Validating models from v20.12 in the new release(v21.06)


There was a problem, all the models that I trained before the release of v21.06 now do not work out, I have to retrain. What could be the mistake?

I attach the text of the logs
problem (9.6 KB)

There has been a few changes to the neural network architectures and tensorflow variables in v21.06, and you won’t be able to load your checkpoints trained with the previous versions unless you manually make some changes to the SimNet source. Here the network expects a global_step variable of type int64, but the saved one in your network checkpoint is of type int32. We have changed the type of this variable from int32 to int64 in v.21.06.