v1.5.0

Incorporating Few-Shot Learning best practices into EasyFSL!

Feature centering and normalization

Centering and normalizing features after passing through the backbone but before inference improves performance by a couple of percent with almost every method and on all benchmarks, so we added the option to extensions of the FewShotClassifier base class.

classifier = PrototypicalNetworks(
	backbone=my_backbone,
	feature_centering=average_base_features,
	feature_normalization=2,
)

Hyperparameter search

We added the tools to perform hyperparameter selection in scripts/hyperparameter_search.py. You can now launch your own grid search, or reproduce ours!

Improved and extended results

We ran a new benchmark on _mini_ImageNet and _tiered_ImageNet with every available method using the best hyperparameters selected on _mini_ImageNet's validation set. You now have everything you need to select the best method for your use case!

v1.4.0

🤯 Huge release including 4 new methods, new standardized backbones and modules, improvements to existing methods and some datasets API, and numbered reproducible benchmarks on miniImageNet and tieredImageNet!

Incredible New Features

• Add 4 new methods: SimpleShot, FEAT, LaplacianShot, PT-MAP
• Add new modules:
  • MultiHeadAttention: default attention module used in FEAT
  • feat_resnet12: default backbone used in FEAT
  • utility feat_resnet12_from_checkpoint to load pre-trained weights for feat_resnet12
• Add scripts to evaluate methods:
  • predict_embeddings to extract all embeddings from a dataset using a pre-trained backbone
  • benchmark_methods to evaluate a few-shot method on a dataset of pre-extracted features
  • config files to ease reproduction
• Add result tables to README for miniImageNet and tieredImageNet
• Add utilities:
  • Add a torch implementation of k_nearest_neighbours()
  • Add a strip_prefix() method that removes a specified prefix from the keys of an OrderectDict (can be used for torch state dicts)

Minor changes

• TIM now uses cosine distance instead of L2 distance
• Add parameterization of the softmax temperature in TransductiveFinetuning, Finetune, and TIM
• Add some non-breaking changes to the datasets:
  • Add option image_file_extension to facilitate the use of the small version of the DanishFungi dataset
  • Make CUB and TieredImageNet extensions of EasySet instead of methods returning an instance of EasySet to improve typing consistency between datasets

Next steps

This release is a huge step because it's the first time we commit on reproducible evaluations of the Few-Shot Learning methods in EasyFSL. We still need to improve and augment these benchmarks.

1. Complete benchmark with Matching and Relation Nets, PT-MAP, Transductive Finetuning (est. July 23)
2. Add explicit hyperparameter selection (est. July 23)
3. Add feature normalization as it's been proven to have a huge impact on the results (est. July 23)
4. Add cross-domain benchmarks (CUB, Fungi) and using other backbones (est. September 23)

Any help is welcome!

v1.3.0

New minor release with great tools to help you load all the data you want!

🥚 What's New

• Add FeaturesDataset (#100): you can now use Few-Shot Learning models directly on embeddings!
  • FeaturesDataset can be initialized from embeddings in various forms (directly from tensors, from a dataframe or a dictionary)
  • FewShotClassifier models default is now to be initialized with a nn.Identity backbone: if you don't specify a backbone, it is equipped to work directly on features
  • New notebook to help understand how to use these new tools
• Add WrapFewShotDataset (#101): any dataset can become a FewShotDataset and be used with EasyFSL
• Added Python 3.11 support and removed Python 3.6 support (since this version is in end-of-life) (#99)
• Tensor shapes are now explicit in docstrings (#102)

🗒️ Full Changelog: v1.2.1...v1.3.0

v1.2.1

🔧 Quickfix: make typing in TaskSampler compatible with MyPy.

v1.2.0

🎉 EasyFSL is now fully tested for Python 3.10!

There were also a lot of issues before related to TaskSampler being too permissive, causing downstream errors that were hard to understand. So now you get an explicit error with TaskSampler if you:

• initialize it with a dataset that's too small (number of classes smaller than n_way, or number of instances in a class smaller than n_shot + n_query)
• call TaskSampler.episodic_collate_fn() with input data that is not a (Tensor, int) or (Tensor, Tensor[0-dim int]) tuple.

Thanks to @Aml-Hassan-Abd-El-hamid for their good will and contribution.

v1.1.0

You can now create a support set from a file structure and easily feed it to Few-Shot Learning methods with SupportSetFolder. Thanks @diego91964 !

v1.0.1

There were some things to fix after the v1 release, so we fixed them:

• EasySet's format check is now case unsensitive (thanks @mgmalana 😄 )
• TaskSampler used to yield torch.Tensor objects which caused errors. So now it yields lists of integers, as is standard in PyTorch's interface.
• When EasySet's initialization didn't find any images in the specified folders, it just built an empty dataset with no warning, which caused silent errors. Now EasySet.__init__() raises the following warning if no image is found: "No images found in the specified directories. The dataset will be empty"

v1.0.0

🎂 Exactly 1 year after the first release of Easy FSL, we have one more year of experience in Few-Shot Learning research. We capitalize on this experience to make Easy FSL easier, cleaner, smarter.

No more episodic training logic inside Few-Shot Learning methods: you can train them however you want.
And more content! 4 additional methods; several ResNet architecture as they're often used in FSL research; and 4 ready-to-use datasets.

🗞️ What's New

• Few-Shot Learning methods
  • Finetune
  • Transductive Finetuning
  • BD-CSPN
  • Transductive Information Maximization
• Pre-designed ResNet architecutres for Few-Shot Learning
• Most common few-shot classification datasets
  • _tiered_ImageNet
  • _mini_ImageNet
  • CU-Birds
  • Danish Fungi (not common but new, and really great)
  • And also an abstract class FewShotDataset to ease your developement or novel or modified datasets
• Example notebooks to perform both episodic training and classical training for your Few-Shot Learning methods
• Support Python 3.9

🔩 What's Changed

• AbstractMetaLearner is renamed FewShotClassifier. All the episodic training logic has been removed from this class and moved to the example notebook episodic_training.ipynb
• FewShotClassifier now supports non-cuda devices
• FewShotClassifier can now be initialized with a backbone on GPU
• Relation module in RelationNetworks can now be parameterized
• Same for embedding modules in Matching Networks
• Same for image preprocessing in pre-designed datasets like EasySet
• EasySet now only collects image files

Full Changelog: v0.2.2...v1.0.0

v0.2.2

Small fixes in EasySet and AbstractMetaLearner

• Sort data instances for each class in EasySet

• Add EasySet.number_of_classes()

• Fix best validation accuracy update

• Move switch to train mode inside fit_on_task()

• Make AbstractMetaLearner.fit() return average loss

v0.2.1

We fixed a bug that caused validation to not occur during training.