# Optionally, do length normalization.
batch_token_counts = [numpy.count_nonzero(s) for s in y_mask.T]
if normalization_alpha:
adjusted_lens = [n**normalization_alpha for n in batch_token_counts]
batch_ce_vals /= numpy.array(adjusted_lens)
ce_vals += list(batch_ce_vals)
token_counts += batch_token_counts
logging.info("Seen {}".format(len(ce_vals)))
assert len(ce_vals) == len(token_counts)
return ce_vals, token_counts
if __name__ == "__main__":
# Parse command-line arguments.
config = read_config_from_cmdline()
logging.info(config)
# TensorFlow 2.0 feature needed by ExponentialSmoothing.
tf.enable_resource_variables()
# Create the TensorFlow session.
tf_config = tf.ConfigProto()
tf_config.allow_soft_placement = True
# Train.
with tf.Session(config=tf_config) as sess:
train(config, sess)
model.inputs.y_mask: y_mask,
model.inputs.training: False}
batch_ce_vals = session.run(model.loss_per_sentence, feed_dict=feeds)
# Optionally, do length normalization.
batch_token_counts = [numpy.count_nonzero(s) for s in y_mask.T]
if normalization_alpha:
adjusted_lens = [n**normalization_alpha for n in batch_token_counts]
batch_ce_vals /= numpy.array(adjusted_lens)
ce_vals += list(batch_ce_vals)
token_counts += batch_token_counts
logging.info("Seen {}".format(len(ce_vals)))
assert len(ce_vals) == len(token_counts)
return ce_vals, token_counts
if __name__ == "__main__":
# Parse command-line arguments.
config = read_config_from_cmdline()
logging.info(config)
# Create the TensorFlow session.
tf_config = tf.ConfigProto()
tf_config.allow_soft_placement = True
# Train.
with tf.Session(config=tf_config) as sess:
train(config, sess)
