[to #42463204] support Pil.Image for image_captioning_pipeline

modelscope/preprocessors/__init__.py

@@ -5,6 +5,6 @@ from .base import Preprocessor
 from .builder import PREPROCESSORS, build_preprocessor
 from .common import Compose
 from .image import LoadImage, load_image
-from .multi_model import OfaImageCaptionPreprocessor
+from .multi_modal import OfaImageCaptionPreprocessor
 from .nlp import *  # noqa F403
 from .text_to_speech import *  # noqa F403

modelscope/preprocessors/multi_model.py → modelscope/preprocessors/multi_modal.py

@@ -73,7 +73,7 @@ class OfaImageCaptionPreprocessor(Preprocessor):
         self.eos_item = torch.LongTensor([task.src_dict.eos()])
         self.pad_idx = task.src_dict.pad()
 
-    @type_assert(object, (str, tuple))
+    @type_assert(object, (str, tuple, Image.Image))
     def __call__(self, data: Union[str, tuple]) -> Dict[str, Any]:
 
         def encode_text(text, length=None, append_bos=False, append_eos=False):
@@ -89,8 +89,8 @@ class OfaImageCaptionPreprocessor(Preprocessor):
                 s = torch.cat([s, self.eos_item])
             return s
 
-        if isinstance(input, Image.Image):
-            patch_image = self.patch_resize_transform(input).unsqueeze(0)
+        if isinstance(data, Image.Image):
+            patch_image = self.patch_resize_transform(data).unsqueeze(0)
         else:
             patch_image = self.patch_resize_transform(
                 load_image(data)).unsqueeze(0)