allow specifying --checkpoint_path with t2t-decoder

and allow keeping timestamp in that case.
This is needed for t2t-translate-all + t2t-bleu to work
as expected (I forgot to add this commit to tensorflow#488).

Author: martinpopel

tensor2tensor/bin/t2t_decoder.py

@@ -47,10 +47,14 @@
 FLAGS = flags.FLAGS
 
 # Additional flags in bin/t2t_trainer.py and utils/flags.py
+flags.DEFINE_string("checkpoint_path", None,
+                    "Path to the model checkpoint. Overrides output_dir.")
 flags.DEFINE_string("decode_from_file", None,
                     "Path to the source file for decoding")
 flags.DEFINE_string("decode_to_file", None,
                     "Path to the decoded (output) file")
+flags.DEFINE_bool("keep_timestamp", True,
+                  "Set the mtime of the decoded file to the checkpoint_path+'.index' mtime.")
 flags.DEFINE_bool("decode_interactive", False,
                   "Interactive local inference mode.")
 flags.DEFINE_integer("decode_shards", 1, "Number of decoding replicas.")
@@ -76,7 +80,11 @@ def decode(estimator, hparams, decode_hp):
     decoding.decode_interactively(estimator, hparams, decode_hp)
   elif FLAGS.decode_from_file:
     decoding.decode_from_file(estimator, FLAGS.decode_from_file, hparams,
-                              decode_hp, FLAGS.decode_to_file)
+                              decode_hp, FLAGS.decode_to_file,
+                              checkpoint_path=FLAGS.checkpoint_path)
+    if FLAGS.checkpoint_path and FLAGS.keep_timestamp:
+      ckpt_time = os.path.getmtime(FLAGS.checkpoint_path + '.index')
+      os.utime(FLAGS.decode_to_file, (ckpt_time, ckpt_time))
   else:
     decoding.decode_from_dataset(
         estimator,

tensor2tensor/utils/decoding.py

@@ -219,7 +219,8 @@ def decode_from_file(estimator,
                      filename,
                      hparams,
                      decode_hp,
-                     decode_to_file=None):
+                     decode_to_file=None,
+                     checkpoint_path=None):
   """Compute predictions on entries in filename and write them out."""
   if not decode_hp.batch_size:
     decode_hp.batch_size = 32
@@ -248,7 +249,7 @@ def input_fn():
     return _decode_input_tensor_to_features_dict(example, hparams)
 
   decodes = []
-  result_iter = estimator.predict(input_fn)
+  result_iter = estimator.predict(input_fn, checkpoint_path=checkpoint_path)
   for result in result_iter:
     if decode_hp.return_beams:
       beam_decodes = []