Momentum ResNets

Official library for using Momentum Residual Neural Networks [1]. These models extend any Residual architecture (for instance it also work with Transformers) to a larger class of deep learning models that consume less memory. They can be initialized with the same weights as a pretrained ResNet and are promising in fine-tuning applications.

Installation

To install momentumnet, you first need to install its dependencies:

$ pip install numpy matplotlib torch

Then install momentumnet:

$ pip install momentumnet

If you do not have admin privileges on the computer, use the --user flag with pip. To upgrade, use the --upgrade flag provided by pip.

To check if everything worked fine, you can do:

$ python -c 'import momentumnet'

and it should not give any error message.

Quickstart

The main class is MomentumNet. It creates a Momentum ResNet that iterates

\[\begin{split}v_{t + 1} &= \gamma \times v_t + (1 - \gamma) \times f_t(x_t) \\ x_{t + 1} &= x_t + v_{t + 1}\end{split}\]

These forward equations can be reversed in closed-form. This enables backpropagation without standard memory consumption, since activations do not have to be stored. This process trades memory for computations.

To get started, you can create a toy Momentum ResNet by specifying the functions f for the forward pass and the value of the momentum term, gamma.

>>> from torch import nn
>>> from momentumnet import MomentumNet
>>> hidden = 8
>>> d = 500
>>> function = nn.Sequential(nn.Linear(d, hidden), nn.Tanh(), nn.Linear(hidden, d))
>>> mresnet = MomentumNet([function,] * 10, gamma=0.9)

Momentum ResNets are a drop-in replacement for ResNets

We can transform a ResNet into a MomentumNet with the same parameters in two lines of codes. For instance, the following code instantiates a Momentum ResNet with weights of a pretrained Resnet-101 on ImageNet. We set “use_backprop” to False so that activations are not saved during the forward pass, allowing smaller memory consumption.

>>> import torch
>>> from momentumnet import transform_to_momentumnet
>>> from torchvision.models import resnet101
>>> resnet = resnet101(pretrained=True)
>>> mresnet101 = transform_to_momentumnet(resnet, gamma=0.9, use_backprop=False)

Importantly, this method also works with Pytorch Transformers module, specifying the residual layers to be turned into their Momentum version.

>>> import torch
>>> from momentumnet import transform_to_momentumnet
>>> transformer = torch.nn.Transformer(num_encoder_layers=6, num_decoder_layers=6)
>>> mtransformer = transform_to_momentumnet(transformer, sub_layers=["encoder.layers", "decoder.layers"],
>>>                                         gamma=0.9, use_backprop=False, keep_first_layer=False)

This initializes a Momentum Transformer with the same weights as the original Transformer.

Memory savings when applying Momentum ResNets to Transformers

Here is a short tutorial showing the memory gains using Momentum Transformers.

Dependencies

These are the dependencies to use momentumnet:

  • numpy (>=1.8)

  • matplotlib (>=1.3)

  • torch (>= 1.9)

  • memory_profiler

  • torchvision

  • vit_pytorch

Bug reports

Use the github issue tracker to report bugs.

Cite

[1] Michael E. Sander, Pierre Ablin, Mathieu Blondel, Gabriel Peyre. Momentum Residual Neural Networks.

Proceedings of the 38th International Conference on Machine Learning, PMLR 139:9276-9287

https://arxiv.org/abs/2102.07870

Fork me on GitHub