cbmira01 / featureranking Goto Github PK

Needs tests for:

• each OpenCL kernel
• dataset entropy
• column entropy
• compare OpenCL and Numpy results

Do this issue before refactoring issues

What will docs look like?

What will device driver setup look like?

How will project deployment work?

What will different Linux distributions look like? Or just target one (Mint/Ubuntu, CentOS)?

We want to measure computation effort, not trial runner management.

Cannot iterate over device list.

Also: make sure each module can run at console, either to run a test or to direct to the main module.

Feature ranking is completely wrong.

Re-read Kantardzic p68: Implement this properly in the trial runner.

Fix Wiki Report, restate properly how features are selected for removal in the round-robin trial.
Also, get rid of existing workstation results in the report.

Fix other wording as encountered.

Look at kernel double-precision floats (cl_khr_fp64 option), use if possible.

Look for platforms that DO NOT have this option, and see how that works.

This project needs a wiki!

Is there any way to make the host-code portion look any less ugly...??
Or is OpenCL host code just ugly to begin with?

'Ranking protocol' code is duplicated in the feature reduction test runners, and needs to be pulled out into its own 'trial runner' module. That's also the place to make a 'trial context' object that knows everything about how to run a test.

• Put feature names in the CSVs
• Do better job indicating key and target fields
• Clean up descriptive text

Fix equation rendering in Wiki report section.

Here is a sample codecog:

https://latex.codecogs.com/svg.image?D_i_j&space;=&space;\left&space;[&space;\sum_{k=1}^{n}&space;((x_i_k&space;-&space;x_j_k)&space;/&space;(max_k&space;-&space;min_k))^{2}&space;\right&space;]^{1/2}

This project should have had logging early, to support refactoring and optimization.
Log files should be ignored by git.

Pre-compute column entropies and value ranges.
Big performance gain here.
Do the 'trial runner' first, because that's where these pre-computations will go.
Work on logging and OpenCL reduction before pre-computation.

A NAN is being returned in OpenCL get_entropy function. Track this down please.
At least guard and fail nicely for it before going deeper.

We need Open-CL sum reduction to work properly, because at least two big calculations depend on it.

A fix here may generalize to min/max reductions.

Make sure the menu input is less confusing