Create app.py

app.py

@@ -0,0 +1,49 @@
+import gradio as gr
+import spaces
+from cumo.model.builder import load_pretrained_model
+from cumo.mm_utils import process_images, tokenizer_image_token
+from cumo.constants import IMAGE_TOKEN_INDEX, DEFAULT_IMAGE_TOKEN
+import torch
+from PIL import Image
+
+model_path = "BenkHel/CumoThesis"  # <-- Das ist dein Modell!
+device = "cuda" if torch.cuda.is_available() else "cpu"
+
+tokenizer, model, image_processor, context_len = load_pretrained_model(
+    model_path, None, None, False, False, device, use_flash_attn=False
+)
+
+PROMPT = "What material is this item and how is it disposed of?"
+PROMPT_WITH_IMAGE = f"{DEFAULT_IMAGE_TOKEN} {PROMPT}"
+
+@spaces.GPU
+def classify_image(image):
+    if image is None:
+        return "Please upload an image."
+    if not isinstance(image, Image.Image):
+        image = Image.fromarray(image)
+    images = process_images([image], image_processor, model.config)
+    images = [img.to(device, dtype=torch.float16) for img in images]
+    image_args = {"images": images}
+    input_ids = tokenizer_image_token(PROMPT_WITH_IMAGE, tokenizer, IMAGE_TOKEN_INDEX, return_tensors='pt').unsqueeze(0).to(device)
+    with torch.no_grad():
+        outputs = model.generate(
+            inputs=input_ids,
+            max_new_tokens=128,
+            pad_token_id=tokenizer.eos_token_id,
+            **image_args
+        )
+    output_text = tokenizer.decode(outputs[0], skip_special_tokens=True)
+    answer = output_text[len(PROMPT):].strip() if output_text.startswith(PROMPT) else output_text
+    return answer
+
+iface = gr.Interface(
+    fn=classify_image,
+    inputs=gr.Image(type="pil", label="Upload an image of a waste item"),
+    outputs=gr.Textbox(label="Classification & Disposal Recommendation"),
+    title="CuMo Waste Classifier",
+    description="Upload a photo of a household waste item. The model will classify the material and recommend how to dispose of it."
+)
+
+if __name__ == "__main__":
+    iface.launch()