Two questions on creating a backwards pre-trained AWD LSTM model

fastai
21

Yes.

Remember that the actually flipping of the tokens happens during mini-batch creation (see the pad_collate function here). So one way of verifying this is by creating a batch and looking at the tokens … for example:

b = next(iter(data.train_dl))
print(b[0])
1 Like
22

I followed this and then used the following to confirm.

vocab = data.vocab

b = next(iter(data.train_dl))
print(b[0][0])

then paste the output of b into this:

vocab.textify()

This confirmed for me that the text was indeed being presented backwards.