EfficientNetV2 Functional subclass

[IMvision12]

https://colab.research.google.com/drive/1YDJTN4P4LArfWjXGW1u7ke_TTL5a8SfB?usp=sharing
Fix : #1425
@ianstenbit

[ianstenbit]

Looks good -- thank you!

[ianstenbit]

/gcbrun

[ianstenbit]

Test failure looks real. Please take a look -- thanks!

blocks = float(sum(args["num_repeat"] for args in blocks_args))
KeyError: 'num_repeat'

[IMvision12]

blocks_args = DEFAULT_BLOCKS_ARGS["efficientnetv2-b0"]
blocks = float(sum(args["num_repeat"] for args in blocks_args))
print(blocks)

Not sure why the test is failing, I tested the above code and it works fine

[ianstenbit]

@IMvision12 based on the error message (are you able to see it?), this is related to serialization / from_config. I'd guess the blocks_args param isn't being serialized correctly (possibly because in this case what's passed in is "default" but what we serialize is the full dictionary -- we could fix this by making a copy of the input blocks_args param and store that on self.blocks_args)

You should be able to run this test locally with PyTest to test further.

[IMvision12]

Oh, yeah i didn't see that

[IMvision12]

=========================================== test session starts ===========================================
platform win32 -- Python 3.9.10, pytest-7.2.2, pluggy-1.0.0
rootdir: F:\effi\keras-cv, configfile: setup.cfg
collected 6 items

keras_cv\models\efficientnet_v2_test.py .....s                                                       [100%]

====================================== 5 passed, 1 skipped in 14.60s ======================================

[ianstenbit]

/gcbrun

[IMvision12]

@ianstenbit, would we be interested in integrating SSD (Single Shot Detector) models to kcv? I was thinking about working on it.

[ianstenbit]

@ianstenbit, would we be interested in integrating SSD (Single Shot Detector) models to kcv? I was thinking about working on it.

I'd open an issue about it and tag LukeWood -- I think in general new OD models are something we're interested in, but prioritizing YOLOx and YOLOv8 first.