This is a minimal example for reproducing a backprop error in the MinkowskiUnion layer in MinkowskiEngine. The model architecture is included in model.py, a single example input is included in data/example.pt, and the script example.py should reproduce the error when executed with Python. This minimal example was tested using the standard docker.io/pytorch/pytorch:1.6.0-cuda10.1-cudnn7-devel PyTorch image, with MinkowskiEngine v0.4.3 installed using pip.
This network was developed for a particle physics application, in which two independent pixel map representations are used simultaneously to classify an event. The two pixel maps are independently convolved, and then the Union function is used to combine them before a final set of convolutions.
Note: This example also includes an error in the MinkowskiChannelwiseConvolution module. This layer throws an error when included in the model, but since it convolves in place without changing the dimensionality of the tensor, it has been commented out in the model file – this second error can be reproduced by uncommenting lines 40-44 of model.py.