apeterswu / rl4nmt Goto Github PK

I was looking through the code a bit and found that delta_reward had a typo here:

Don't think this affects the experiments, but the unused_kwargs is pretty dangerous so worth checking.

I went through tensor2tensor and looked for data generation but couldn't find the problem which will prep the data for training for your code, zh-en is not supported. Can you please help?
The script says that putting _rev will make the data for zh-en task but it does not work.Only en-zh work, So ?

In the paper you maintained about verifying the baseline reward approach and the baseline reward estimator was pretrained. But i couldn't see the pretraining code !. Can you help me with the pretraining code ??

请问这个代码的环境依赖是什么呢？
我的环境是python3.6 tensorflow1.6 tensor2tensor1.5 代码无法运行

As mentioned in your paper the cumulative future reward is used to update the policy at timestep t. My understand is that this is done at the following line inside the compute_sentence_bleu function.

It seems that delta_results[::-1] reverses the batch dimension instead of the time dimension. Shouldn't this be?:

delta_results = delta_results[:, ::-1].cumsum(axis=1)[:, ::-1]

First of all, thanks for putting this repo up, we enjoy your paper and would like to reproduce it in our lab. Have a few questions:

1. is this repo complete? I see only training scripts, not sure how the data is generated. And not sure if I find the RL models and how they are integrated... maybe I am missing something?
2. I see you have a complete tensor2tensor source code copied in this repo instead of referencing to the t2t official package. Does this mean you have changed t2t code base in some way? or we can safely use t2t official package (perhaps with an older version)?
3. Would really appreciate if you can provide a more detailed example on how to produce the unified model (zhen_bi_src_tgt_mono) from this code base.