config.py

dict(
    z = 20, # the dimension of latent variable
    batch_size = 16, # mini-batch size
    label_size = 2, # the number of discourse relation label, we model it as one-vs-all
    L = 1, # the sample number
    learning_rate = 0.001, # lerning rate for Adam
    max_iter = 100, # maximum iteration number, here we set to 100, we use 1000 in our paper
    clip_c = 1., # gradient clip
    is_load = False, # whether load model parameters, during testing, this should be corrected
    seed = 1473769786, # the seed for random, same seed for same results.

    train_arg1 = './data/com/train.arg1.dta.pkl',
    train_arg2 = './data/com/train.arg2.dta.pkl',
    train_lbl = './data/com/train.lbl.dta.pkl',
    dev_arg1 = './data/com/dev.arg1.dta.pkl',
    dev_arg2 = './data/com/dev.arg2.dta.pkl',
    dev_lbl = './data/com/dev.lbl.dta.pkl',
    test_arg1 = './data/com/test.arg1.dta.pkl',
    test_arg2 = './data/com/test.arg2.dta.pkl',
    test_lbl = './data/com/test.lbl.dta.pkl',

    mode = 'train', # train or test, if test, make the is_load True
)