Tang-Poetry-Generator

In this project, a multi-layer Char-level (optional bidirectional) RNN, LSTM, GRU framework based on MXNET is built up. This project is inspired by the interactive deep learning course Dive into Deep Learning. Users can modify the model setting to create their own Tang-Poetry generator. A trained one-layer LSTM model is also provided. Users can apply this trained model to create five-characters eight-lines poems("五言律诗"), a special form of tang-poetry.

Data

The provided dataset came from chinese-poetry project, a great database contains almost all the ancient poetries in Chinese. The provided trained LSTM model is based on 14k five-characters eight-lines poems from the database. Users can change the data_cleaning.py file to create your own training data.

Model training

User can modify the train_model.py and run the file to train your own model. The default optimizer is "SGD", users can also change the optimizer to "Adam" or other optimizers supported by MXNET in the modeling.py. More specific parameters details are provided in the file. Below is the setting parameters for the trained one-layer LSTM model.

model_ind, num_hiddens, num_layer, bidirectional = "LSTM", 360, 1, False
ctx, lr, num_epochs, data_iteror = mx.gpu(), 1e2, 1500, data_iter # you can change the ctx to mx.cpu()
num_steps, batch_size, clipping_theta = 40, 128, 1e-2
pred_period, prefix = 50, '書湖河美人山萬鄉' # you can define your own prefix

Model Predict

User can make the trained model into a poetry generator by specifying the model parameters, the directory of the trained model and running the predict.py file.

One example poems generated by the model:

書信來相問，
湖上雲爲雨。
河漢日駸沒，
美人朝夕望。
人間無別離，
山中有幾處。
萬里長江水，
鄉關無夢歸。