keras-cv icon indicating copy to clipboard operation
keras-cv copied to clipboard

Adding RegNets to keras_cv

Open AdityaKane2001 opened this issue 2 years ago • 4 comments

@sayakpaul @LukeWood

Added RegNets to keras_cv as per discussions in #679, #530 , #552 , #590

/auto Closes #530

AdityaKane2001 avatar Aug 29 '22 09:08 AdityaKane2001

@AdityaKane2001 can you fix code formatting and tests?

LukeWood avatar Sep 01 '22 18:09 LukeWood

@LukeWood could you take a look again? I believe the tests were failing due to timeout so I split the Regnet tests into two files.

AdityaKane2001 avatar Sep 02 '22 09:09 AdityaKane2001

/gcbrun

LukeWood avatar Sep 17 '22 22:09 LukeWood

/gcbrun

LukeWood avatar Sep 22 '22 21:09 LukeWood

Is this PR ready to merge?

qlzh727 avatar Oct 21 '22 18:10 qlzh727

Is this PR ready to merge?

Where is the training script?

bhack avatar Oct 21 '22 18:10 bhack

Where is the training script?

Here they go: https://github.com/AdityaKane2001/regnets_trainer

DavidLandup0 avatar Nov 09 '22 14:11 DavidLandup0

@qlzh727 @LukeWood

Could you please take a look at this one?

AdityaKane2001 avatar Nov 15 '22 06:11 AdityaKane2001

It would be interesting to benchmark the basic trainer and the regnets trainer on this arch. I'd expect the specialized version to outperform the general one, but we have been toying with the idea of doing a general script that's "good enough" vs the idea of doing specialized scripts for each arch, if I'm not mistaken?

DavidLandup0 avatar Nov 15 '22 16:11 DavidLandup0

@DavidLandup0

I think the main focus of having training scripts is reproducibility. They have mentioned it in a PR or RFC, I believe. @ianstenbit please correct me if I'm wrong.

AdityaKane2001 avatar Nov 15 '22 16:11 AdityaKane2001

@DavidLandup0 @AdityaKane2001 you are both correct WRT the purpose of our training script.

We want reproducible weights, and we also want a training script that is reasonably generalizable and demonstrates the use of KerasCV training utilities like image augmentation layers. I think in the long run, we'll end up having more bespoke training scripts as many models require specific techniques to achieve SOTA.

I will definitely run the basic_training script against at least one of these models as a benchmark.

@AdityaKane2001 -- for migrating your training script to KerasCV, a few questions:

  • is it feasible to update the script to use KerasCV preprocessing layers and follow the general high-level pattern of basic_training?
  • does basic_training provide most of the functionality that your script used? If not, might it make sense to update basic_training?

ianstenbit avatar Nov 15 '22 17:11 ianstenbit

/gcbrun

ianstenbit avatar Nov 15 '22 17:11 ianstenbit

@ianstenbit

I'll take a look and raise a PR, most likely towards the end of this week.

AdityaKane2001 avatar Nov 15 '22 17:11 AdityaKane2001

@ianstenbit

I'll take a look and raise a PR, most likely towards the end of this week.

Sounds great -- thank you!

ianstenbit avatar Nov 15 '22 17:11 ianstenbit

@ianstenbit

The accelerator test failed - is that expected?

AdityaKane2001 avatar Nov 16 '22 19:11 AdityaKane2001

I guess the models tests have passed, other tests have failed due to timeout.

AdityaKane2001 avatar Nov 16 '22 19:11 AdityaKane2001

@ianstenbit

The accelerator test failed - is that expected?

The tests all passed, but it timed out. We've been having some issues with this lately and it's not caused by your PR.

ianstenbit avatar Nov 16 '22 19:11 ianstenbit