💬 Open source machine learning framework to automate text- and voice-based conversations: NLU, dialogue management, connect to Slack, Facebook, and more - Create chatbots and voice assistants
added config var for "drop small last batch" for DIET which is off by default and results in the original batching logic
turned off the logic permanently for TED- and UnexpecTEDIntentPolicy
added tests that batching logic reacts to the parameter
did not add test for UnexpecTEDIntentPolicy not failing because it was way to complex to create the needed precomputation objects to really simulate a full training. In a naive test the policy did not fail at all. Reproduced the error locally as follows and confirmed that it is now gone
Error reproduction
initialize a new bot repo using rasa init
add UnexpecTEDIntentPolicy to the policies, with batch size 5
remove rules
replace stories.yml with the yml below containing 12 stories that will lead to two being dropped with batch size 5 and triggering the problem
run rasa train --force with and without the changes
Proposed changes:
Error reproduction
rasa init
rasa train --force
with and without the changesStories
Status (please check what you already did):
black
(please check Readme for instructions)