNet | ResNet18 | 42.2% | imagenet-resnet18-flops-rr-42.2.pth.tar |
| ImageNet | ResNet18 | 45.4% | imagenet-resnet18-flops-rr-45.4.pth.tar |
| ImageNet | ResNet34 | 41.4% | imagenet-resnet34-flops-rr-41.4.pth.tar |
| ImageNet | ResNet34 | 45.4% | imagenet-resnet34-flops-rr-45.4.pth.tar |
| ImageNet | ResNet50 | 42.3% | imagenet-resnet50-flops-rr-42.3.pth.tar |
| ImageNet | ResNet50 | 53.9% | imagenet-resnet50-flops-rr-53.9.pth.tar |
| ImageNet | MobileNetV2 | 30.0% | imagenet-mobilenetv2-flops-rr-30.0.pth.tar |
| ImageNet | MobileNetV2 | 54.5% | imagenet-mobilenetv2-flops-rr-54.5.pth.tar |
Download and evaluate
Install the Hub client with pip install huggingface_hub. For example, download the CIFAR-10 VGG16 checkpoint with the 34.6% FLOPs reduction label.
from huggingface_hub import hf_hub_download
path = hf_hub_download(
repo_id="ghimiredhikura/LAASP",
filename="CIFAR-10/cifar10-vgg16-flops-rr-34.6.pth.tar",
)
print(path)
Use the downloaded path as --pruned_path in the matching GitHub evaluation script. The project documents Python 3.9.7, PyTorch 1.10.2, and torchvision 0.11.2. The checkpoints use the original PyTorch serialization format. Some contain serialized model objects and require the original model classes from the project code.
Citation
@article{ghimire2023laasp,
title={Loss-aware automatic selection of structured pruning criteria for deep neural network acceleration},
author={Ghimire, Deepak and Lee, Kilho and Kim, Seong-heum},
journal={Image and Vision Computing},
volume={136},
pages={104745},
year={2023},
doi={10.1016/j.imavis.2023.104745},
eprint={2506.20152},
archivePrefix={arXiv}
}