Closed sj-li closed 2 years ago
Hi, This is currently not supported. I have support for this in our internal version of the repo, but I need to test it intensively to verify that it doesn't blow up for people who have been using it like this already. If it was just training code, it would be just changing a couple of lines, but the problem is mostly in the "gather" operation generated in the onnx graph when we want to infer with TensorRT. I'll have support for this in the next release, along with the inference backend in C++ and more tasks. I'll leave this open an assign this to myself
Hi, is it possible that you provide your current fix? It should work for those not using TensorRT, if I understood correctly.
Hi,
The quickest hotfix for this is to change this line from:
stub = torch.zeros((1,
self.backbone.get_input_depth(),
self.ARCH["dataset"]["sensor"]["img_prop"]["height"],
self.ARCH["dataset"]["sensor"]["img_prop"]["width"]))
to
stub = torch.zeros((1,
5,
self.ARCH["dataset"]["sensor"]["img_prop"]["height"],
self.ARCH["dataset"]["sensor"]["img_prop"]["width"]))
Great, thank you!
Hi, I want to only only use xyz and remission data to train network, so I modify config file as:
And the program report:
Do you know what's the problem? Thanks!