import gradio as gr from transformers import AutoProcessor, AutoModelForCausalLM, BlipForQuestionAnswering, ViltForQuestionAnswering import torch 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_base = AutoProcessor.from_pretrained("microsoft/git-base-vqav2") git_model_base = AutoModelForCausalLM.from_pretrained("microsoft/git-base-vqav2") git_processor_large = AutoProcessor.from_pretrained("microsoft/git-large-vqav2") git_model_large = AutoModelForCausalLM.from_pretrained("microsoft/git-large-vqav2") blip_processor_base = AutoProcessor.from_pretrained("Salesforce/blip-vqa-base") blip_model_base = BlipForConditionalGeneration.from_pretrained("Salesforce/blip-vqa-base") blip_processor_large = AutoProcessor.from_pretrained("Salesforce/blip-vqa-large") blip_model_large = BlipForConditionalGeneration.from_pretrained("Salesforce/blip-vqa-large") vilt_processor = AutoProcessor.from_pretrained("dandelin/vilt-b32-finetuned-vqa") vilt_model = ViltForQuestionAnswering.from_pretrained("dandelin/vilt-b32-finetuned-vqa") device = "cuda" if torch.cuda.is_available() else "cpu" git_model_base.to(device) blip_model_base.to(device) git_model_large.to(device) blip_model_large.to(device) vilt_model.to(device) def generate_answer_git(processor, model, image, question): # prepare image pixel_values = processor(images=image, return_tensors="pt").pixel_values # prepare question input_ids = processor(text=question, add_special_tokens=False).input_ids input_ids = [processor.tokenizer.cls_token_id] + input_ids input_ids = torch.tensor(input_ids).unsqueeze(0) generated_ids = model.generate(pixel_values=pixel_values, input_ids=input_ids, max_length=50) generated_answer = processor.batch_decode(generated_ids, skip_special_tokens=True) return generated_answer def generate_answer_blip(processor, model, image, question): # prepare image + question inputs = processor(images=image, text=text, return_tensors="pt") generated_ids = model.generate(**inputs, max_length=50) generated_answer = processor.batch_decode(generated_ids, skip_special_tokens=True) return generated_answer def generate_answer_vilt(processor, model, image, question): # prepare image + question encoding = processor(image, text, return_tensors="pt") with torch.no_grad(): outputs = model(**encoding) predicted_class_idx = outputs.logits.argmax(-1).item() return model.config.id2label[predicted_class_idx] def generate_answers(image, question): answer_git_base = generate_caption(git_processor_base, git_model_base, image, question) answer_git_large = generate_caption(git_processor_large, git_model_large, image, question) answer_blip_base = generate_caption(blip_processor_base, blip_model_base, image, question) answer_blip_large = generate_caption(blip_processor_large, blip_model_large, image, question) answer_vilt = generate_answer_vilt(vilt_processor, vilt_model, image, question) return answer_git_base, answer_git_large, answer_blip_base, answer_blip_large, answer_vilt examples = [["cats.jpg", "How many cats are there?"], ["stop_sign.png", "What's behind the stop sign?"], ["astronaut.jpg", "What's the astronaut riding on?"]] outputs = [gr.outputs.Textbox(label="Answer generated by GIT-base"), gr.outputs.Textbox(label="Answer generated by GIT-large"), gr.outputs.Textbox(label="Answer generated by BLIP-base"), gr.outputs.Textbox(label="Answer generated by BLIP-large"), gr.outputs.Textbox(label="Answer generated by ViLT")] title = "Interactive demo: comparing visual question answering (VQA) models" description = "Gradio Demo to compare GIT, BLIP and ViLT, 3 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 = "

BLIP docs | GIT docs

" interface = gr.Interface(fn=generate_answers, inputs=[gr.inputs.Image(type="pil"), gr.inputs.Textbox(label="Question")], outputs=outputs, examples=examples, title=title, description=description, article=article, enable_queue=True) interface.launch(debug=True)