Overview

A modular library for few-shot learning, implementing classic and state-of-the-art meta-learning algorithms

It's designed to be flexible, extensible, and easy to integrate into your research or production pipeline

• Plug-and-play support for multiple few-shot learning models
• Episodic training and evaluation setup
• Clean dataset interface for custom datasets
• Support for custom backbones like ResNet, ViT, and more
• Built-in data augmentation options (basic, mixup, cutmix)

Available Models

The following model results are based on the original papers, evaluated on Mini-Imagenet, using 5-way 5-shot tasks.

Model	Paper Reference	Input parameter name	Mini-imagenet accuracy (%)
ProtoNet	https://arxiv.org/abs/1703.05175	proto_net	68.20
RelationNet	https://arxiv.org/abs/1711.06025	relation_net	65.32
MatchingNet	https://arxiv.org/abs/1606.04080	matching_net	60.00
MetaOptNet	https://arxiv.org/abs/1904.03758	metaopt_net	78.63
TapNet	https://arxiv.org/abs/1905.06549	tapnet	76.36
TADAM	https://arxiv.org/abs/1805.10123	tadam	76.70
DN4	https://arxiv.org/abs/1903.12290	dn4	71.02
MSENet	https://arxiv.org/abs/2409.07989v2	msenet	84.42
FEAT	https://arxiv.org/abs/1812.03664	feat	82.05
DSN	https://ieeexplore.ieee.org/document/9157772	dsn	78.83
METAQDA	https://arxiv.org/abs/2101.02833	metaqda	84.28
Negative Margin	https://arxiv.org/abs/2003.12060	negativemargin	81.57
R2D2	https://arxiv.org/abs/1805.08136	r2d2	68.40
MAML	https://arxiv.org/abs/1703.03400	maml	63.10
MTL	https://arxiv.org/abs/1812.02391	mtl	74.30

All models are implemented as separate .py files under the models/ directory

It is possible to use ResNet and Swin as backbones for the models. The respective input parameters are resnet and swin.

Getting Started

Install the package directly from GitHub:

pip install git+https://github.com/victor-nasc/SimpleFewShot.git

A ready-to-run Colab example is available:

Custom Dataset Format

To use a custom dataset, provide a .csv file with the following structure:

• image_id: Path to the image file
• One column per class (binary: 1 if image belongs to class, 0 otherwise)

Example:

idx	image_id	class1	class2	class3
1	/data/img1.jpg	1	0	0
2	/data/img2.jpg	0	1	0
3	/data/img3.jpg	0	0	1
...	...	...	...	...

You can define your training and test class splits programmatically using the FewShotManager class

Obs.: This method supports also multi-labeled datasets

Training a model

In order to train a model with your data, run:

python3 train.py \
  --model <model_name> \
  --backbone <backbone_name> \
  --ways <number_of_classes_per_task> \
  --shots <number_of_examples_per_class_per_task> \
  --gpu <which_gpu_to_run> \
  --lr <lr_value> \
  --l2_weight <l2_weight_value> \
  --augment <cutmix/mixup/None>

Model name is required and should be used as informed on the models table, on the Input parameter name column. Standard values for the other parameters are given below:

ways	shots	gpu	lr	l2_weight
2	5	0	0.0001	0.0

Generate GradCAM

All heatmaps are generated for images from a class that was not presented during training. Not all models can generate GradCAM explainability, due to their architecture. In order to generate a GradCAM for an image, run:

python explain.py \
  --model <model_name> \
  --weights <weights.pth> \
  --ways <number_of_classes_per_task> \
  --shots <number_of_examples_per_class_per_task> \
  --gpu <gpu_index> \
  --image <image_file_name_with_no_extension> \
  --target-class <target_class_name> \
  --classes <"target_class,other_test_class_for_fine_tuning"> \
  --augment <cutmix/mixup/None>

Citation

If you find this library useful in your research or project, please consider citing:

...