renovamen / metallic

A clean, lightweight and modularized PyTorch meta-learning library.

Metallic

Metallic is a meta-learning library based on PyTorch.

Different from other meta-learning libraries, Metallic tries to:

• Provide a clean, lightweight and modularized toolbox for setting up meta-learning experiments quickly with the least amount of code.
• For gradient-based meta-learning methods (like MAML), support more optimizers instead of SGD only using higher.

The library is work in progress.

 

Features

Algorithms

The supported interface algorithms currently include:

Gradient-based

• Model-Agnostic Meta-Learning (MAML), including first-order version (FOMAML)
• Reptile
• Minibatch Proximal Update
• Almost No Inner Loop (ANIL)

Metric-based

• Matching Networks
• Prototypical Networks

Datasets

The supported datasets currently include:

• Omniglot
• Mini-ImageNet

 

Installation

git clone https://github.com/Renovamen/metallic.git
cd metallic
python setup.py install

or

pip install git+https://github.com/Renovamen/metallic.git --upgrade

 

Documentations

Check the API documentation here: metallic-docs.vercel.app

 

Acknowledgements

Metallic is highly inspired by the following awesome libraries:

• learn2learn
• Torchmeta
• higher

 

License

Metallic is MIT licensed, see the LICENSE file for more details.

metallic's Issues

MAML or FOMAML implement

in the maml.py you compute the grad manully and map the grad from higher to the original model. but you only use the grad[-1] to the original model. is it works well?

                if meta_train == True:
                    # query_loss.backward()
                    outer_grad = torch.autograd.grad(query_loss / n_tasks, fmodel.parameters(time=0))
                    grad_list.append(outer_grad)

        # When in the meta-training stage, update the model's meta-parameters to
        # optimize the query losses across all of the tasks sampled in this batch.
        if meta_train == True:
            # apply gradients to the original model parameters
            apply_grads(self.model, grad_list[-1])                                      ---- about this line.
            # outer loop update

also in reptile.py