FASTAI: 'LSTM' object has no attribute 'out' & attributeerror 'tuple' object has no attribute 'view'

Question

I received these errors in my NLP model by running fit_one_cycle on a Learner object. I only managed to come across This source here with someone coming across the same problem. It helped me find a solution for me, which I am sharing here in case someone else come across this problem.

The tuple issue is due to the fact that the forward method returns a tuple of three tensors in the FastAI example whereas just a tensor is expected in the __call__ method. The issued appeared after some update of the functionality, this seems to be the case

Answer 1

You can solve this in two ways:

1. Simply use a TextLearner object without specifying the callback methods rather than the Learner object (Then the callbacks should automatically be set and these are ajour with functionality). (Recommended)
2. Replace the RNNRegularizer with the correct callback function (requries more customization)