prajitr / fast-pixel-cnn Goto Github PK

I have no idea of this line about sampling from softmax. Can you give me a help?

sel = tf.one_hot(tf.argmax(logit_probs - tf.log(-tf.log(tf.random_uniform(logit_probs.get_shape(), minval=1e-5, maxval=1. - 1e-5, seed=seed))), 3), depth=nr_mix, dtype=tf.float32)

I want to commend the team on their fantastic work. This addition to the PixelCNN++ makes it much more usable for integrators like myself.

I've been attempting to use your package to produce conditioned images. That is, I used the CIFAR-10 pre-trained model available on PixelCNN++, and passed a one-hot vector to 'h' in your model. I receive the following error:

NotFoundError (see above for traceback): Tensor name "model/conditional_weights_11/hw/ExponentialMovingAverage" not found in checkpoint files /home/da/tensorflow3/pixel-cnn/params_cifar.ckpt

Is this because the pretrained model provided by PixelCNN++ was not trained with the conditional labels? I can see you're restoring variables using variables_to_restore around like 70 of generate.py, but the behavior of this function is complex. As an example, if I restore all variables instead of using the method you employ, the model cannot load v_stack_cache.

Your advice is greatly appreciated.

hello,I have a problem when I download the pretrain data,I found when I download 45%,the Download speed dropped to 0. Can you give me a help?