meta-learning

Overview

I have been working on a side project over the past few days. Along the way, I implemented the following meta-learning algorithms in pytorch:

• MAML (paper)
• MAML++ (paper)
• Reptile (paper)
• MetaSGD (paper)

The MAML++ implementation only includes the multi-step loss optimization meant to improve gradient stability.

These algorithms are toy implementations. The code is decently documented if you want to take a look, but meta_learning_algorithms.py lacks abstractions across the algorithm implementations.

Example: MetaSGD on the Sinusoid Dataset

Requirements

• torch==2.5.0
• higher==0.2.1 for MAML and MAML++

Example Usage

from meta_learning_algorithms import MAML, Reptile, MetaSGD

...  # loading datasets, creating the models, etc.

maml = MAML(
    model=maml_model,
    loss=nn.MSELoss(),
    maml_plus_plus=False,
    inner_lr=1e-2,
    meta_lr=1e-1,
    device="cpu"
)
maml_plus_plus = MAML(
    model=maml_pp_model,
    loss=nn.MSELoss(),
    maml_plus_plus=True,
    inner_lr=1e-2,
    meta_lr=1e-1,
    device="cpu"
)
meta_sgd = MetaSGD(
    model=meta_sgd_model,
    loss=nn.MSELoss(),
    inner_lr=1e-3,
    meta_lr=1e-3,
    device="cpu"
)
reptile = Reptile(
    model=reptile_model,
    loss=nn.MSELoss(),
    inner_lr=1e-1,
    meta_lr=1e-1,
    clipping=4.0,
    device="cpu"
)

and then training would look like

maml.train(train_data_loader, val_data_loader, epochs=15)

Random Footnotes

• The MetaSGD implementation uses the functional_call from torch.func instead of the higher monkey-patched functional module function. I did not realize that pytorch had such a function-- this makes it way easier to do meta-learning
• The get_per_step_loss_importance_vector in MAML++ is taken from the How To Train Your MAML source code. Maybe I didn't read the paper close enough, but I could not find how they calculated $v_i$ in the paper
• The code is not abstracted well because I originally wrote these in a set of messy jupyter notebooks before putting them together