The Sparsely Gated Mixture of Experts Layer for PyTorch

This repository contains the PyTorch re-implementation of the sparsely-gated MoE layer described in the paper Outrageously Large Neural Networks for PyTorch.

from moe import MoE
import torch
# instantiate the MoE layer
model = MoE(input_size=1000, output_size=20, num_experts=10,hidden_size=66, k= 4, noisy_gating=True)
X = torch.rand(32, 1000)
#train
model.train()
# forward
y_hat, aux_loss = model(X)
# evaluation
model.eval()
y_hat, aux_loss = model(X)

Requirements

To install the requirements run:

pip install -r requirements.py

Example

The file example.py contains a minimal working example illustrating how to train and evaluate the MoE layer with dummy inputs and targets. To run the example:

python example.py

CIFAR 10 example

The file cifar10_example.py contains a minimal working example of the CIFAR 10 dataset. It achieves an accuracy of 39% with arbitrary hyper-parameters and not fully converged. To run the example:

python cifar10_example.py

Used by

FastMoE: A Fast Mixture-of-Expert Training System This implementation was used as a reference PyTorch implementation for single-GPU training.

Acknowledgements

The code is based on the TensorFlow implementation that can be found here.

Citing

@misc{rau2019moe,
    title={Sparsely-gated Mixture-of-Experts PyTorch implementation},
    author={Rau, David},
    journal={https://github.com/davidmrau/mixture-of-experts},
    year={2019}
}