Meta-Learning & Computer Vision

This directory contains meta-learning examples and reproductions for common computer vision benchmarks.

MAML

The following files reproduce MAML on the Omniglot and mini-ImageNet datasets. The FOMAML results can be obtained by setting first_order=True in the MAML wrapper. On Omniglot, the CNN results can be obtained by swapping OmniglotFC with OmniglotCNN.

maml_omniglot.py - MAML on the Omniglot dataset with a fully-connected network.
maml_miniimagenet.py - MAML on the mini-ImageNet dataset with the standard convolutional network.

Note that the original MAML paper trains with 5 fast adaptation step, but tests with 10 steps. This implementation only provides the training code.

Results

When adapting the code to different datasets, we obtained the following results. Only the fast-adaptation learning rate needs a bit of tuning, and good values usually lie in a 0.5-2x range of the original value.

Dataset	Architecture	Ways	Shots	Original	learn2learn
Omniglot	FC	5	1	89.7%	88.9%
Omniglot	CNN	5	1	98.7%	99.1%
mini-ImageNet	CNN	5	1	48.7%	48.3%
mini-ImageNet	CNN	5	5	63.1%	65.4%
CIFAR-FS	CNN	5	5	71.5%	73.6%
FC100	CNN	5	5	n/a	49.0%

Usage

Manually edit the respective files and run:

python examples/vision/maml_omniglot.py

or

python examples/vision/maml_miniimagenet.py

Prototypical Networks

The file protonet_miniimagenet.py reproduces Prototypical Networks on the mini-ImageNet dataset.

This implementation provides training and testing code.

Results

Dataset	Architecture	Ways	Shots	Original	learn2learn
mini-ImageNet	CNN	5	1	49.4%	49.1%
mini-ImageNet	CNN	5	5	68.2%	66.5%

Usage

For 1 shot 5 ways:

python examples/vision/protonet_miniimagenet.py

For 5 shot 5 ways:

python examples/vision/protonet_miniimagenet.py --shot 5 --train-way 20

ANIL

The file anil_fc100.py implements ANIL on the FC100 dataset.

Results

While ANIL only used mini-ImageNet as a benchmark, we provide results for CIFAR-FS and FC100 as well.

Dataset	Architecture	Ways	Shots	Original	learn2learn
mini-ImageNet	CNN	5	5	61.5%	63.2%
CIFAR-FS	CNN	5	5	n/a	68.3%
FC100	CNN	5	5	n/a	47.6%

Usage

Manually edit the above file and run:

python examples/vision/anil_fc100.py

Reptile

The file reptile_miniimagenet.py reproduces Reptile on the mini-ImageNet dataset.

Results

The mini-ImageNet file can easily be adapted to obtain results on Omniglot and CIFAR-FS as well.

Dataset	Architecture	Ways	Shots	Original	learn2learn
Omniglot	CNN	5	5	99.5%	99.5%
mini-ImageNet	CNN	5	5	66.0%	65.5%
CIFAR-FS	CNN	10	3	n/a	46.3%

Usage

Manually edit the above file and run:

python examples/vision/reptile_miniimagenet.py

Baseline

The file supervised_pretraining.py reproduces the pretraining baseline of Dhillon et al. and extends to different architectures, datasets, and data augmentation.

The pretrained weights can be downloaded using l2l.vision.models.get_pretrained_backbone().

Results

The mini-ImageNet file can easily be adapted to obtain results on Omniglot and CIFAR-FS as well.

Dataset	Architecture	Ways	Shots	Original	learn2learn
CIFAR-FS	CNN4	5	5	n / a	73.13%
FC100	CNN4	5	5	n / a	52.18%
mini-ImageNet	ResNet12	5	5	73.31%	77.38%
tiered-ImageNet	ResNet12	5	5	82.88%	83.80%

Usage

Manually edit the above file and run:

python examples/vision/supervised_pretraining.py

Also see examples/vision/Makefile for reproducible commands.