from PIL import Image import requests from transformers import Blip2Processor, Blip2ForConditionalGeneration import torch device = "cuda" if torch.cuda.is_available() else "cpu" print(device) def init_BLIP(device): processor = Blip2Processor.from_pretrained("Salesforce/blip2-opt-2.7b") if device == 'cuda': model = Blip2ForConditionalGeneration.from_pretrained( "Salesforce/blip2-opt-2.7b", load_in_8bit=True,torch_dtype=torch.float16, device_map = 'auto') else: print('Using CPU model') model = Blip2ForConditionalGeneration.from_pretrained( "Salesforce/blip2-opt-2.7b",device_map={"": device}, torch_dtype=torch.float32,low_cpu_mem_usage=True) model.eval() if torch.__version__ >= "2": model = torch.compile(model) processor = processor return model,processor def infer_BLIP2(model,processor,image,device): outputs= '' prompts = [ "This is a picture of", "Question: What is in the picture? Answer:", "Question: Where is this image depicting? Answer:", "Question: Who is in this picture? Answer:", "Question: What are the things in the picture doing? Answer:", "Question: Why do you think they are doing it? Answer:", "Question: What emotion does the person or animal in the image feel? Answer:", ] for prompt in prompts: if device == 'cuda': inputs = processor(images=image, text=prompt, return_tensors="pt").to(device, torch.float16) else: inputs = processor(images=image, text=prompt, return_tensors="pt") generated_ids = model.generate(**inputs) generated_text = processor.batch_decode(generated_ids, skip_special_tokens=True)[0].strip() outputs+= prompt+generated_text+' ' return outputs ''' Testing model,processor = init_BLIP(device) image = Image.open('/home/spooky/Downloads/IMG20221214012021.jpg') infer_BLIP2(model,processor,image,device)'''