tf-lstm-char-cnn icon indicating copy to clipboard operation
tf-lstm-char-cnn copied to clipboard

Published 20 hours ago • verified publisher icon mkroutikov

fixing non-determenistic op ordering reported as issue #20

Open mkroutikov opened this issue 6 years ago • 0 comments

@mkudinov please review.

This is somewhat theoretical fix.

  1. removed initial cleanup of padding in char embedding matrix from train.py. This should not affect training as each training step will be clearing it anyway
  2. used tf.control_dependencies to schedule embedding cleanup just after gradients were applied (in train.py)

Please, confirm it is working correctly now.

mkroutikov avatar May 15 '18 12:05 mkroutikov