Sequential generation error

[carpedm20]

Current implementation successfully works as a autoencoder but it can't generate from the beginning, from a blank image. This error should be related to masking.

[carpedm20]

(Up: ground truth, Down: prediction)

Fix with 4c459c1

[carpedm20]

Even the one-shot recovering of the original image can be learned quickly as shown in previous comment, the sequential sampling of an image is still not working like igul222/pixel_rnn.

[carpedm20]

Update.

Trained model with python main.py --data=mnist --model=pixel_cnn