Suggestion: video tutorial about model configuration

RLlib
1

I suggest to prepare video tutorial how to configure every aspect of the neural network model:
https://docs.ray.io/en/latest/rllib/rllib-models.html#default-model-config-settings

I see that there is a huge opportunity to configure the model as convolution model, recurrent model, model with attention, model with preprocessing etc. but many times it’s not clear how to do that.
Situation is that, I know I have very powerful tool in my hand but I’m not sure how to use it.

Images with graphs with model structure visualization can be very useful too. The graphs with proper descriptions can show how to turn on or turn off every part of the model by specified configuration value. The graphs can show which options can be possible to use together.

I suppose that the majority of the ray users are familiar with PyTorch or Tensorflow neural network basic architectures (dense network, RNN, CNN), so the tool like keras model.summary() or better tf.keras.utils.plot_model(model, to_file=dot_img_file, show_shapes=True) directly in rllib can be a good functionality.

1 Like
2

Hey @Peter_Pirog ,

This is a great idea and you should add it to our office hours document. In the office hours to come, we will regularly go over such topics capture the whole thing on youtube.
But: @kourosh is currently re-designing lots of things that have to do with how you formulate models. So we might do this first and have @kourosh introduce workflows in an instructional video later on!

Until then: Most of these config settings should be covered by our examples and tuned yaml files so you can at least get a sense of what is possible. If in doubt, you can write your own model with the ModelV2 API.

Cheers