G2P

Training

1. Install requirements (see requirements.txt) and get G2P-my_LSTM-act1_save_new_general.py to run locally using python3

2. Generate train,dev,test data and place it in a subfolder in data (say you call this MYDATA). Two sample folders are given in data.

   • NB All data must be in tab seperated format, see examples in data. If necessary, run python3 renderTest.py < test.1k.0 to convert your test data in the same format.

3. In G2P-my_LSTM-act1_save_new_general.py and neuralnets/BiLSTM_proper.py modify paths indicated by "MODIFY" to your local settings.

4. Run G2P-my_LSTM-act1_save_new_general.py with some arguments. Arguments include hyperparameters, but also

   • the path to the train,dev,test split you want to train (indicated by MYDATA)
   • the embeddings you want to use for training (make a new copy of embeddings for each run you make in case your train,dev,test splits differ).

   Sample runs are given in allvsall_odd.

Evaluation

1. Outputs are stored in the path indicated in neuralnets/BiLSTM_proper.py under "MODIFY".

2. Run Eval/eval2.py using the respective output files as arguments.