can-song/FF_unsupervised

An implementation of unsupervised example of the Forward-Forward algorithm proposed by (Hinton, 2020)

FF_unsupervised

A standalone PyTorch implementation of the Forward-Forward algorithm (specifically the unsupervised example) proposed by (Hinton, 2020), Sec 3.2

(make sure to change runtime type to GPU)

File structure:
.
├── main.py
├── main.ipynb
├── utils.py
├── MNIST
├── environment.yml
├── README.md
└── LICENSE

The utils.py file has the functions to generate the negative examples.

The prepare_data.py file downloads the MNIST dataset, generates a dataset of negative images, and saves it as a .pt file that can be loaded directly into the code as a dataset, and then a dataloader. This saves a lot of time compared to generating negative data on the fly.

The main.py file has the Unsupervised_FF class along with the training procedure included inside the main block.

Contributions are welcome through pull requests :)