[tune] Trial reporter fix (#3951)

Fixes #3949.

python/ray/tune/trainable.py

@@ -18,9 +18,9 @@ import uuid
 
 import ray
 from ray.tune.logger import UnifiedLogger
-from ray.tune.result import (DEFAULT_RESULTS_DIR, TIME_THIS_ITER_S,
-                             TIMESTEPS_THIS_ITER, DONE, TIMESTEPS_TOTAL,
-                             EPISODES_THIS_ITER, EPISODES_TOTAL)
+from ray.tune.result import (
+    DEFAULT_RESULTS_DIR, TIME_THIS_ITER_S, TIMESTEPS_THIS_ITER, DONE,
+    TIMESTEPS_TOTAL, EPISODES_THIS_ITER, EPISODES_TOTAL, TRAINING_ITERATION)
 from ray.tune.trial import Resources
 
 logger = logging.getLogger(__name__)
@@ -181,6 +181,7 @@ class Trainable(object):
         # self._timesteps_total should not override user-provided total
         result.setdefault(TIMESTEPS_TOTAL, self._timesteps_total)
         result.setdefault(EPISODES_TOTAL, self._episodes_total)
+        result.setdefault(TRAINING_ITERATION, self._iteration)
 
         # Provides auto-filled neg_mean_loss for avoiding regressions
         if result.get("mean_loss"):
@@ -191,7 +192,6 @@ class Trainable(object):
             experiment_id=self._experiment_id,
             date=now.strftime("%Y-%m-%d_%H-%M-%S"),
             timestamp=int(time.mktime(now.timetuple())),
-            training_iteration=self._iteration,
             time_this_iter_s=time_this_iter,
             time_total_s=self._time_total,
             pid=os.getpid(),