Better t2t-translate-all functionality (#676)

* wait until the directory is created (e.g. with t2t-translate-all)

* t2t-translate relies on the translated files to have correct mtime

"correct" here means the time of the checkpoint creation,
not time when the evaluation was done.
(In my original PR, --keep_timestamp was True by default,
but I agree this is a bit non-standard.)

Author: martinpopel

tensor2tensor/bin/t2t_bleu.py

@@ -57,6 +57,7 @@
 from __future__ import print_function
 
 import os
+import time
 
 # Dependency imports
 
@@ -110,6 +111,14 @@ def main(_):
     raise ValueError(
         "Either --translation or --translations_dir must be specified.")
   transl_dir = os.path.expanduser(FLAGS.translations_dir)
+  if not os.path.exists(transl_dir):
+    exit_time = time.time() + FLAGS.wait_minutes * 60
+    tf.logging.info("Translation dir %s does not exist, waiting till %s."
+                    % (transl_dir, time.asctime(time.localtime(exit_time))))
+    while not os.path.exists(transl_dir):
+      time.sleep(10)
+      if time.time() > exit_time:
+        raise ValueError("Translation dir %s does not exist" % transl_dir)
 
   last_step_file = os.path.join(FLAGS.event_dir, "last_evaluated_step.txt")
   if FLAGS.min_steps == -1:

tensor2tensor/bin/t2t_translate_all.py

@@ -97,7 +97,7 @@ def main(_):
           "--decode_hparams=beam_size={FLAGS.beam_size},alpha={FLAGS.alpha} "
           "--model={FLAGS.model} --hparams_set={FLAGS.hparams_set} "
           "--checkpoint_path={model.filename} --decode_from_file={source} "
-          "--decode_to_file={out_file}"
+          "--decode_to_file={out_file} --keep_timestamp"
       ).format(**locals_and_flags)
       command = FLAGS.decoder_command.format(**locals())
       tf.logging.info("Running:\n" + command)