fix lazy load (#328)

(cherry picked from commit 326d9a8b3fab1b4da35b6c04bd7c27dbbe9cf14c)

swift/llm/utils/dataset.py

@@ -1121,20 +1121,7 @@ def load_dataset_from_local(
         dataset = HfDataset.from_dict(df.to_dict(orient='list'))
         dataset_list.append(preprocess_func(dataset))
 
-    dataset = concatenate_datasets(dataset_list)
-
-    def load_image(row):
-        from PIL import Image
-        import requests
-        if not os.path.exists(row['image']):
-            row['image'] = requests.get(row['image'], stream=True).raw
-        row['image'] = Image.open(row['image'])
-        return row
-
-    if 'image' in dataset.features and isinstance(dataset[0]['image'], str):
-        dataset = HfDataset.from_list(
-            dataset_map(dataset, load_image, num_proc=4).data)
-    return dataset
+    return concatenate_datasets(dataset_list)
 
 
 def get_custom_dataset(_: str, train_subset_split_list: Union[str, List[str]],

swift/llm/utils/template.py

@@ -1,4 +1,5 @@
 # Copyright (c) Alibaba, Inc. and its affiliates.
+import os
 from copy import deepcopy
 from typing import Any, Dict, List, Literal, Optional, Tuple, Union
 
@@ -554,6 +555,13 @@ def build_conversation_input_ids(
 
     def encode(self, example: Dict[str,
                                    Any]) -> Dict[str, Optional[List[int]]]:
+        if 'image' in example and isinstance(example['image'], str):
+            from PIL import Image
+            import requests
+            if not os.path.exists(example['image']):
+                example['image'] = requests.get(
+                    example['image'], stream=True).raw
+            example['image'] = Image.open(example['image'])
         return self.build_conversation_input_ids(
             self.tokenizer,
             query=example['query'],