running tasks using saved PYTORCH model checkpoints as opposed to BLUEBERT TF checkpoints

[diegoolano]

hello, just a question if there is code for trying out the tasks via pytorch saved model checkpoints as opposed to TF checkpoints that were used by the original BlueBert evals? I was initially thinking of converting from a saved pytorch checkpoint to onnx and then to TF, but that will result in a .pb file which doesn't include the weights necessary to load I don't think. If not, do you have suggestions? thanks for your time.