|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| """Pretrains a recurrent language model.
|
|
|
| Computational time:
|
| 2 days to train 100000 steps on 1 layer 1024 hidden units LSTM,
|
| 256 embeddings, 400 truncated BP, 256 minibatch and on single GPU (Pascal
|
| Titan X, cuDNNv5).
|
| """
|
| from __future__ import absolute_import
|
| from __future__ import division
|
| from __future__ import print_function
|
|
|
|
|
|
|
| import tensorflow as tf
|
|
|
| import graphs
|
| import train_utils
|
|
|
| FLAGS = tf.app.flags.FLAGS
|
|
|
|
|
| def main(_):
|
| """Trains Language Model."""
|
| tf.logging.set_verbosity(tf.logging.INFO)
|
| with tf.device(tf.train.replica_device_setter(FLAGS.ps_tasks)):
|
| model = graphs.get_model()
|
| train_op, loss, global_step = model.language_model_training()
|
| train_utils.run_training(train_op, loss, global_step)
|
|
|
|
|
| if __name__ == '__main__':
|
| tf.app.run()
|
|
|