Add ResNeSt to keras.applications

[innat]

System information.

TensorFlow version (you are using): 2.7 Are you willing to contribute it (Yes/No): No

Describe the feature and the current behavior/state.

Requesting to add the following model -

• Paper / Model: ResNeSt: Split-Attention Networ
• Official PyTorch Code: https://github.com/zhanghang1989/ResNeSt
• Citation (till now): ~ 400

3rd party code:

• TensorFlow : https://github.com/QiaoranC/tf_ResNeSt_RegNet_model
• Jax: https://github.com/n2cholas/jax-resnet/

Will this change the current api? How? Yes.

tensorflow.keras.applications.ResNeSt50
tensorflow.keras.applications.ResNeSt101
tensorflow.keras.applications.ResNeSt200
tensorflow.keras.applications.ResNeSt269

Who will benefit from this feature? TensorFlow/Keras users.

(It's better to have it in KerasCV`).

[dathudeptrai]

@innat Please refer this implementation (https://github.com/RichardXiao13/TensorFlow-ResNets).

[innat]

@dathudeptrai no, it's a 3rd party implementation and comparatively much better tf-implementation is already mentioned above.

[dathudeptrai]

@innat The implementation I mentioned above has almost the same structure and coding style as the models in tf.keras.applications

[innat]

@dathudeptrai you're missing the point. I'm not looking for model implementation. Please read the first post.

[rchao]

Triage note: We'll consider this once KerasCV is ready which we believe is a better place for the application to live.

[innat]

@rchao is there any possibility that all the models under keras.applicaitons will move to the keras-cv project?

[innat]

@QiaoranC @RichardXiao13

[rchao]

@LukeWood do you have an answer to the keras-cv question?

[LukeWood]

@LukeWood do you have an answer to the keras-cv question?

if this is referring to: "@rchao is there any possibility that all the models under keras.applicaitons will move to the keras-cv project?"

I believe that the answer to this is that in the long term: yes. In the short/mid term: probably no as it would break many users.

The code of applications will likely exist in both locations for a period of time, with some small differences (i.e. defaulting the weights argument to None for applications instead of imagenet)

[LukeWood]

I'd personally be in favor of including this in KerasCV instead of keras.applications. This would allow us to avoid duplicating some code. @fchollet may have some thoughts here as to where the best place to contribute this in the near term would be.