Thomas Werkmeister
Thomas Werkmeister
Hey @dakshvar22 I think you could review the current status now. I solved the feature length problem by calculating the IDs adhoc and passing on some of the labelling config...
I am thinking, potentially it's also a matter of the label batch size. training with 6percent of data is working on main, and the loss is going down. Given that...
Have found a bug, there was a mistake in the indentation from copy and pasting your code. And It would only fill a single embedding instead of all sampled embeddings....
Ok loss is looking good now during training! rasa test results look v similar to main
One thing that is still at the back of my mind and might be worth addressing: We make sure to sample negative candidates according to the proper distribution of label...
> > On top of that, the DPLL does not sample uniquely, so it might sample the same labels multiple times. > > Maybe better to replace the sampling inside...
Ok this is failing bc some vectors were hard coded in the tests ... This might be remind us that we need to signal to our users that a retraining...
superseded by #11598
closing as superseded by #11628
Hey there, this is Thomas from Berlin. I am a software/ML engineer with some social science training and got into CI after reading the book of why. It seems to...