Lesson 10: Using salesforce/pytorch-qrnn vs fast.ai

I was thinking of using the salesforce/pytorch-qrnn implementation of the AWD LSTM model (mainly because of multi-GPU support) and was wondering if anyone has already tried it and what the main drawbacks would be of using it in place of the fast.ai implementation.

I’d suggest instead adding QRNN support to the fastai version. My guess is it’s just a single line of code to change. The benefits are:

  • AWD LSTM isn’t updated any more (smerity has started a new company in music generation, and doesn’t seem likely to keep AWD LSTM updated now - although I have sent him the info on fixing it for pytorch 0.4 so maybe he’ll get on to it sometime)
  • fastai adds important stuff like classification
  • fastai is integrated with fastai.text, ModelData, etc