generated sample become zero

I have followed mater branch and trained like this
`$ python main.py --data=mnist --model=pixel_rnn`

But after about 400 epochs, the pixel vanished and the generated picture looks like a empty picture:

What's wrong with my procedure?

Start with main.py and reproduce the reported command, `python main.py --data=mnist --model=pixel_rnn`, after roughly 400 epochs. Inspect the training and sample-generation behavior for the point where pixels disappear; done means identifying a reproducible cause and restoring non-empty generated samples.

Written by the indexing model from the issue text.