Python version

[Silencezjl]

Why do you use Python 3.6? Python 3.6 does not support TensorFlow well. I have so many warnings in Python 3.6.

I use Python 3.5 to run all your code. But I got a bad F1-Measure about 60 ~~

Does anybody else really get a 90 points F1-Measure?????

[Silencezjl]

Finally, I use python 3.6 with warnings to run your code. And I got a 92.32 points F1-Measure. It's so magical!!!!!!

[hankcs]

Yes, Python 3.5 won't work and outputs low scores. Currently the code stops training when score is lower than 90, in order to save your time of tuning parameters. Stick to our documentation please (version must >= 3.6).

Generally speaking, in iteration 1 you'll get a dev score around 92.3 with pku cnn, that's exactly what you reported. The final dev score will be around 96.1, and final test score will be around 95.2. You can use the bundled official sighan score script to verify our result like:

$ python3 official_scorer.py --gold-file data/pku/raw/test.txt --test-out result/pku/cnn/2017-11-10_02-30-25/test-out.txt 
Evaluating [pku] using official SIGHAN score script...
pku
=== F MEASURE:  0.953

I attached a training log file for your perusal.
info.log

[Silencezjl]

Thanks a lot!