Argmax code for inference

[pravn]

Could you put up code (pytorch/chainer) for the argmax calculation during inference time? From Tachibana-Uenoyama: https://arxiv.org/pdf/1710.08969.pdf

I am encountering the same problem during synthesis that the attention curve 'breaks' randomly during inference (training samples perfectly aligned). Your scheme of forcing alignment is absolutely brilliant , but coding is slightly problematic for batched attention vectors.