Passing custom features to TED policy

Rasa Open Source
1

Is it possible to pass dense features from a custom featurizer to the TED policy?

2

Hi @kearnsw

Yes, in an end-to-end setting, TED should use the featurizers of your NLU pipeline, so if you create a custom featurizer and add it there, it should be used by TED.

3

Hi @j.mosig

Thanks for responding. Is there a way for me to inspect the model to ensure the features are being passed? For example, through the debug log.

4

I have to check that… But first, did you install Rasa from source?

5

I have both installed in different environments. I can change the source code to allow this, just wondering if there’s a way to use the existing TED policy or if I should raise an issue on GitHub.

6

inspect the model to ensure the features are being passed

Do I get this right that you want to see if the feature arrays seen by TED include the features of you component? I don’t think you can output this by default (e.g. -vv doesn’t print what featurizers are seen by TED). But If you run from code, you can always use a print or tf.print statement. The codebase has changed a lot recently, so you may need to look at different places, but you may start looking around here. Does this help?

7

Yes, I’d like to see which featurizers and features are used by TED. I had traced the code and it appeared that the features from each featurizer are accessed through a dictionary via a key for each featurizer name. I’ll go ahead and open a GitHub issue and submit a PR that includes this additional logging at the debug level.

1 Like