Autograd inside torchscript from java (libtorch)

jit
1

Hello, I am trying to use autograd to get gradients inside a torchscript file I am calling from java using libtorch but am running into an error “RuntimeError: element 0 of tensors does not require grad and does not have a grad_fn”

My torchscript model:

class MyModel(torch.nn.Module):
    def __init__(self):
        super().__init__()
    
    def forward(self, x):
        # Require a gradient
        x.requires_grad_()
        y = x.pow(2).sum()
        print(f'y.requires_grad: {y.requires_grad}')
        # Check if x requires a gradient
        print(f'x.requires_grad: {x.requires_grad}')
        dy_dx = torch.autograd.grad([y], [x])[0]
        return dy_dx

Saving with:

model_scripted = torch.jit.script(MyModel())
model_scripted.save("model.pt")

This throws no errors and returns the expected values when I reload and run the model inside python with:

model_loaded = torch.jit.load("model.pt")
dy_dx = model_loaded(x)
print(dy_dx)

However this fails and throws an error when I load inside java and try to run the model:

Traceback of TorchScript, original code (most recent call last):
  File "/var/folders/g9/ykm_xvjx25s8bq5bkfjqy5240000gn/T/ipykernel_39052/954115032.py", line 12, in forward
        # Check if x requires a gradient
        print(f'x.requires_grad: {x.requires_grad}')
        dy_dx = torch.autograd.grad([y], [x])[0]
                ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ <--- HERE
        return dy_dx
RuntimeError: element 0 of tensors does not require grad and does not have a grad_fn

Not sure why this is a problem when loading the torchscript into java but not a problem when loading into python.

2

I guess you are trying to deploy the model onto a mobile platform?
If so, I think gradient calculation is disabled by default as its main purpose is inference.

3

Hey @ptrblck thanks for the response. No, this is not to a mobile platform. Just trying to run the torchscript model inside a java application.

4

Ah OK, I’m not familiar with this workflow and don’t know what the Java env might set and if it would disable gradient calculation.

5

Does it do the same if you try to load it with the JavaCPP Presets for PyTorch?

6

Attempting to use the JavaCPP Presets I run into an issue with JNITorch:

Warning: Version of org.bytedeco:pytorch could not be found.
Warning: Version of org.bytedeco:openblas could not be found.
 Uncaught exception: exiting with status code 1
java.lang.UnsatisfiedLinkError: no jnitorch in java.library.path
7

If you’re trying this on Android, we’d first need to create builds for that platform.

8

@mhermon I am running into a similar issue. Did you find a solution?