transparency of CRDNN about speechbrain

[pplantinga]

Thanks for the feedback Ju-Chieh, its important that this model be relatively easy to understand. One limitation of defining this model in the recipe is that it cannot be reused across multiple recipes. Perhaps there's something else we could do to improve transparency.

One thing that could potentially help: instead of defining the blocks in three separate yaml files, we could put them all in the same yaml file. I think this would require re-writing the "ReplicateBlock" method in PR #79 based on the latest changes to the yaml (#101).

Do you think that putting all of the parameters in one yaml file would help?

[TParcollet]

I would say that it's already quite transparent ... What is your concern @jjery2243542 ?

[jjery2243542]

My concern is, the recipe might become difficult to modify if we use the whole block of module instead of declaring one by one from nnet. One has to look into the model and try to know how to override this yaml when they want to change the architecture.

[mravanelli]

That's a bit the idea of "lobes". Instead of redefining the same model every time in the experiment.py, we can put in "lobes" some kind of popular models and just use them. If users want to change the current model with another one in the lobes they can do it very easily. This way it is also easier to maintain the recipes: if we wanna modify something in the CRDNN model, for instance, we do it just once and not for all the recipes (that can be a lot at the end of the project). Note that using models in lobes is not mandatory and users can always define their model in experiment.py. But in our recipes we have to be consistent and the "lobes" approach is what we are currently using in all of them. @Gastron any other comment?

[jjery2243542]

I think you are right. Using a customized model is quite easy so we could use lobes to share some general structure.