Just wondering how to save and restore trained deep generative models in zhusuan?

Thanks for the reply! Any examples to show that? If we can save and re

Basically it will look like this: <div class="highlight highlight-source-python no

save and restore models? about zhusuan HOT 4 CLOSED

ifangcheng commented on July 22, 2024

save and restore models?

from zhusuan.

Comments (4)

thjashin commented on July 22, 2024

Hi. Yes, it's just the same. The only thing you need to be careful about is that zs.reuse will add another level of scope.

from zhusuan.

ifangcheng commented on July 22, 2024

Thanks for the reply! Any examples to show that?
If we can save and restore the pre-trained deep generative models in zhusuan, is it possible to use the "fine tune" trick to do transfer learning for DGMs? e.g. whether we can fine tune a DGM trained on cifar10, to get a better DGM for a special domain with my own small dataset?

from zhusuan.

thjashin commented on July 22, 2024

Basically it will look like this:

    saver = tf.train.Saver(max_to_keep=10)

    with tf.Session() as sess:
        sess.run(tf.global_variables_initializer())

        # Restore from the latest checkpoint
        ckpt_file = tf.train.latest_checkpoint(result_path)
        begin_epoch = 1
        if ckpt_file is not None:
            logger.info('Restoring model from {}...'.format(ckpt_file))
            begin_epoch = int(ckpt_file.split('.')[-2]) + 1
            saver.restore(sess, ckpt_file)

        for epoch in range(begin_epoch, epochs + 1):
             ....

            if epoch % save_model_freq == 0:
                logger.info('Saving model...')
                save_path = os.path.join(result_path,
                                         "vae.epoch.{}.ckpt".format(epoch))
                if not os.path.exists(os.path.dirname(save_path)):
                    os.makedirs(os.path.dirname(save_path))
                saver.save(sess, save_path)
                logger.info('Done')