Factor Analysis on pure python

This MR implements the ISV and the JFA algorithms on pure python using numpy/dask-arrays. In the way I'm implementing it, we can easily run this algorithm in a distributed manner, either locally on an HPC.

Closes #41 (closed)

The implementation of ISV is done (docs are missing still).

JFA is on the way

assigned to @tiago.pereira

added 1 commit

• 7fd907ff - JFA, finished update rule for V and U. D is still missing

Compare with previous version

added 1 commit

• e088b20c - Training script done

Compare with previous version

Some updates

• Training ISV and JFA is done.
• Enrollment functions for both stategies implemented

Only the scoring is missing. Hope to get it done by tomorrow

added 1 commit

• e10ee7b5 - Implemented the scoring function

Compare with previous version

added 1 commit

• e9b7f547 - Implemented the scoring function

Compare with previous version

Implemented the scoring function for the ISV and JFA.

Things to be done:

• set the load and save to HDF5
• update the docs
• do some more tests

@amohammadi, do you have a trained UBM and a set of GMM stas from voxforge that I can use?

Thanks

please ask @ydayer he should be able to help you train one. It trains in a few minutes.

Thanks @amohammadi

Hey @ydayer; Do you have an already trained UBM and some GMMStats from voxforge so I can use them?

Thanks

I have a UBM trained on Voxforge if you want (algorithm_trainer.pkl).

I just noticed a bug (trying to regenerate some GMMStats for you) where I cannot load a checkpointed UBM though (I always ran the pipeline in memory lately, so I missed that). I'm looking into it.

Thanks @ydayer,

I'm waiting for it then

added 1 commit

• ba37b6f9 - Optimized some functions

Compare with previous version

added 1 commit

• 1c534380 - Added name

Compare with previous version

added 1 commit

• 0893f03b - Added plotting mechanism

Compare with previous version

added 1 commit

• fef377f4 - Added plotting mechanism

Compare with previous version

My code seems working :-) Below follow some intuition plots from JFA using MNIST. I'm getting the same plots I was getting with the old JFA

added 1 commit

• a8f19ef1 - Fixing plots

Compare with previous version

added 1 commit

• f9ed9c56 - [precommit] fixes

Compare with previous version

Pre-commit works on my computer, but doesn't work on the CI

black....................................................................Failed
- hook id: black
- duration: 0.25s
- exit code: 1
Traceback (most recent call last):
  File "/scratch/builds/bob/bob.pipelines/.cache/pre-commit/repol43ayt_c/py_env-python3/bin/black", line 8, in <module>
    sys.exit(patched_main())
  File "/scratch/builds/bob/bob.pipelines/.cache/pre-commit/repol43ayt_c/py_env-python3/lib/python3.9/site-packages/black/__init__.py", line 1129, in patched_main
    patch_click()
  File "/scratch/builds/bob/bob.pipelines/.cache/pre-commit/repol43ayt_c/py_env-python3/lib/python3.9/site-packages/black/__init__.py", line 1115, in patch_click
    from click import _unicodefun  # type: ignore
ImportError: cannot import name '_unicodefun' from 'click' (/scratch/builds/bob/bob.pipelines/.cache/pre-commit/repol43ayt_c/py_env-python3/lib/python3.9/site-packages/click/__init__.py)
flake8...................................................................Passed

Anyone experiencing this?

It seems that black has a bug.

bob.pipelines is failing too https://gitlab.idiap.ch/bob/bob.pipelines/-/jobs/262684

There is indeed an issue:

https://github.com/psf/black/issues/2964

added 1 commit

• 81b60744 - [pre-commit] fixing issue

Compare with previous version