Tensorflow: how to save/restore a model?

Question

After you train a model in Tensorflow:

1. How do you save the trained model?
2. How do you later restore this saved model?

Answer 1

If it is an internally saved model, you just specify a restorer for all variables as

restorer = tf.train.Saver(tf.all_variables())

and use it to restore variables in a current session:

restorer.restore(self._sess, model_file)

For the external model you need to specify the mapping from the its variable names to your variable names. You can view the model variable names using the command

python /path/to/tensorflow/tensorflow/python/tools/inspect_checkpoint.py --file_name=/path/to/pretrained_model/model.ckpt

The inspect_checkpoint.py script can be found in './tensorflow/python/tools' folder of the Tensorflow source.

To specify the mapping, you can use my Tensorflow-Worklab, which contains a set of classes and scripts to train and retrain different models. It includes an example of retraining ResNet models, located here