learnc: fit the constants in existing C code to data (9.1% smaller JPEGs, +4.9 F1 on WebRTC VAD) [P]

learnc tunes named constants in unmodified C code against a score function you write in C, and writes the same file back with new values.

Results, both on held-out data:

• JPEG, stb_image_write.h (128 quantization-table entries, tuned on 12 Kodak images, tested on the other 12): 9.1% BD-rate at equal SSIM, checked independently in libjpeg with scikit-image. mozjpeg's tables get 2.7%. See image 1.
• Google's WebRTC VAD (every constant in vad_core.c, tuned on LibriParty dev): F1 0.810 to 0.859 on LibriParty eval. See image 2.

Limits:

• No JPEG gain below 1 bpp.
• Best for tens to hundreds of constants, not large weight sets.

Code (MIT): https://github.com/devanmolsharma/learnc ·