Skip to main content

MetaNN provides extensions of PyTorch nn.Module for meta learning

Project description

Documentation Status

1. Introduction

In meta learner scenario, it is common use dependent variables as parameters, and back propagate the gradient of the parameters. However, parameters of PyTorch Module are designed to be leaf nodes and it is forbidden for parameters to have grad_fn. Meta learning coders are therefore forced to rewrite the basic layers to adapt the meta learning requirements.

This module provide an extension of torch.nn.Module, DependentModule that has dependent parameters, allowing the differentiable dependent parameters. It also provide the method to transform nn.Module into DependentModule, and turning all of the parameters of a nn.Module into dependent parameters.

2. Installation

pip install MetaNN

3. Example

PyTorch suggest all parameters of a module to be independent variables. Using DependentModule arbitrary torch.nn.module can be transformed into dependent module.

from metann import DependentModule
from torch import nn
net = torch.nn.Sequential(
    nn.Linear(10, 100),
    nn.Linear(100, 5))
net = DependentModule(net)
print(net)

Higher-level api such as MAML class are more recommended to use.

from metann.meta import MAML, default_evaluator_classification as evaluator
from torch import nn
net = torch.nn.Sequential(
    nn.Linear(10, 100),
    nn.Linear(100, 5))
)
maml = MAML(net, steps_train=5, steps_eval=10, lr=0.01)
output = maml(data_train)
loss = evaluator(output, data_test)
loss.backward()

4. Documents

The documents are available at ReadTheDocs. MetaNN

5. License

MIT

Copyright (c) 2019-present, Hanqiao Yu

Project details


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

MetaNN-0.3.2.tar.gz (9.4 kB view hashes)

Uploaded Source

Built Distribution

MetaNN-0.3.2-py3-none-any.whl (11.1 kB view hashes)

Uploaded Python 3

Supported by

AWS AWS Cloud computing and Security Sponsor Datadog Datadog Monitoring Fastly Fastly CDN Google Google Download Analytics Microsoft Microsoft PSF Sponsor Pingdom Pingdom Monitoring Sentry Sentry Error logging StatusPage StatusPage Status page