Will torch.jit.unused support normal object class?

jit
1

It seems that torch.jit.unused does not support normal object class. Is there any plans to support it? For now, the following codes will cause an error: 'torch.jit.frontend.UnsupportedNodeError: with statements aren’t supported:'

@torch.jit.script
class MyObject(object):
    def __init__(self):
        super().__init__()

    @torch.jit.unused
    def unsupported(self):
        with torch.no_grad():
            print('This is not supported by jit.')

    def supported(self, x):
        print('This is supported by jit.')

By the way, could anyone give some advices to handle this situation where classes with torchscript-unsupported attributes are used during torchscript inference process.

2

Right now everything that is in a @torch.jit.scripted class has to be TorchScript compatible, but we’re working on fixing it so @ignore and @unused work as they should. The best way around it for now is probably to use free-standing functions (which you can @ignore) that take the object as the first argument

3

Got it, thank you very much:)

4

@driazati Thanks for the answer. Is there an issue where we can track the progress of the Python Object class @ignore and @unused support?

5

This feature will be released in TensorFlow 1.7. The pull request can be found here.

6

I thought TF1.7 was already released in 2018 :wink:

1 Like
7

@ptrblck Haha :rofl:, typo, pytorch 1.7 ofcourse :grin:.

2 Likes