refine resnet benchmard print (#4893)

* fix ptb_dy time print for benchmark, test=develop

* refine resnet benchmard print, test=develop

PaddleCV/image_classification/train.py

@@ -120,6 +120,7 @@ def validate(args,
     test_batch_time_record = []
     test_batch_metrics_record = []
     test_batch_id = 0
+
     if int(os.environ.get('PADDLE_TRAINERS_NUM', 1)) > 1:
         compiled_program = test_prog
     else:
@@ -245,6 +246,7 @@ def train(args):
                                                  train_fetch_vars[0], exe)
 
     batch_cost_avg = TimeCostAverage()
+    reader_cost_avg = TimeCostAverage()
 
     #NOTE: this for benchmark
 
@@ -267,11 +269,14 @@ def train(args):
             #NOTE: this is for benchmark
             if args.max_iter and total_batch_num == args.max_iter:
                 return
+            t2 = time.time()
+            reader_cost = t2 - t1
+            reader_cost_avg.record(reader_cost)
             train_batch_metrics = exe.run(compiled_train_prog,
                                           feed=batch,
                                           fetch_list=train_fetch_list)
-            t2 = time.time()
-            train_batch_elapse = t2 - t1
+            t3 = time.time()
+            train_batch_elapse = t3 - t1
             train_batch_time_record.append(train_batch_elapse)
             batch_cost_avg.record(train_batch_elapse)
 
@@ -279,11 +284,18 @@ 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,
-                           batch_cost_avg.get_average(), pass_id,
-                           train_batch_id, args.print_step)
+                print_info(
+                    "batch",
+                    train_batch_metrics_avg,
+                    batch_cost_avg.get_average(),
+                    pass_id,
+                    train_batch_id,
+                    args.print_step,
+                    reader_cost=reader_cost_avg.get_average(),
+                    ips=args.batch_size / batch_cost_avg.get_average())
                 sys.stdout.flush()
                 if train_batch_id % args.print_step == 0:
+                    reader_cost_avg.reset()
                     batch_cost_avg.reset()
             train_batch_id += 1
             t1 = time.time()

PaddleCV/image_classification/utils/utility.py

@@ -420,7 +420,9 @@ def print_info(info_mode,
                batch_id=0,
                print_step=1,
                device_num=1,
-               class_dim=5):
+               class_dim=5,
+               reader_cost=0.0,
+               ips=0.0):
     """print function
 
     Args:
@@ -439,25 +441,28 @@ def print_info(info_mode,
             if len(metrics) == 2:
                 loss, lr = metrics
                 logger.info(
-                    "[Pass {0}, train batch {1}] \tloss {2}, lr {3}, elapse {4}".
+                    "[Pass {0}, train batch {1}] \tloss {2}, lr {3}, reader_cost: {5}, batch_cost: {4}, ips: {6}".
                     format(pass_id, batch_id, "%.5f" % loss, "%.5f" % lr,
-                           "%2.4f sec" % time_info))
+                           "%2.4f sec" % time_info, "%.5f sec" % reader_cost,
+                           "%.5f images/sec" % ips))
             # train and no mixup output
             elif len(metrics) == 4:
                 loss, acc1, acc5, lr = metrics
                 logger.info(
-                    "[Pass {0}, train batch {1}] \tloss {2}, acc1 {3}, acc{7} {4}, lr {5}, elapse {6}".
+                    "[Pass {0}, train batch {1}] \tloss {2}, acc1 {3}, acc{7} {4}, lr {5}, reader_cost: {8}, batch_cost: {6}, ips: {9}".
                     format(pass_id, batch_id, "%.5f" % loss, "%.5f" % acc1,
                            "%.5f" % acc5, "%.5f" % lr, "%2.4f sec" % time_info,
-                           min(class_dim, 5)))
+                           min(class_dim, 5), "%.5f sec" % reader_cost,
+                           "%.5f images/sec" % ips))
             # test output
             elif len(metrics) == 3:
                 loss, acc1, acc5 = metrics
                 logger.info(
-                    "[Pass {0}, test  batch {1}] \tloss {2}, acc1 {3}, acc{6} {4}, elapse {5}".
+                    "[Pass {0}, test  batch {1}] \tloss {2}, acc1 {3}, acc{6} {4}, reader_cost: {7}, batch_cost: {5}, ips: {8}".
                     format(pass_id, batch_id, "%.5f" % loss, "%.5f" % acc1,
                            "%.5f" % acc5, "%2.4f sec" % time_info,
-                           min(class_dim, 5)))
+                           min(class_dim, 5), "%.5f sec" % reader_cost,
+                           "%.5f images/sec" % ips))
             else:
                 raise Exception(
                     "length of metrics {} is not implemented, It maybe caused by wrong format of build_program_output".
@@ -525,8 +530,9 @@ def best_strategy_compiled(args,
             fluid.require_version(min_version='1.7.0')
             build_strategy.fuse_bn_act_ops = args.fuse_bn_act_ops
         except Exception as e:
-            logger.info("PaddlePaddle version 1.7.0 or higher is "
-            "required when you want to fuse batch_norm and activation_op.")
+            logger.info(
+                "PaddlePaddle version 1.7.0 or higher is "
+                "required when you want to fuse batch_norm and activation_op.")
         build_strategy.fuse_elewise_add_act_ops = args.fuse_elewise_add_act_ops
 
         exec_strategy = fluid.ExecutionStrategy()