import gradio as gr from transformers import AutoProcessor, AutoTokenizer, AutoImageProcessor, AutoModelForCausalLM, BlipForConditionalGeneration, Blip2ForConditionalGeneration, VisionEncoderDecoderModel, InstructBlipForConditionalGeneration import torch import open_clip from huggingface_hub import hf_hub_download device = "cuda" if torch.cuda.is_available() else "cpu" torch.hub.download_url_to_file('http://images.cocodataset.org/val2017/000000039769.jpg', 'cats.jpg') torch.hub.download_url_to_file('https://huggingface.co/datasets/nielsr/textcaps-sample/resolve/main/stop_sign.png', 'stop_sign.png') torch.hub.download_url_to_file('https://cdn.openai.com/dall-e-2/demos/text2im/astronaut/horse/photo/0.jpg', 'astronaut.jpg') git_processor_large_coco = AutoProcessor.from_pretrained("microsoft/git-large-coco") git_model_large_coco = AutoModelForCausalLM.from_pretrained("microsoft/git-large-coco").to(device) blip_processor_large = AutoProcessor.from_pretrained("Salesforce/blip-image-captioning-large") blip_model_large = BlipForConditionalGeneration.from_pretrained("Salesforce/blip-image-captioning-large").to(device) blip2_processor = AutoProcessor.from_pretrained("Salesforce/blip2-opt-6.7b-coco") blip2_model = Blip2ForConditionalGeneration.from_pretrained("Salesforce/blip2-opt-6.7b-coco", device_map="auto", load_in_4bit=True, torch_dtype=torch.float16) instructblip_processor = AutoProcessor.from_pretrained("Salesforce/instructblip-vicuna-7b") instructblip_model = InstructBlipForConditionalGeneration.from_pretrained("Salesforce/instructblip-vicuna-7b", device_map="auto", load_in_4bit=True, torch_dtype=torch.float16) def generate_caption(processor, model, image, tokenizer=None, use_float_16=False): inputs = processor(images=image, return_tensors="pt").to(device) if use_float_16: inputs = inputs.to(torch.float16) generated_ids = model.generate(pixel_values=inputs.pixel_values, num_beams=3, max_length=20, min_length=5) if tokenizer is not None: generated_caption = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0] else: generated_caption = processor.batch_decode(generated_ids, skip_special_tokens=True)[0] return generated_caption def generate_caption_blip2(processor, model, image, replace_token=False): prompt = "A photo of" inputs = processor(images=image, text=prompt, return_tensors="pt").to(device=model.device, dtype=torch.float16) generated_ids = model.generate(**inputs, num_beams=5, max_length=50, min_length=1, top_p=0.9, repetition_penalty=1.5, length_penalty=1.0, temperature=1) if replace_token: # TODO remove once https://github.com/huggingface/transformers/pull/24492 is merged generated_ids[generated_ids == 0] = 2 return processor.batch_decode(generated_ids, skip_special_tokens=True)[0] def generate_captions(image): caption_git_large_coco = generate_caption(git_processor_large_coco, git_model_large_coco, image) caption_blip_large = generate_caption(blip_processor_large, blip_model_large, image) caption_blip2 = generate_caption_blip2(blip2_processor, blip2_model, image).strip() caption_instructblip = generate_caption_blip2(instructblip_processor, instructblip_model, image, replace_token=True) return caption_git_large_coco, caption_blip_large, caption_blip2, caption_instructblip examples = [["cats.jpg"], ["stop_sign.png"], ["astronaut.jpg"]] outputs = [gr.outputs.Textbox(label="Caption generated by GIT-large fine-tuned on COCO"), gr.outputs.Textbox(label="Caption generated by BLIP-large"), gr.outputs.Textbox(label="Caption generated by BLIP-2 OPT 6.7b"), gr.outputs.Textbox(label="Caption generated by InstructBLIP"), ] title = "Interactive demo: comparing image captioning models" description = "Gradio Demo to compare GIT, BLIP, BLIP-2 and InstructBLIP, 4 state-of-the-art vision+language models. To use it, simply upload your image and click 'submit', or click one of the examples to load them. Read more at the links below." article = "
" interface = gr.Interface(fn=generate_captions, inputs=gr.inputs.Image(type="pil"), outputs=outputs, examples=examples, title=title, description=description, article=article, enable_queue=True) interface.launch(debug=True)