Export libtorch model which support variable input tensor

deployment
1

Do libtorch model support variable size of the tensor? Trying to train a classification on medical images(CT scans), but the number of slices are different for each file([512, 512, variable]). Is it possible to export the model witch support input of [1, variable, 512, 512]?

Every examples of libtorch use fix size input before export the model

img = torch.zeros((1,3,16,112,112)) #64,3,16,112,112; 16, 3, 1, 112, 112
traced_script_module.save("video_r3d_18.pt")

Thanks

2

Yes, it should be possible via torch.jit.script. Tracing via torch.jit.trace uses the passed input to trace all operations and might fail, if you are using different inputs in the deployment use case.

1 Like