A pytorch implementation of the regularized non-negative matrix tri-factorization algorithm. Compatible with CUDA enabled GPU's. The code was developed in an enviroment with the following packages. This uses the cuda11.6 package.
- python 3.8.1
- pytorch 1.12.0 py3.8_cuda11.6_cudnn8.3.2_0
- torchaudio 0.12.0
- torchvision 0.13.0
- cudatoolkit 11.6.0
- numpy 1.22.3
- pandas 1.4.3
The goal of this function is to take a matrix X of dimension n x m and factor it into three matrices, U in n x k1, S in k1 x k2, and V in k2 x m, such that the difference between X and the product U x S x V is minimized. Formally this is resolved by minimizing the objective || X - U S V ||2^2. The program outputs the three product matrices.
The model allows for the addition of orthoganility regularization on the U and/or the V matrix. The idea of this regularization is to generate unique factors. It is particularly useful in resolving unique clusters from the lower dimensional embedding. It is recommend to put orthoganility reg. on either or both factors such that they are unique. If both are selected, then the S matrix will have more overlap.
This works simililarly to the ortho regularization. In this case, the number of non-zero entries in each factor in penalized. It again will result in more unique factors although.
| Argument | is Required? | Description | additional Info |
|---|---|---|---|
| --in_file: | required | The file containing the tab delimited X matrix if full form. | example: test/A.txt |
| --k1: | required | The lower dimension of the U matrix. | example 2 |
| --k2: | required | The lower dimension of the V matrix. | example 3 |
| --lU: | optional | strength of ortho reg on U. | example 1000 (value dependent on number of features along rows) |
| --lV: | optional | strength of ortho reg on V. | example 1000 (value dependent on number of features along cols) |
| --aU: | optional | strength of sparsity reg on U. | example 1000 (value dependent on number of features along rows) |
| --aV: | optional | strength of sparsity reg on V. | example 1000 (value dependent on number of featurs along cols) |
| --verbose | optional | if included, time and convergence info will be printed to terminal | |
| --seed | optional | random seed used to initialize U, S, and V. | Default 1010 |
| --max_iter | optional | number of epochs to attempt prior to output. | Default 100 |
| --term_tol | optional | the relative change in error prior to completion. | Default 1e-5 |
| --out_dir | optional | output directory to print U, S, and V. | |
| --cpu | optional | if included, defuaults to using the CPU |
./run_example.sh 1. Defaults to the CPU ./run_example.sh 0. Runs on GPU if available.