train.py: val logger on gpu 0 only

train.py

@@ -142,8 +142,9 @@ def validate(model, criterion, valset, iteration, batch_size, n_gpus,
         val_loss = val_loss / (i + 1)
 
     model.train()
-    print("Validation loss {}: {:9f}  ".format(iteration, reduced_val_loss))
-    logger.log_validation(reduced_val_loss, model, y, y_pred, iteration)
+    if rank == 0:
+        print("Validation loss {}: {:9f}  ".format(iteration, reduced_val_loss))
+        logger.log_validation(reduced_val_loss, model, y, y_pred, iteration)
 
 
 def train(output_directory, log_directory, checkpoint_path, warm_start, n_gpus,
@@ -236,9 +237,9 @@ def train(output_directory, log_directory, checkpoint_path, warm_start, n_gpus,
                     reduced_loss, grad_norm, learning_rate, duration, iteration)
 
             if not overflow and (iteration % hparams.iters_per_checkpoint == 0):
-                validate(model, criterion, valset, iteration, hparams.batch_size,
-                         n_gpus, collate_fn, logger, hparams.distributed_run, rank)
-
+                validate(model, criterion, valset, iteration,
+                         hparams.batch_size, n_gpus, collate_fn, logger,
+                         hparams.distributed_run, rank)
                 if rank == 0:
                     checkpoint_path = os.path.join(
                         output_directory, "checkpoint_{}".format(iteration))