/matrix-analysis-final

Primary LanguageJupyter Notebook

This repo contains codes needed for NMF based on MU and HALS. It implements algorithms from the paper https://arxiv.org/abs/1107.5194

This codes require numpy, cv2 and matplotlib for NMF and plots part.

Deep learning models require tensorflow, tritonclient, protobuf. These packages are not needed for run.py.

NMF

Two NMF algorithms are implemented in nmf.py, if alpha=0, they reduce to non-accelerated versions.

# Accelerated Multiplicative Update
accelerated_MU(X, L, alpha, epsillon, max_iter, update_func="gillis", save_every_iter=-1, save_with_index=False, save_cb=None, weight_path="", restart=False)
# Accelerated Hierarchical ALS
accelerated_HALS(X, L, alpha, epsillon, max_iter, update_func="gillis", save_every_iter=-1, save_with_index=False, save_cb=None, weight_path="", restart=False)

Results

Facial components from HALS, L=30 Reconstructed faces from HALS, L=30

Main function

Main codes contained in run.py.

python run.py --help
usage: run.py [-h] [--datadir DATADIR] [--algorithm ALGORITHM] [--accelerated]
              [--update-func UPDATE_FUNC] [--L L] [--alpha ALPHA] [--eps EPS]
              [--max-iter MAX_ITER] [--outputdir OUTPUTDIR]
              [--output-recon-images] [--use-sample] [--restart]
              [--save-every SAVE_EVERY] [--save-with-index] [--seed SEED]

Accelerated MU/HALS

optional arguments:
  -h, --help            show this help message and exit
  --datadir DATADIR     path to directory containing standardized data
                        (default: None)
  --algorithm ALGORITHM
                        NMF algorithm to use, MU or HALS (default: None)
  --accelerated         whether or not to use the accelerated version
                        True/False (default: False)
  --update-func UPDATE_FUNC
                        which update function to use for hals, `paper` or
                        `gillis` (default: gillis)
  --L L                 low rank dimension (default: 30)
  --alpha ALPHA         control parameter for number of inner loop iterations,
                        ignored if accelerated=False (default: 2)
  --eps EPS             control parameter for stopping condition, ignored if
                        accelerated=False (default: 0.1)
  --max-iter MAX_ITER   number of iterations to update W and H (default: 2)
  --outputdir OUTPUTDIR
                        where to export the output (default: None)
  --output-recon-images
                        if true, will generate reconstructed images (default:
                        False)
  --use-sample          if true, make use of the list in meta.py to output
                        report (default: False)
  --restart             if true, restart from files available in outputdir
                        (default: False)
  --save-every SAVE_EVERY
                        save weight every # iterations (default: 0)
  --save-with-index     save weight every # iterations, without replacement
                        (default: False)
  --seed SEED           random seed (default: 42)

Samples

Commands to run can be found in script_all.sh, e.g. to perform NMF with L=30 using accelerated HALS for 100 iterations:

python run.py --datadir /mnt/i/dataset-processed --algorithm hals --accelerated --update-func gillis --L 60 --alpha 0.5 --eps 0.1 --max-iter 100 --outputdir ./HALS_Accel_L60

Figures

Other Python files help generating some figures, such as compression ratio

Compression ratio