Reproduce “Fixing the train-test resolution discrepancy: FixEfficientNet” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Hugo Touvron, Andrea Vedaldi, Matthijs Douze, Hervé Jégou
The paper demonstrates that applying the computationally efficient FixRes method to EfficientNet architectures consistently yields significant performance improvements on image classification benchmarks. FixEfficientNet-L2 achieved a new state-of-the-art of 88.5% Top-1 accuracy on ImageNet, alongside robust gains on ImageNet-v2 and ImageNet Real Labels, indicating improved generalization.
Key insights
- Correcting the train-test resolution discrepancy with FixRes consistently boosts performance across all EfficientNet variants, providing a simple yet powerful enhancement.
- FixRes improves the generalization capabilities of models, as evidenced by sustained performance gains on the ImageNet-v2 dataset, which is designed to identify overfitting.
- The method is computationally efficient and architecture-agnostic, making it a practical and accessible improvement for virtually any CNN classification model.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.