Create app.py

app.py

@@ -0,0 +1,24 @@
+# trocr_infer.py  -- paste into your app.py and call ocr_with_trocr(pil_image)
+from transformers import TrOCRProcessor, VisionEncoderDecoderModel
+import torch
+
+device = "cuda" if torch.cuda.is_available() else "cpu"
+
+# Model choices: "microsoft/trocr-base-printed" or "microsoft/trocr-base-handwritten"
+MODEL_NAME = "microsoft/trocr-base-printed"
+
+processor = TrOCRProcessor.from_pretrained(MODEL_NAME)
+model = VisionEncoderDecoderModel.from_pretrained(MODEL_NAME).to(device)
+model.eval()
+
+def ocr_with_trocr(pil_image):
+    """
+    Input: PIL.Image (RGB)
+    Returns: recognized text string
+    """
+    # Preprocess
+    pixel_values = processor(images=pil_image, return_tensors="pt").pixel_values.to(device)
+    # Generate (greedy; tune generation params if desired)
+    generated_ids = model.generate(pixel_values, max_length=128, num_beams=1)
+    text = processor.batch_decode(generated_ids, skip_special_tokens=True)[0]
+    return text.strip()