|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
"""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() |
|
|