added scheduled sampling for batch model

[patrick-dd]

I added scheduled sampling for the batched attention model.

I don't have cuda or sconce set up on my computer, but it runs very slowly on a cpu.

[koustuvsinha]

Hi @patrick-dd , couldn't understand your implementation of scheduled sampling. You are creating output_placeholder but you are not using it at all, you are passing the same single batched best word as the next input?

decoder_input = Variable(ni)