Now, some parts of the model are not initialized due to the lazy @tensor annotations. Therefore, their initialization must be forced for example by printing out the affected tensors.
This must be either avoided or moved to the code of the runners, instead of the __init__ methods of the model parts.
We need to make sure that all output-generating model parts call their output in their output in the __init__ method, so at least the inference part of the graph is always initialized.
Now, some parts of the model are not initialized due to the lazy
@tensor
annotations. Therefore, their initialization must be forced for example by printing out the affected tensors.This must be either avoided or moved to the code of the runners, instead of the
__init__
methods of the model parts.