Metadata-Version: 2.1
Name: pynsm
Version: 0.9.0
Summary: A PyTorch implementation of non-negative similarity matching
Author-email: Tiberiu Tesileanu <ttesileanu@gmail.com>, Shagesh Sridharan <shagesh1996@gmail.com>, Yanis Bahroun <ybahroun@flatironinstitute.org>
License: MIT License
        
        Copyright (c) 2022-2023 Shagesh Sridharan, Tiberiu Tesileanu
        
        Permission is hereby granted, free of charge, to any person obtaining a copy
        of this software and associated documentation files (the "Software"), to deal
        in the Software without restriction, including without limitation the rights
        to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
        copies of the Software, and to permit persons to whom the Software is
        furnished to do so, subject to the following conditions:
        
        The above copyright notice and this permission notice shall be included in all
        copies or substantial portions of the Software.
        
        THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
        IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
        FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
        AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
        LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
        OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
        SOFTWARE.
        
Project-URL: Source, https://github.com/Shagesh/pytorch-NSM
Keywords: neural network,biology,research
Classifier: License :: OSI Approved :: MIT License
Classifier: Programming Language :: Python :: 3
Classifier: Intended Audience :: Science/Research
Requires-Python: >=3.6
Description-Content-Type: text/markdown
Provides-Extra: dev
License-File: LICENSE

# Non-negative similarity matching in PyTorch

[![Python 3.8](https://img.shields.io/badge/python-3.8-green.svg)](https://www.python.org/downloads/release/python-380/)
[![license: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](https://opensource.org/licenses/MIT)

This is an implementation of non-negative similarity matching (NSM) for PyTorch focusing on ease of use, extensibility, and speed.

## Table of Contents

- [Installation](#installation)
- [Example usage](#example-usage)
- [Features](#features)
- [Questions?](#questions)

## Installation (development)

These instructions are for development use (i.e., if you want to make changes to the package). Instructions for users will be added soon.

It is strongly recommended to use a virtual environment when working with this code. The installation instructions below include the commands for creating the virtual environment, using either `conda` (recommended) or `venv`.

### Using `conda`

If you do not have `conda` installed, the easiest way to get started is with [Miniconda](https://docs.conda.io/en/latest/miniconda.html). Follow the installation instructions for your system.

Next, create a new environment and install for CPU using

```sh
conda env create -f environment.yml
```

For using an NVIDIA GPU run

```sh
conda env create -f environment-cuda.yml
```

Note that most Macs do not have an NVIDIA GPU, so you should use the first invocation shown above. If your Mac uses the newer Apple chips, you may be able to use ``device = mps`` to get GPU acceleration (the installation procedure remains unchanged).

The commands above automatically perform an "editable" install — this means that changes made to the code will automatically take effect without having to reinstall the package.

### Using `venv`

Before creating a new virtual environment, it is best to ensure you're not using the system version of Python — this is often badly out of date. Some options for doing this are outlined in [The Hitchhiker's Guide to Python](https://docs.python-guide.org/starting/installation/#installation-guides), although many options exist. One advantage of using `conda` is that this is done for you.

Once you have a proper Python install, create a new virtual environment by running the following command in a terminal inside the main folder of the repository:

```sh
python -m venv env
```

This creates a subfolder called `env` containing the files for the virtual environment. Next we need to activate the environment and install the package with its pre-requisites:

```sh
source env/bin/activate
pip install -e ".[dev]"
```

The `-e` marks this as an "editable" install — this means that changes made to the code will automatically take effect without having to reinstall the package.

## Example Usage

See the notebooks in the [`examples`](examples) folder to get started with the package.

## Features

### Similarity matching with arbitrary encoder

The code defines a neural network module called `SimilarityMatching`. This model takes an arbitrary encoder module and generates a "competitor" layer -- a linear layer that enforces competition between the different output channels from the encoder. Natural examples of encoder layers are fully-connected layers and convolutional layers. The `forward` method performs the forward pass of the model, which involves running to convergence the dynamics involving the lateral competitor connections. The model also includes a method for calculating the loss from which the plasicity rules can be derived.

### ZCA Whitening

The code includes a ZCA whitening function `computeZCAMatrix` that computes the ZCA matrix for a set of input observations `X`. The function performs normalization, reshaping, covariance computation, singular value decomposition (SVD), and builds the ZCA matrix. The whitening transformation is implemented in the `ZCATransformation` class, which takes the ZCA matrix and transformation mean as inputs and applies the transformation to a given tensor image.

### Training the Model

One of the key design goals was to make using the `SimilarityMatching` modules be as seamless as possible. Indeed, training the module can be done in precisely the same way as training any other PyTorch module. The only difference is that there is a canonical choice for the loss function, and that can be obtained from the `loss()` method of the module. See the example notebooks for details on the training.

### Classification

The pre-trained network is used in a supervised fashion to perform classification. Specifically, the embeddings generated by the `SimilarityMatching` module are followed by a max-pooling layer and then passed through an SVM (using `SGDClassifier` from `sklearn`). We obtain good accuracy on the test set.

### Supervised NSM

The code also introduces a module called `SupervisedSimilarityMatching` class, which implements supervised learning. It includes additional functionality for handling labeled data during training. It is a special case of `MultiSimilarityMatching`, which attempts to maximize the similarity between the circuit's output and multiple inputs.

## Questions?

Please contact us by opening an issue on GitHub.
