debug

app.py

@@ -136,8 +136,8 @@ def patch_resize_transform(patch_image_size=480, is_document=False):
 
 
 # Construct input for caption task
-def construct_sample(image: Image, patch_image_size=480):
-    patch_image = patch_resize_transform(patch_image_size)(image).unsqueeze(0)
+def construct_sample(image: Image, patch_image_size=480, is_document=False):
+    patch_image = patch_resize_transform(patch_image_size, is_document=is_document)(image).unsqueeze(0)
     patch_mask = torch.tensor([True])
     src_text = encode_text("图片上的文字是什么?", append_bos=True, append_eos=True).unsqueeze(0)
     src_length = torch.LongTensor([s.ne(pad_idx).long().sum() for s in src_text])
@@ -177,7 +177,7 @@ def ocr(img, task_type):
     ocr_result = []
     for i, (box, image) in enumerate(zip(box_list, image_list)):
         image = Image.fromarray(image)
-        sample = construct_sample(task, image, cfg.task.patch_image_size)
+        sample = construct_sample(image, cfg.task.patch_image_size, is_document=(task_type=='Document'))
         sample = utils.move_to_cuda(sample) if use_cuda else sample
         sample = utils.apply_to_sample(apply_half, sample) if use_fp16 else sample