Reputation: 13
My goal is save an tensorflow model and restore it, but with different shape variables (I want add new classes and new training examples). In this question it's explained that the option reshape of Saver works if the shapes have the same total number of elements, and it's necessary to build a new graph. My question is: Is possible to build a new graph with the old trained model? So that resume the training with the new class.
Upvotes: 0
Views: 219
Reputation: 1080
You can initialize shared part of the model from old checkpoint. Please take a look to: tf.contrib.framework.init_from_checkpoint
Upvotes: 1