Autogenerate preset tables on keras.io

[mattdangerw]

This will be a fairly involved contribution, but one that will help out the project quite a bit!

We are hoping to generate our preset checkpoint tables on keras.io directly from our library.

There will be two main steps to this.

1. In this repo, we need to update our xx_presets.py files so they contain all the metadata they need to generate the documentation tables. We should add a "metadata" dictionary for each preset, which can contain the model "description" and "params" count, with the ability to add more metadata later.
2. On the keras.io repo, we need to add code to traverse through our model presets and render the metadata as a markdown table on the models page. I've written a starter commit showing how this could look here.

[mattdangerw]

Another useful code pointer, this script shows how you can programmatically walk through our presets. This will be useful for writing the rendering code on keras.io.

[durganshyadav]

can i contribute in this

[Cyber-Machine]

@mattdangerw I would like to work on this issue, I have gone through the model's folder and found out there are 6 models that contain <model>_presets.py, and each file includes a dictionary item as

backbone_presets = {
<model_name> : {
 "config" : {  <model_layers> },
 "preprocessing_config" : {},
  "description" : ( < model_desc > ),
  "weights_url" : <URL>,
  "weights_hash" : <HASH>,
  "spm_photo_url" : <URL> ,
  "spm_photo_hash" : <HASH>

 }
}

So in the first task, I am going to add metadata as:

backbone_presets = {
<model_name> : {
 "config" : {  <model_layers> },
 "preprocessing_config" : {},
  "description" : ( <model_desc> ),
  "metadata": {
    "description" : ( <model_desc> ),
    "params" : <no of parameters>
  },
  "weights_url" : <URL>,
  "weights_hash" : <HASH>,
  "spm_photo_url" : <URL> ,
  "spm_photo_hash" : <HASH>

 }
}

[mattdangerw]

Actually @Cyber-Machine is it ok if @durganshyadav take this one on? I referred @durganshyadav over to this issue on http://go/kerasnlp/issues/650, they have been looking for an issue for a bit.

We will continue opening contributions welcome issues regularly!

And thanks everyone for patience as we scale up to more contributors here! We are getting a lot more new contributors which is wonderful but will be a growing process for us 😄

[Cyber-Machine]

Sure @mattdangerw, I will be ready for making contributions! 🤗 😄

[mattdangerw]

Actually apologies, I just saw that @durganshyadav is assigned on #658

So in the interesting of spreading our new issue around, @Cyber-Machine let's leave this one for you!

Sorry about the confusion. @Cyber-Machine if you would like to take this on feel free. And please ping if any questions.

[Cyber-Machine]

Sure @mattdangerw, I will start working on this issue. Thanks!

[Cyber-Machine]

Hi @mattdangerw ,

This weekend I went through the code suggested by you and the Keras-io repo and made the appropriate changes in the xx_params.py files where backbone_presets and classifier_presets are defined.

Looking at the original site, the table also needed a model link along with description and params, so I added that in metadata except for Albert and DebertaV3.

I have shared my approach in this colab notebook and have made a pull request #690 for the changes I have made till now.

Could you review my PR and suggest changes that need to be made?