search

AdityaKane2001 / regnety

Implementation of RegNetY in TensorFlow 2
Apache License 2.0
20 stars 1 forks source link

Changes in input pipeline #6

Closed AdityaKane2001 closed 3 years ago

AdityaKane2001 commented 3 years ago

This issue summarizes the tested changes in the input pipeline.

This issue can be more or less treated as a log file. This issue will be tracked by the project board in "Done" section.

Date Feature Commit Result (at the time of change) Retained? Comments
June 14 Vanilla, as in pycls https://github.com/AdityaKane2001/regnety/pull/1/commits/40f93df36d97c005ceb0eb8d8a0f6c66878dede4 NA No -
June 26 tf vision RandAugment https://github.com/AdityaKane2001/regnety/commit/3275b4db165379b1e0b62d4aeffa50aa2a6888d6 ~40imgs/sec No Non vectorized
June 26 Resize images to 512 https://github.com/AdityaKane2001/regnety/commit/25a75b5353b36a4cd4a8e642d249ebedd4c33319 ~40imgs/sec Yes Improved later
June 27 GCS, interleave, determinism=False https://github.com/AdityaKane2001/regnety/commit/80d558dbddbda788956231dcfc35ee4ec35ccc1e ~50imgs/sec Yes Retained until final version
June 28 Vectorized random_sized_crop https://github.com/AdityaKane2001/regnety/commit/9365f3a25363b0e5b933a46f0adf93a6c17d764f ~60imgs/sec No Removed in favor of random_crop
June 29 WeakRandAugment - All augmentations https://github.com/AdityaKane2001/regnety/commit/4394214e0235c374344e3e52bad01c20420d313d ~50imgs/sec No All numbers henceforth are reported with cache
July 2 Caching NA, removing and adding back frequently NA Yes -
July 3 Using AI Platform notebooks NA ~110imgs/sec Yes n1-standard-8 with Tesla T4
July 5 Removing redundant augs: invert, sharpen, cutout, color_degrade, shear, translate https://github.com/AdityaKane2001/regnety/commit/897cdb070fbcc9626eeb5cc9a2288b9698041453 ~200imgs/sec Yes Best performing, as least augmentations and cache were involved
July 7 Added back random rotate and crop https://github.com/AdityaKane2001/regnety/commit/37f1607deef583472e8acb41b4abe0b0c60d9159 ~140imgs/sec Yes Final speed, with random rotate and random crop
August 1 Added Inception crop, pca jitter Numerous ~1800 (on final hardware) imgs/sec Yes Final, in the repo

IMPORTANT: The first few versions were extremely slow and thus the numbers are approximate. The later numbers are relatively more accurate due to use of Tensorboard and AI Platform notebooks.

I will update this table as and when required.

AdityaKane2001 commented 3 years ago

Closing this one as input pipeline is now freezed.