An LSTM neural network model written in TensorFlow.
RUN
Just run train.py withpython train.py
And then use tensorboard --logdir=(tensorboard data file)
like tensorboard --logdir=2018-04-21 to launch TensorBoard in your browser.
Also, you can use automatically saved model to run a demonstration of prediction. Comment out (I googled this word) the training code in if __name__ == "__main__":, with only demonstrate() left. It is surprising that the result seems good.
PROBLEMS
- It is weird that accuracy of test set is much higher than that of train set, 70%+ and 40%+ respectively. I have looked it upon google and zhihu, it is said that the reason is that the amount of train set is too big and it will take a while for training to make it act as normal. And right now I am working on it.
- The macro-F1 is really low.
THANKS
@PTSTS for dataset and advice for model.