Implementation of Transformer model using Tensorflow

• Prepare a directory containing tokenized training data (e.g. train_ja, train_en), validation data (e.g. validate_ja, validate_en) and vocabulary files (e.g. vocab_ja, vocab_en). I used RakutenMA to tokenize Japanese text and nltk to tokenize English text.

• Modify config.py to change hyperparameters.
• Run python3 train.py -d *data_directory* -l *source_language-target_language* -m *model_number* to start training.
• Press Ctrl-c to stop training.

• Modify beam_size, coverage_penalty_weight, length_penalty_weight in config.py if necessary.
• Run python3 translate.py -d *data_directory* -l *source_language-target_language* -m *model_number*. An image of attention_weights is plotted after each translation.

• Run python3 server.py -d *data_directory* -l *source_language-target_language* -m *model_number* -i *ip_address* -p *port_number* on a server.
• Put client.py and ipag.ttf file in the same directory. Run python3 client.py -i *server_ip_address* -p *server_port_number*. Add --plot if you want to plot attention_weights.