how can I generate after training?

[daiwk]

how can I use the checkpoint to generate images?
e.g. I wanna
`
saver.restore(sess, ckt_path)

        print("restored")
        x, _ = self.dataset.test.next_batch(self.batch_size)
        feed_dict = {self.input_tensor: x}
        sess.run(init)

        print("testing.....")
        z_var = self.model.latent_dist.sample_prior(self.batch_size)
        fake_x, _ = self.model.generate(z_var)
        fake_d, _, fake_reg_z_dist_info, _ = self.model.discriminate(fake_x)
        generator_loss = - tf.reduce_mean(tf.log(fake_d + TINY))

        sess.run(init)

        ##print(sess.run(z_var, feed_dict)) ## ok
        print(sess.run(fake_x, feed_dict))`

but it says
tensorflow.python.framework.errors.FailedPreconditionError: Attempting to use uninitialized value fc_batch_norm_8/batch_norm/fc_batch_norm_8/batch_norm/moments/normalize/mean/ExponentialMovingAverage
[[Node: fc_batch_norm_8/batch_norm/fc_batch_norm_8/batch_norm/moments/normalize/mean/ExponentialMovingAverage/read = IdentityT=DT_FLOAT, _class=["loc:@fc_batch_norm_8/batch_norm/fc_batch_norm_8/batch_norm/moments/normalize/mean/ExponentialMovingAverage"], _device="/job:localhost/replica:0/task:0/cpu:0"]]
Caused by op u'fc_batch_norm_8/batch_norm/fc_batch_norm_8/batch_norm/moments/normalize/mean/ExponentialMovingAverage/read', defined at:

[NHDaly]

I'd love to hear the official answer to this question!

That said, I managed to get this to work by making this tiny change to infogan_trainer.py:
In 8fca21e, I changed:

fake_x, _ = self.model.generate(z_var)

to

self.fake_x, _ = self.model.generate(z_var)

so that the InfoGANTrainer holds a reference to the piece of the graph that generates fake data samples from the trained generator graph. This way, you can reuse the graph to generate data once it's trained.

---

To get access to the trained graph, I think you could probably do something similar to what you've done above by loading from a checkpoint, but instead in b4602c9 I also changed InfoGANTrainer.train() to accept a tf.Session() parameter, so that you can use it immediately without having to reload the weights from a checkpoint file.

---

So finally, for mnist, in bb8ad8e I've changed the last few lines to now look like this:

    sess = tf.Session()
    algo.train(sess)

    # Use the trained model to generate one batch of images!
    generated_images = sess.run(algo.fake_x)

    print("Saving generated images to " + gen_dir)

    for i,img in enumerate(generated_images):
        img_file = os.path.join(gen_dir, 'generated_%05d.png' % (i))
        scipy.misc.toimage(img.reshape(dataset.image_shape[:2]), cmin=0.0, cmax=1.0).save(img_file)

---

I hope this helps! 😸

[zdx3578]

can ref his code: https://github.com/RutgersHan/InfoGAN/

[wxdai]

@daiwk Have the same problem! Did you manage to solve it?