A-project-for-gogame-have-bug-need-to-solve

The project which is to imitate the gogame. I create a cnn model you can see the CnnModel file is used for outputing a rough possibilties of the next step. The next step is to combine the data of the rough possibilties and past chess moves send to the rnn model to train. In the rnn model the context_attention model will let the rnn model put attention to the rough possibilites and concat the hidden layer of lstm and output the context and then send the context and a previous step to the lstm layer and the lstm will memory the past each steps by using a loop. The bug is that using the deep lstm or rnn network and sgd for training the gradient will disappear.