Reputation: 99
I know that there's a way to examine TF checkpoints (such as described in inspect_checkpoint.py).
However, is there a way to update certain weights in this checkpoint and write it back, without having to construct the whole model? For example, I want to quickly edit the weights of one of the layers and write it back, and don't want to write the code for the deep network this model corresponds to.
Upvotes: 1
Views: 2558
Reputation: 5808
Take a look at MetaGraph saving and loading. There are a few examples there; you should be able to define the model, save the MetaGraph, then load and modify it without re-defining the variables manually.
Upvotes: 1