Android and PyTorch

Mobile
1

I have trained a custom model and wanna add it to my android app. I saved the model with jet but when I run it on my android device I get this error:

com.facebook.jni.CppException: empty not implemented for TensorTypeSet(VariableTensorId, CUDATensorId) (empty at aten/src/ATen/Functions.h:3999)

Any fixes ?

2

Hi,

It seems like you’re trying to use a CUDA Tensor in your mobile app. I’m pretty sure that is not supported. You might want to make sure that your model is on the CPU before doing the export.

1 Like
4

Any code advice how to do that ?
When I trained my model I moved it to the cpu and then extracted with this code:

model.cpu()
example = torch.rand(1, 3, 224, 224)
traced_script_module = torch.jit.trace(model, example)
traced_script_module.save("app/src/main/assets/model.pt")
5

That looks right. And when you use this model you see the error above?
How do you create the input to the model on the mobile side?

6

Im using the sample PyTorchDemo app from pytorch. Just changed the model’s name in the code nothing else.

7

This is weird, maybe someone from the mobile team will have a better idea?

You can also try to double check your model to make sure it does not contain any weight that is on the gpu?

8

hmm okay thanks for your help.
is there any easy way how to check the location of the weights ?

9

You can print the .device field of any Tensor to check where it lives.
Also you can try to install a cpu-only build of pytorch and try to load your model in there. If there is cuda weights in it, it will fail as well, but with a friendlier error message.

10

Hi! I have the same problem, even with model.cpu().
Did you find a solution?