theoqian

https://github.com/ne7ermore/deeping-flow/blob/433296c0a2cd1ebd6db8524aa9d20bbc59d0f31f/deep-reinforced-sum-model/attention.py#L46 这里说dec_out的形状是bsz * time * dec_hsz。而dec_out是model.py中tf.nn.dynamic_rnn的输出。因为你每个时间步的对dynamic_rnn的输入形状是bsz * 1 * emb_dim，所以dec_out的形状应该是bsz * 1 * dec_hsz，他不会保存前几个时间步的状态。麻烦您看看我的疑问是否正确。

I try to train a mask_align model with default config in the repo (only change data paths) and DE-EN training data from https://github.com/lilt/alignment-scripts. In some of training steps the losses...