Got stuck in TorchScript Trace

jit
1

Hi,

Basically, I have built the deep learning model which takes dynamic inputs with different shapes,

so how do I trace the model here?

usually, we trace the model using the below command

example = torch.rand(1, 3, 224, 224)
traced_script_module = torch.jit.trace(model, example)

but then, in my scenario, I pass dynamic input ndarray with different shapes
eg:

1. example = torch.rand(1, 3, 224, 224)
2. example = torch.rand(1, 3, 224, 550)
3. example = torch.rand(1, 3, 640, 480)

could you please help me understand and how to implement this?

@ptrblck @ppwwyyxx @driazati
#jit #deployment

2

You would have to torch.jit.script the model instead of tracing.
The next stable release will use the nvfuser backend during scripting which has dynamic shape support.

3

thanks for the reply @ptrblck , I will try to save the model by torch.jit.script and get back to you.