Flatness Improves Backbone Generalisation in Few-shot Classification

Implementation for the following paper: Rui Li, Martin Trapp, Marcus Klasson, and Arno Solin (2025). Flatness Improves Backbone Generalisation in Few-shot Classification. In Winter Conference on Applications of Computer Vision (WACV)

We introduce a simple yet effective training protocol for the backbone in few-shot classification. We show that flatness-aware backbone training combined with vanilla fine-tuning results in a simpler yet competitive baseline compared to the state-of-the-art. We present theoretical and empirical results indicating that careful backbone training is crucial in FSC.

Average test accuracy on the Meta-Dataset benchmark for different backbone training under the same adaptation: empirical risk minimisation (ERM) without information fusion, with fine-tuning, or with knowledge distillation; sharpness-aware minimisation (SAM) without information fusion or with fine-tuning.

Dependencies

This code requires the following:

• PyTorch 1.13.1
• TensorFlow 2.8.1

Usage

• Clone or download this repository.
• Setup Meta-Dataset:
  • Follow the the "User instructions" in the Meta-Dataset repository for "Installation" and "Downloading and converting datasets".
• After setting up Meta-Dataset, backbone can be trained with SAM using train_vanilla_sam.py
  • Change line 54 to load MetaDatasetEpisodeReader, MetaDatasetBatchReader.
• To select backbone for evaluation, run select_backbone.py.
  • Save the trained backbones in saved_model/sam/{dataset}.pth or change line 22.
  • Change line 12 to load MetaDatasetEpisodeReader.

Acknowledge

We thank authors of Meta-Dataset, PARC, SAM and SUR for their source code.