Skip to content


Repository files navigation

Test Status arXiv

learn2learn is a software library for meta-learning research.

learn2learn builds on top of PyTorch to accelerate two aspects of the meta-learning research cycle:

  • fast prototyping, essential in letting researchers quickly try new ideas, and
  • correct reproducibility, ensuring that these ideas are evaluated fairly.

learn2learn provides low-level utilities and unified interface to create new algorithms and domains, together with high-quality implementations of existing algorithms and standardized benchmarks. It retains compatibility with torchvision, torchaudio, torchtext, cherry, and any other PyTorch-based library you might be using.

To learn more, see our whitepaper: arXiv:2008.12284


  • Taskset and transforms to create few-shot tasks from any PyTorch dataset.
  • Models, datasets, and benchmarks for computer vision and few-shot learning.
  • learn2learn.gym: Environment and utilities for meta-reinforcement learning.
  • learn2learn.algorithms: High-level wrappers for existing meta-learning algorithms.
  • learn2learn.optim: Utilities and algorithms for differentiable optimization and meta-descent.



pip install learn2learn

Snippets & Examples

The following snippets provide a sneak peek at the functionalities of learn2learn.

High-level Wrappers

Few-Shot Learning with MAML

For more algorithms (ProtoNets, ANIL, Meta-SGD, Reptile, Meta-Curvature, KFO) refer to the examples folder. Most of them can be implemented with with the GBML wrapper. (documentation).

maml = l2l.algorithms.MAML(model, lr=0.1)
opt = torch.optim.SGD(maml.parameters(), lr=0.001)
for iteration in range(10):
    task_model = maml.clone()  # torch.clone() for nn.Modules
    adaptation_loss = compute_loss(task_model)
    task_model.adapt(adaptation_loss)  # computes gradient, update task_model in-place
    evaluation_loss = compute_loss(task_model)
    evaluation_loss.backward()  # gradients w.r.t. maml.parameters()
Meta-Descent with Hypergradient

Learn any kind of optimization algorithm with the LearnableOptimizer. (example and documentation)

linear = nn.Linear(784, 10)
transform = l2l.optim.ModuleTransform(l2l.nn.Scale)
metaopt = l2l.optim.LearnableOptimizer(linear, transform, lr=0.01)  # metaopt has .step()
opt = torch.optim.SGD(metaopt.parameters(), lr=0.001)  # metaopt also has .parameters()

error = loss(linear(X), y)
opt.step()  # update metaopt
metaopt.step()  # update linear

Learning Domains

Custom Few-Shot Dataset

Many standardized datasets (Omniglot, mini-/tiered-ImageNet, FC100, CIFAR-FS) are readily available in (documentation)

dataset =  # any PyTorch dataset
transforms = [  # Easy to define your own transform, n=5),, k=1),,
taskset = Taskset(dataset, transforms, num_tasks=20000)
for task in taskset:
    X, y = task
    # Meta-train on the task
Environments and Utilities for Meta-RL

Parallelize your own meta-environments with AsyncVectorEnv, or use the standardized ones. (documentation)

def make_env():
    env = l2l.gym.HalfCheetahForwardBackwardEnv()
    env = cherry.envs.ActionSpaceScaler(env)
    return env

env = l2l.gym.AsyncVectorEnv([make_env for _ in range(16)])  # uses 16 threads
for task_config in env.sample_tasks(20):
    env.set_task(task)  # all threads receive the same task
    state = env.reset()  # use standard Gym API
    action = my_policy(env)

Low-Level Utilities

Differentiable Optimization

Learn and differentiate through updates of PyTorch Modules. (documentation)

model = MyModel()
transform = l2l.optim.KroneckerTransform(l2l.nn.KroneckerLinear)
learned_update = l2l.optim.ParameterUpdate(  # learnable update function
        model.parameters(), transform)
clone = l2l.clone_module(model)  # torch.clone() for nn.Modules
error = loss(clone(X), y)
updates = learned_update(  # similar API as torch.autograd.grad
l2l.update_module(clone, updates=updates)
loss(clone(X), y).backward()  # Gradients w.r.t model.parameters() and learned_update.parameters()


A human-readable changelog is available in the file.


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

Arnold, Sebastien M. R., Praateek Mahajan, Debajyoti Datta, Ian Bunner, and Konstantinos Saitas Zarkias. 2020. “learn2learn: A Library for Meta-Learning Research.” arXiv [cs.LG].

You can also use the following Bibtex entry.

  title         = "learn2learn: A Library for {Meta-Learning} Research",
  author        = "Arnold, S{\'e}bastien M R and Mahajan, Praateek and Datta,
                   Debajyoti and Bunner, Ian and Zarkias, Konstantinos Saitas",
  month         =  aug,
  year          =  2020,
  url           = "",
  archivePrefix = "arXiv",
  primaryClass  = "cs.LG",
  eprint        = "2008.12284"

Acknowledgements & Friends

  1. TorchMeta is similar library, with a focus on datasets for supervised meta-learning.
  2. higher is a PyTorch library that enables differentiating through optimization inner-loops. While they monkey-patch nn.Module to be stateless, learn2learn retains the stateful PyTorch look-and-feel. For more information, refer to their ArXiv paper.
  3. We are thankful to the following open-source implementations which helped guide the design of learn2learn: