I believe the only thing missing is a PyTorch version of the model for the regression test.

I tried dp convert-backend ./W.pb ./W.pth but it failed with the following error:

KeyError: "The name 'train_attr/training_script:0' refers to a Tensor which does not exist. The operation, 'train_attr/training_script', does not exist in the graph."

@njzjz, could you help?

This model seems to be generated by the old version and doesn't store the training parameters. This is necessary to convert between backends, though.

Ok, I'm now using examples/water/dpa2/frozen_model.pth.

Unfortunately, it still doesn't work with leak sanitizer - presumably because PyTorch does some weird dynamic linking.