Few-shot learning

The aim for this repository is to contain clean, readable and tested code to reproduce few-shot learning research.

This project is written in python 3.6 and Pytorch and assumes you have a GPU.

See these Medium articles for some more information

1. Theory and concepts
2. Discussion of implementation details

Setup

Requirements

Listed in requirements.txt. Install with pip install -r requirements.txt preferably in a virtualenv.

Data

Edit the DATA_PATH variable in config.py to the location where you store the Omniglot and miniImagenet datasets.

After acquiring the data and running the setup scripts your folder structure should look like

DATA_PATH/
    Omniglot/
        images_background/
        images_evaluation/
    miniImageNet/
        images_background/
        images_evaluation/

Omniglot dataset. Download from https://github.com/brendenlake/omniglot/tree/master/python, place the extracted files into DATA_PATH/Omniglot_Raw and run scripts/prepare_omniglot.py

miniImageNet dataset. Download files from https://drive.google.com/file/d/0B3Irx3uQNoBMQ1FlNXJsZUdYWEE/view, place in data/miniImageNet/images and run scripts/prepare_mini_imagenet.py

Tests (optional)

After adding the datasets run pytest in the root directory to run all tests.

Results

The file experiments/experiments.txt contains the hyperparameters I used to obtain the results given below.

Prototypical Networks

Run experiments/proto_nets.py to reproduce results from Prototpyical Networks for Few-shot Learning (Snell et al).

Arguments

• dataset: {'omniglot', 'miniImageNet'}. Whether to use the Omniglot or miniImagenet dataset
• distance: {'l2', 'cosine'}. Which distance metric to use
• n-train: Support samples per class for training tasks
• n-test: Support samples per class for validation tasks
• k-train: Number of classes in training tasks
• k-test: Number of classes in validation tasks
• q-train: Query samples per class for training tasks
• q-test: Query samples per class for validation tasks

	Omniglot
k-way	5	5	20	20
n-shot	1	5	1	5
Published	98.8	99.7	96.0	98.9
This Repo	98.2	99.4	95.8	98.6

	miniImageNet
k-way	5	5
n-shot	1	5
Published	49.4	68.2
This Repo	48.0	66.2

Matching Networks

A differentiable nearest neighbours classifier.

Run experiments/matching_nets.py to reproduce results from Matching Networks for One Shot Learning (Vinyals et al).

Arguments

• dataset: {'omniglot', 'miniImageNet'}. Whether to use the Omniglot or miniImagenet dataset
• distance: {'l2', 'cosine'}. Which distance metric to use
• n-train: Support samples per class for training tasks
• n-test: Support samples per class for validation tasks
• k-train: Number of classes in training tasks
• k-test: Number of classes in validation tasks
• q-train: Query samples per class for training tasks
• q-test: Query samples per class for validation tasks
• fce: Whether (True) or not (False) to use full context embeddings (FCE)
• lstm-layers: Number of LSTM layers to use in the support set FCE
• unrolling-steps: Number of unrolling steps to use when calculating FCE of the query sample

I had trouble reproducing the results of this paper using the cosine distance metric as I found the converge to be slow and final performance dependent on the random initialisation. However I was able to reproduce (and slightly exceed) the results of this paper using the l2 distance metric.

	Omniglot
k-way	5	5	20	20
n-shot	1	5	1	5
Published (cosine)	98.1	98.9	93.8	98.5
This Repo (cosine)	92.0	93.2	75.6	77.8
This Repo (l2)	98.3	99.8	92.8	97.8

	miniImageNet
k-way	5	5
n-shot	1	5
Published (cosine, FCE)	44.2	57.0
This Repo (cosine, FCE)	42.8	53.6
This Repo (l2)	46.0	58.4

Model-Agnostic Meta-Learning (MAML)

I used max pooling instead of strided convolutions in order to be consistent with the other papers. The miniImageNet experiments using 2nd order MAML took me over a day to run.

Run experiments/maml.py to reproduce results from Model-Agnostic Meta-Learning (Finn et al).

Arguments

• dataset: {'omniglot', 'miniImageNet'}. Whether to use the Omniglot or miniImagenet dataset
• distance: {'l2', 'cosine'}. Which distance metric to use
• n: Support samples per class for few-shot tasks
• k: Number of classes in training tasks
• q: Query samples per class for training tasks
• inner-train-steps: Number of inner-loop updates to perform on training tasks
• inner-val-steps: Number of inner-loop updates to perform on validation tasks
• inner-lr: Learning rate to use for inner-loop updates
• meta-lr: Learning rate to use when updating the meta-learner weights
• meta-batch-size: Number of tasks per meta-batch
• order: Whether to use 1st or 2nd order MAML
• epochs: Number of training epochs
• epoch-len: Meta-batches per epoch
• eval-batches: Number of meta-batches to use when evaluating the model after each epoch

NB: For MAML n, k and q are fixed between train and test. You may need to adjust meta-batch-size to fit your GPU. 2nd order MAML uses a lot more memory.

	Omniglot
k-way	5	5	20	20
n-shot	1	5	1	5
Published	98.7	99.9	95.8	98.9
This Repo (1)	95.5	99.5	92.2	97.7
This Repo (2)	98.1	99.8	91.6	95.9

	miniImageNet
k-way	5	5
n-shot	1	5
Published	48.1	63.2
This Repo (1)	46.4	63.3
This Repo (2)	47.5	64.7

Number in brackets indicates 1st or 2nd order MAML.