VLM.py 952 B

1234567891011121314151617181920212223242526272829
  1. from transformers import BlipProcessor, BlipForQuestionAnswering
  2. import torch
  3. from PIL import Image
  4. device = "cuda" if torch.cuda.is_available() else "cpu"
  5. # Chargement du modèle BLIP pré-entraîné (VQA)
  6. processor = BlipProcessor.from_pretrained("Salesforce/blip-vqa-base", use_fast=True)
  7. model = BlipForQuestionAnswering.from_pretrained("Salesforce/blip-vqa-base").to(device)
  8. model.eval()
  9. def ask_blip(image_path, question):
  10. image = Image.open(image_path).convert("RGB")
  11. # Préparer les inputs (image + question)
  12. inputs = processor(image, question, return_tensors="pt").to(device)
  13. # Générer la réponse
  14. with torch.no_grad():
  15. outputs = model.generate(**inputs)
  16. answer = processor.tokenizer.decode(outputs[0], skip_special_tokens=True)
  17. return answer
  18. # Exemple d'utilisation
  19. if __name__ == "__main__":
  20. img_path = "co.jpg"
  21. question = "explain the image?"
  22. print(ask_blip(img_path, question))