keras-cv
keras-cv copied to clipboard
Adding RegNets to keras_cv
@sayakpaul @LukeWood
Added RegNets to keras_cv as per discussions in #679, #530 , #552 , #590
/auto Closes #530
@AdityaKane2001 can you fix code formatting and tests?
@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.
/gcbrun
/gcbrun
Is this PR ready to merge?
Is this PR ready to merge?
Where is the training script?
Where is the training script?
Here they go: https://github.com/AdityaKane2001/regnets_trainer
@qlzh727 @LukeWood
Could you please take a look at this one?
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
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.
@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 updatebasic_training
?
/gcbrun
@ianstenbit
I'll take a look and raise a PR, most likely towards the end of this week.
@ianstenbit
I'll take a look and raise a PR, most likely towards the end of this week.
Sounds great -- thank you!
@ianstenbit
The accelerator test failed - is that expected?
I guess the models tests have passed, other tests have failed due to timeout.
@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.