heinsaar/aimx

Aimx

Various AI / DL audio experiments & exploration.

aimx's Issues

Dataprep & train pipeline to optionally use intermediate JSON file

Try Keras mixed precision training on GPUs

Apply & explore results of the NNs on the larger datasets

Make sure scripts work cross-platform (test on Linux)

Originally they were developed under Windows. Some code (like dir.split("\\" and other) needs to be made to work for Linux, etc.

Execution timing, duration measurement and reports

Create code metrics test to have reports on LOC, Lint and other useful code stats

Make Inferex work real-time on a mic stream

Python static code checking / lint

Fast cleanup of uninteresting contents of gen_* dirs

"Uninteresting" are usually those that are produced as a result of fast dataprep and/or fast training, for example, sanitest output and quick experimentation results. Some markers: very few epochs, very small depth, no meaningful accuracy, etc.

Possible parameter names: -notrash

Hoist hard-coded parameters into clargs

Related TODOs highlighted in comments in this commit: 5986181

Automate dir-label relationship

Create Inferex module by integrating OpenVINO for inferencing

Integrate TensorBoard for faster experimentation

CLI option to stop sanitest pipeline on error

Current default is to continue.

Make dir-label relationship work correctly with dataset views

Replace print_info() with differentiated and proper logging

INFO (white), WARNING (brown), ERROR (red), etc

Likely candidate: Python logging

Create tests for messaging verification

Error texts, warnings, etc. This is better be done later, when messaging is stabilized.

Detect all words in audio files of arbitrary length

Initial setup of GitHub Actions for the first stable CI pipeline link

Sanitest: automatically detect changed files and run only affected scripts

Add accuracy test sanitest_accuracy.py

Create a common utils module and move the relevant code there.

Add help texts for command line arguments.

Clarg for views into datasets instead of actual physical splits

Automate classification layer neuron count deduction based on labels

Mentex yardstone for audio data augmentation

To include things like mixing, trimming, etc. — exact actions necessary will become clear and specified accordingly as we go forward.

Expected first yardstone is ability to call:
mentex.py -mapolicy random -signals ../workdir/signals -noises ../workdir/noises
Which will result in a new directory /workdir/signoises containing signals that have been paired with noises according to the random mapping policy (mapolicy). List of initial policies to include:

• random (in which case every signal will be mixed with a noise picked randomly at each runtime)
• direct (in which case every signal will be mixed with a corresponding noise picked deterministically at each runtime)

Add a minimum acceptable accuracy test

This test will ensure that a new dataset or model produce at least as good an accuracy on certain tasks as the old accepted standard.