Update nlp.py

[todtom]

Fix some bugs like tensorflow/tensorflow#5118.

[todtom]

I run it for the Chinese-English translation task. When use 'basic_tokenizer' it often reports the error like 'TypeError: cannot use a bytes pattern on a string-like object'. It works for me after fix the bugs. I'm not sure whether the modified 'initital_vocabulary' works for some other tokenizers

[zsdonghao]

@todtom Hi, thank you for your contribution. I wonder which TF version are you using? As the code seem to work fine previously, we need to make sure other users can also run their scripts. Thanks~

[wagamamaz]

@todtom after you fix the bugs, can it work well in English-French translation task?

[todtom]

@zsdonghao the enviroment as follows:

Windows10, python 3.5.2.3, tensorflow 0.12.0, tensorlayer 1.2.8

maybe the error caused by system? I haven't run it in Linux yet
@wagamamaz I train a toy 'en-fr' model which use 1k sentences for training data, it seems work well.

[zsdonghao]

@todtom great, it may caused by python3.5.