Spaces:
Sleeping
Sleeping
| import spaces | |
| import torch | |
| import re | |
| import gradio as gr | |
| from transformers import AutoTokenizer, AutoModelForCausalLM | |
| from PIL import Image | |
| if torch.cuda.is_available(): | |
| device, dtype = "cuda", torch.float16 | |
| else: | |
| device, dtype = "cpu", torch.float32 | |
| model_id = "vikhyatk/moondream2" | |
| revision = "2024-04-02" | |
| tokenizer = AutoTokenizer.from_pretrained(model_id, revision=revision) | |
| moondream = AutoModelForCausalLM.from_pretrained( | |
| model_id, trust_remote_code=True, revision=revision, torch_dtype=dtype | |
| ).to(device=device) | |
| moondream.eval() | |
| def answer_questions(image_tuples, prompt_text): | |
| result = "" | |
| prompts = [p.strip() for p in prompt_text.split(',')] # Splitting and cleaning prompts | |
| print(f"prompts\n{prompts}\n") | |
| image_embeds = [img[0] for img in image_tuples if img[0] is not None] # Extracting images from tuples, ignoring None | |
| # Check if the lengths of image_embeds and prompts are equal | |
| #if len(image_embeds) != len(prompts): | |
| #return ("Error: The number of images input and prompts input (seperate by commas in input text field) must be the same.") | |
| answers = [] | |
| for prompt in prompts: | |
| image_answers = moondream.batch_answer( | |
| images=[img.convert("RGB") for img in image_embeds], | |
| prompts=[prompt] * len(image_embeds), | |
| tokenizer=tokenizer, | |
| ) | |
| answers.append(image_answers) | |
| data = [] | |
| for i in range(len(image_tuples)): | |
| image_name = f"image{i+1}" | |
| image_answers = [answer[i] for answer in answers] | |
| print(f"image{i+1}_answers \n {image_answers} \n") | |
| data.append([image_name] + image_answers) | |
| for question, answer in zip(prompts, answers): | |
| Q_and_A += (f"Q: {question}\nA: {answer}\n\n") | |
| print(f"\n\n{Q_and_A}\n\n") | |
| result = {'headers': prompts, 'data': data} | |
| return result | |
| ''' | |
| answers = moondream.batch_answer( | |
| images=image_embeds, | |
| prompts=prompts, | |
| tokenizer=tokenizer, | |
| ) | |
| for question, answer in zip(prompts, answers): | |
| result += (f"Q: {question}\nA: {answer}\n\n") | |
| return result | |
| ''' | |
| with gr.Blocks() as demo: | |
| gr.Markdown("# moondream2 unofficial batch processing demo") | |
| gr.Markdown("1. Select images\n2. Enter one or more prompts separated by commas. Ex: Describe this image, What is in this image?\n\n") | |
| gr.Markdown("**Currently each image will be sent as a batch with the prompts thus asking each promp on each image**") | |
| gr.Markdown("*Running on free CPU space tier currently so results may take a bit to process compared to duplicating space and using GPU space hardware*") | |
| gr.Markdown("## 🌔 moondream2\nA tiny vision language model. [GitHub](https://github.com/vikhyatk/moondream)") | |
| with gr.Row(): | |
| img = gr.Gallery(label="Upload Images", type="pil") | |
| with gr.Row(): | |
| prompt = gr.Textbox(label="Input Prompts", placeholder="Enter prompts (one prompt for each image provided) separated by commas. Ex: Describe this image, What is in this image?", lines=8) | |
| with gr.Row(): | |
| submit = gr.Button("Submit") | |
| output = gr.TextArea(label="Questions and Answers", lines=30) | |
| output2 = gr.Dataframe(label="Structured Dataframe", type="array",wrap=True) | |
| submit.click(answer_questions, [img, prompt], output, output2) | |
| demo.queue().launch() | |