What is your collate function doing? In most cases, you can use before_batch instead of the collate function (for instance all the padding). Otherwise, you need to pass your own create_batch indeed.
Actually you’re right … was able to do with with before_batch (had a bug in my code).
Kinda related but is there a way to tell my Learner that there is only ONE input even if my DataBlock is setup so that there are TWO? I’m actually merging the two inputs from my DataBlock into a single input in a function I pass to before_batch. Unfortunately, Learner still expects a nn.Module with a forward() function that includes TWO arguments.
You can set learn.n_inp=1 to force the behavior, but it will probably mess with the show methods, as everything is setup to have the number of inputs of your data be the same for the Learner.