fix resnet50 usetime statistics (#4838)

PaddleCV/image_classification/train.py

@@ -234,6 +234,7 @@ def train(args):
         train_batch_id = 0
         train_batch_time_record = []
         train_batch_metrics_record = []
+        train_batch_time_print_step = []
 
         if not args.use_dali:
             train_iter = train_data_loader()
@@ -256,8 +257,19 @@ def train(args):
                 np.array(train_batch_metrics), axis=1)
             train_batch_metrics_record.append(train_batch_metrics_avg)
             if trainer_id == 0:
-                print_info("batch", train_batch_metrics_avg, train_batch_elapse,
-                           pass_id, train_batch_id, args.print_step)
+                if train_batch_id % args.print_step == 0:
+                    if len(train_batch_time_print_step) == 0:
+                        train_batch_time_print_step_avg = train_batch_elapse
+                    else:
+                        train_batch_time_print_step_avg = np.mean(
+                            train_batch_time_print_step)
+                    train_batch_time_print_step = []
+                    print_info("batch", train_batch_metrics_avg,
+                               train_batch_time_print_step_avg, pass_id,
+                               train_batch_id, args.print_step)
+                else:
+                    train_batch_time_print_step.append(train_batch_elapse)
+
                 sys.stdout.flush()
             train_batch_id += 1
             t1 = time.time()