Skip to content
PyTorch Meta-learning Framework for Researchers
Python Makefile
Branch: master
Clone or download
Type Name Latest commit message Commit time
Failed to load latest commit information.
.github Add template for future PRs. (#80) Oct 11, 2019
docs Adding, updating, and fixing docs. (#53) Sep 12, 2019
examples Add information to the vision examples + link to TorchMeta. (#68) Sep 23, 2019
learn2learn typo (#83) Oct 8, 2019
.travis.yml Lint codebase (#50) Sep 10, 2019
Makefile Add citation information and link to higher to README. (#81) Oct 8, 2019
setup.cfg Updated Sep 15, 2019

Build Status

learn2learn is a PyTorch library for meta-learning implementations.

The goal of meta-learning is to enable agents to learn how to learn. That is, we would like our agents to become better learners as they solve more and more tasks. For example, the animation below shows an agent that learns to run after a only one parameter update.


learn2learn provides high- and low-level utilities for meta-learning. The high-level utilities allow arbitrary users to take advantage of exisiting meta-learning algorithms. The low-level utilities enable researchers to develop new and better meta-learning algorithms.

Some features of learn2learn include:

  • Modular API: implement your own training loops with our low-level utilities.
  • Provides various meta-learning algorithms (e.g. MAML, FOMAML, MetaSGD, ProtoNets, DiCE)
  • Task generator with unified API, compatible with torchvision, torchtext, torchaudio, and cherry.
  • Provides standardized meta-learning tasks for vision (Omniglot, mini-ImageNet), reinforcement learning (Particles, Mujoco), and even text (news classification).
  • 100% compatible with PyTorch -- use your own modules, datasets, or libraries!


pip install learn2learn

API Demo

The following is an example of using the high-level MAML implementation on MNIST. For more algorithms and lower-level utilities, please refer to the documentation or the examples.

import learn2learn as l2l

mnist = torchvision.datasets.MNIST(root="/tmp/mnist", train=True)

mnist =
task_generator =,
                                        classes=[0, 1, 4, 6, 8, 9],
model = Net()
maml = l2l.algorithms.MAML(model, lr=1e-3, first_order=False)
opt = optim.Adam(maml.parameters(), lr=4e-3)

for iteration in range(num_iterations):
    learner = maml.clone()  # Creates a clone of model
    adaptation_task = task_generator.sample(shots=1)

    # Fast adapt
    for step in range(adaptation_steps):
        error = compute_loss(adaptation_task)

    # Compute evaluation loss
    evaluation_task = task_generator.sample(shots=1,
    evaluation_error = compute_loss(evaluation_task)

    # Meta-update the model parameters


To cite the learn2learn repository in your academic publications, please use the following reference.

Sébastien M.R. Arnold, Praateek Mahajan, Debajyoti Datta, Ian Bunner. "learn2learn"., 2019.

You can also use the following Bibtex entry.

    author       = {Sébastien M.R. Arnold, Praateek Mahajan, Debajyoti Datta, Ian Bunner},
    title        = {learn2learn},
    month        = sep,
    year         = 2019,
    url          = {}

Acknowledgements & Friends

  1. The RL environments are adapted from Tristan Deleu's implementations and from the ProMP repository. Both shared with permission, under the MIT License.
  2. TorchMeta is similar library, with a focus on supervised meta-learning. If learn2learn were missing a particular functionality, we would go check if TorchMeta has it. But we would also open an issue ;)
  3. higher is a PyTorch library that also enables differentiating through optimization inner-loops. Their approach is different from learn2learn in that they monkey-patch nn.Module to be stateless. For more information, refer to their ArXiv paper.
You can’t perform that action at this time.