Fix the exception of DALI

tools/static/program.py

@@ -96,7 +96,7 @@ def create_model(architecture, image, classes_num, config, is_train):
     if "is_test" in params:
         params['is_test'] = not is_train
     model = architectures.__dict__[name](class_dim=classes_num, **params)
-    
+
     out = model(image)
     return out
 
@@ -289,8 +289,9 @@ def create_strategy(config):
     exec_strategy = paddle.static.ExecutionStrategy()
 
     exec_strategy.num_threads = 1
-    exec_strategy.num_iteration_per_drop_scope = (10000 if 'AMP' in config and
-                    config.AMP.get("use_pure_fp16", False) else 10)
+    exec_strategy.num_iteration_per_drop_scope = (
+        10000
+        if 'AMP' in config and config.AMP.get("use_pure_fp16", False) else 10)
 
     fuse_op = True if 'AMP' in config else False
 
@@ -357,7 +358,8 @@ def mixed_precision_optimizer(config, optimizer):
     if 'AMP' in config:
         amp_cfg = config.AMP if config.AMP else dict()
         scale_loss = amp_cfg.get('scale_loss', 1.0)
-        use_dynamic_loss_scaling = amp_cfg.get('use_dynamic_loss_scaling', False)
+        use_dynamic_loss_scaling = amp_cfg.get('use_dynamic_loss_scaling',
+                                               False)
         use_pure_fp16 = amp_cfg.get('use_pure_fp16', False)
         optimizer = paddle.static.amp.decorate(
             optimizer,
@@ -501,7 +503,21 @@ def run(dataloader,
     use_dali = config.get('use_dali', False)
     dataloader = dataloader if use_dali else dataloader()
     tic = time.time()
-    for idx, batch in enumerate(dataloader):
+
+    idx = 0
+    batch_size = None
+    while True:
+        # The DALI maybe raise RuntimeError for some particular images, such as ImageNet1k/n04418357_26036.JPEG
+        try:
+            batch = next(dataloader)
+        except StopIteration:
+            break
+        except RuntimeError:
+            logger.warning(
+                "Except RuntimeError when reading data from dataloader, try to read once again..."
+            )
+            continue
+        idx += 1
         # ignore the warmup iters
         if idx == 5:
             metric_list["batch_time"].reset()