How to load checkpoint as an independent TF/Keras model for utilizing?

[xinghua-qu]

Hi,

Is there any example code about how to load the checkpoint (https://github.com/google/dopamine/tree/master/docs#downloads) that can be utilized independently? For instance, I use openAI gym to create an environment and utilize your checkpoint model to predict actions.

Many thanks and stay safe.

[nhewitt99]

I've been working on the same thing. Try this, changing the paths and the "24" (the iteration checkpoint to use) as necessary:

import tensorflow.compat.v1 as tf
with tf.Session() as sess:
    saver = tf.train.import_meta_graph('{path/to/base/dir}/checkpoints/tf_ckpt-24.meta', clear_devices=True)
    saver.restore(sess, tf.train.latest_checkpoint('{path/to/base/dir}/checkpoints'))

Using Dopamine is my first trial-by-fire with Tensorflow, so I'm not sure of meaningful next steps yet after opening the session, but hopefully this first step can help you some.