Update app.py

app.py

@@ -17,28 +17,6 @@ device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
 model = PaliGemmaForConditionalGeneration.from_pretrained(model_id, torch_dtype=torch.bfloat16).eval().to(device)
 processor = PaliGemmaProcessor.from_pretrained(model_id)
 
-def resize_and_pad(image, target_dim):
-    # Calculate the aspect ratio
-    scale_factor = 1
-    aspect_ratio = image.width / image.height
-    if aspect_ratio > 1:
-        # Width is greater than height
-        new_width = int(target_dim * scale_factor)
-        new_height = int((target_dim / aspect_ratio) * scale_factor)
-    else:
-        # Height is greater than width
-        new_height = int(target_dim * scale_factor)
-        new_width = int(target_dim * aspect_ratio * scale_factor)
-
-    resized_image = image.resize((new_width, new_height), Image.LANCZOS)
-
-    # Create a new image with the target dimensions and a white background
-    new_image = Image.new("RGB", (target_dim, target_dim), (255, 255, 255))
-    new_image.paste(resized_image, ((target_dim - new_width) // 2, (target_dim - new_height) // 2))
-
-    return new_image
-
-
 ###### Transformers Inference
 @spaces.GPU
 def infer(
@@ -46,7 +24,7 @@ def infer(
         text: str,
         max_new_tokens: int
 ) -> str:
-    inputs = processor(text=text, images=resize_and_pad(image, 448), return_tensors="pt", padding="longest", do_convert_rgb=True).to(device).to(dtype=model.dtype)
+    inputs = processor(text=text, images=image, return_tensors="pt", padding="longest", do_convert_rgb=True).to(device).to(dtype=model.dtype)
     with torch.no_grad():
         generated_ids = model.generate(
             **inputs,
@@ -57,16 +35,8 @@ def infer(
 
 ######## Demo
 
-INTRO_TEXT = """## Curacel Handwritten Arabic demo\n\n
+INTRO_TEXT = """## Curacel Auto Damage demo\n\n
 Finetuned from: google/paligemma-3b-pt-448
-
-
-Translation model demo at: https://prod.arabic-gpt.ai/
-
-Prompts:
-Translate the Arabic to English: {model output}
-
-The following is a diagnosis in Arabic from a medical billing form we need to translate to English. The transcriber is not necessariily accurate so one or more characters or words may be wrong. Given what is written, what is the most likely diagnosis. Think step by step, and think about similar words or mispellings in Arabic. Give multiple arabic diagnoses along with the translation in English for each, then finally select the diagnosis that makes the most sense given what was transcribed and print the English translation as your most likely final translation. Transcribed text:  {model output}
 """
 
 with gr.Blocks(css="style.css") as demo: