Fix/ncf eval default (tensorflow#5438)

Taylor Robie · web-flow · commit aec1fec62873 · 2018-10-04T17:00:36.000-07:00
* improve default handling for eval_batch_size

* return eval_batch_size default to None

* fix syntax error
diff --git a/official/recommendation/ncf_main.py b/official/recommendation/ncf_main.py
@@ -128,8 +128,9 @@ def run_ncf(_):
   batch_size = distribution_utils.per_device_batch_size(
       int(FLAGS.batch_size), num_gpus)
 
-  eval_batch_size = int(FLAGS.eval_batch_size or FLAGS.batch_size)
   eval_per_user = rconst.NUM_EVAL_NEGATIVES + 1
+  eval_batch_size = int(FLAGS.eval_batch_size or
+                        max([FLAGS.batch_size, eval_per_user]))
   if eval_batch_size % eval_per_user:
     eval_batch_size = eval_batch_size // eval_per_user * eval_per_user
     tf.logging.warning(
@@ -365,7 +366,8 @@ def define_ncf_flags():
   @flags.validator("eval_batch_size", "eval_batch_size must be at least {}"
                    .format(rconst.NUM_EVAL_NEGATIVES + 1))
   def eval_size_check(eval_batch_size):
-    return int(eval_batch_size) > rconst.NUM_EVAL_NEGATIVES
+    return (eval_batch_size is None or
+            int(eval_batch_size) > rconst.NUM_EVAL_NEGATIVES)
 
 
 if __name__ == "__main__":
