HesScale: Scalable Computation of Hessian Diagonals

HesScale is built on top of Pytorch and BackPack. It allows for Hessian diagonals to backpropagate through the layers of the network.

Installation:

1. You need to have environemnt with python 3.7:

python3.7 -m venv .hesscale
source .hesscale/bin/activate

2. Install Dependencies:

python -m pip install --upgrade pip
pip install .

Run a minimal example:

We added a couple of minimal examples in the examples directory for easier understanding of how to use this package. Here is a minimal example:

import torch
from backpack import backpack, extend
from optimizers.adahesscale import AdaHesScale

hidden_units = 128
n_obs = 6
n_classes = 10
lr = 0.0004
batch_size = 1

model = torch.nn.Sequential(
    torch.nn.Linear(n_obs, hidden_units),
    torch.nn.Sigmoid(),
    torch.nn.Linear(hidden_units, hidden_units),
    torch.nn.Tanh(),
    torch.nn.Linear(hidden_units, n_classes),
)
loss_func = torch.nn.CrossEntropyLoss()
optimizer = AdaHesScale(model.parameters(), lr=lr)

extend(model)
extend(loss_func)

inputs = torch.randn((batch_size, n_obs))
target_class = torch.randint(0, n_classes, (batch_size,))

prediction = model(inputs)
optimizer.zero_grad()
loss = loss_func(prediction, target_class)

with backpack(optimizer.method):
    loss.backward()

optimizer.step()

Contributing

We appreciate any help to extend HesScale to recurrent neural networks. If you consider contributing, please fork the repo and create a pull request.

License

Distributed under the MIT License. See LICENSE for more information.

Reproduction

To reproduce the experiments in the paper, you run the scripts in experiemnts directory for reproducing the approximation-quality experiment and the computational-cost experiment. For reproducing the training plots, please refer to the our other repo that uses DeepOBS for complete reproduction of our optimization results.

How to cite

If you use our code, please consider citing our paper too.

Elsayed, M., & Mahmood, A. R. (2022). HesScale: Scalable Computation of Hessian Diagonals. arXiv preprint arXiv:2210.11639.

mohmdelsayed/HesScale