What is the easiest way to get a PyTorch model into an android app?

1

I have my PyTorch model and I need to make an android app using it for a demo. I was thinking of somehow converting it into a tensorflow model by manually copying weights using a dictionary but is there any easier way? Thanks in advance

3 Likes
2

Yeah, you’ll have to use something like TensorFlow or Caffe2. We’re working on ways to improve exporting models, but for now you’ll have to:

  1. Get the weights via model.state_dict()
  2. Convert the Tensors to NumPy arrays
  3. Copy the weights into your TensorFlow or Caffe2 model
2 Likes
3

I’ve been having lots and lots of trouble trying to convert my model to tensorflow. Does torch-android still have support? I’m unable to build it for some reason.

4

maybe it would be better to have isolated model on backend of your application since you don’t have ability to run in directly on devices?

5

Could you give a walk-through example for those steps? Thanks!

6

@colesbury: Have there been any recent developments on this topic since you posted your last answer (Aug’17)? Or is the process still roughly the same? Thank you so much.