Rasa: Training data generator hangs in Interactive Learning

[tahamr83]

Rasa version:
1.0.2
Python version:
3.6.6
Operating system (windows, osx, ...):
MacOS
Issue:
So the issue I am facing while migrating to Rasa 1.0.x is that during interactive learning
training.load_data() runs indefinitely. This was seemingly not an issue in previous rasa version, since I have been using the interactive learning feature a lot (RASA 0.14.x and 0.15.x). My stories are using checkpoints. But the checkpointing isn't complex. Each story starts with only one common checkpoint. But on the other hand I do have a lot of OR statements in my story. Why is this happening? Also regular story training doesn't hang. It does complete in a reasonable time.

Content of configuration file (config.yml):

# Configuration for Rasa NLU.
# https://rasa.com/docs/rasa/nlu/components/
language: en_core_web_md
pipeline:
- name: "SpacyNLP"
- name: "SpacyTokenizer"
- name: "SpacyFeaturizer"
- name: "RegexFeaturizer"
- name: "CRFEntityExtractor"
- name: "EntitySynonymMapper"
- name: "SklearnIntentClassifier"
- name: "DucklingHTTPExtractor"
  url: http://0.0.0.0:8000

# Configuration for Rasa Core.
# https://rasa.com/docs/rasa/core/policies/
policies:
  - name: FormPolicy
    priority: 7
  - name: FallbackPolicy
    fallback_action_name: 'action_custom_fallback'
    priority: 6
  - name: MappingPolicy
    priority: 3
  - name: MemoizationPolicy
    max_history: 5
    priority: 2
  - name: KerasPolicy
    epochs: 300
    max_history: 3
    priority: 1

[akelad]

@tahamr83 is there any chance you could share your stories?

[tahamr83]

Apologies for the delay. Here it is

stories.md.zip

Also another thing I noticed on our other bot is that the loading does not hang but the action execution in the interactive command shell is quite slow. These stories are simpler than the one above.I think it's somehow related to the trackers generated for visualization since if I change the interactive.py line 1335 to trackers = [] it becomes quite fast.
stories2.md.zip

I'd like to mention again that this was not an issue in the previous versions.

[akelad]

@imLew can you take a look at this?

[akelad]

@tabergma could you take a look at this instead?

[tabergma]

Thanks for pointing this out. I was able to reproduce the error. I will have a closer look at it and come back to you if I find anything.

[msamogh]

@tabergma Did you reproduce it on @tahamr83's stories or something simpler?

[tabergma]

@msamogh I took his stories.

[wochinge]

this might be related to https://github.com/RasaHQ/rasa/issues/4612

[wochinge]

@tmbo Is that the one you tackled recently?

[tmbo]

No not really. I think the issues in interactive learning are either in the visualisation (the graph, I think that might do another round of story step processing) or in the data loading for the training.

When I was working on this for the botsociety integration I fixed it by not using any training data generation there as it wasn't needed in that case.

[wochinge]

@tahamr83 Is this still happening with rasa 1.4? I tried to reproduce but couldn't. However, I found out that you currently can't disable the visualization properly.

[wochinge]

@tahamr83 Never mind. Found the error and will do a PR later. In the meanwhile you can fix it by using the --skip-visualization flag