Update app.py

app.py

@@ -5,7 +5,9 @@ import torch
 
 print("Loading TrOCR model...")
 device = "cuda" if torch.cuda.is_available() else "cpu"
-processor = TrOCRProcessor.from_pretrained('microsoft/trocr-base-handwritten')
+
+# use_fast=False completely bypasses the transformers tokenizer conversion bug!
+processor = TrOCRProcessor.from_pretrained('microsoft/trocr-base-handwritten', use_fast=False)
 model = VisionEncoderDecoderModel.from_pretrained('microsoft/trocr-base-handwritten').to(device)
 print("Model loaded successfully!")
 
@@ -14,16 +16,13 @@ def predict(image):
     if image is None:
         return ""
     
-    # Process the image and send it to exactly where the model is currently living
     pixel_values = processor(image, return_tensors="pt").pixel_values.to(model.device)
     
-    # Generate the text extraction
     generated_ids = model.generate(pixel_values)
     generated_text = processor.batch_decode(generated_ids, skip_special_tokens=True)[0]
     
     return generated_text
 
-# The final interface
 demo = gr.Interface(
     fn=predict, 
     inputs=gr.Image(type="pil"),