app.py

import gradio as gr
import os
import torch
from transformers import FuyuForCausalLM, AutoTokenizer
from transformers.models.fuyu.processing_fuyu import FuyuProcessor
from transformers.models.fuyu.image_processing_fuyu import FuyuImageProcessor

model_id = "adept/fuyu-8b"
revision = "refs/pr/3"
dtype = torch.bfloat16
device = "cuda"

tokenizer = AutoTokenizer.from_pretrained(model_id, revision=revision)
model = FuyuForCausalLM.from_pretrained(model_id, device_map="auto", torch_dtype=dtype, revision=revision)
processor = FuyuProcessor(image_processor=FuyuImageProcessor(), tokenizer=tokenizer)

caption_prompt = "Generate a coco-style caption.\\n"

def predict(image, prompt):
    # image = image.convert('RGB')
    model_inputs = processor(text=prompt, images=[image])
    model_inputs = {k: v.to(dtype=dtype if torch.is_floating_point(v) else v.dtype, device=device) for k,v in model_inputs.items()}

    generation_output = model.generate(**model_inputs, max_new_tokens=40)
    prompt_len = model_inputs["input_ids"].shape[-1]
    return tokenizer.decode(generation_output[0][prompt_len:], skip_special_tokens=True)

def caption(image):
    return predict(image, caption_prompt)

def set_example_image(example: list) -> dict:
    return gr.Image.update(value=example[0])



css = """
  #mkd {
    height: 500px; 
    overflow: auto; 
    border: 1px solid #ccc; 
  }
"""

with gr.Blocks(css=css) as demo:
    gr.HTML(
        """
            <h1 id="title">Fuyu Multimodal Demo</h1>
            <h3><a href="https://hf.co/adept/fuyu-8b">Fuyu-8B</a> is a multimodal model that supports a variety of tasks combining text and image prompts.</h3>
            For example, you can use it for captioning by asking it to describe an image. You can also ask it questions about an image, a task known as Visual Question Answering, or VQA. This demo lets you explore captioning and VQA, with more tasks coming soon :)
            Learn more about the model in <a href="https://www.adept.ai/blog/fuyu-8b">our blog post</a>.
            <br>
          	<br>
            <strong>Note: This is a raw model release. We have not added further instruction-tuning, postprocessing or sampling strategies to control for undesirable outputs. The model may hallucinate, and you should expect to have to fine-tune the model for your use-case!</strong>
            <h3>Play with Fuyu-8B in this demo! 💬</h3>
        """
    )
    with gr.Tab("Visual Question Answering"):
        with gr.Row():
            with gr.Column():
                image_input = gr.Image(label="Upload your Image")
                text_input = gr.Textbox(label="Ask a Question")
            vqa_output = gr.Textbox(label="Output")
            
        vqa_btn = gr.Button("Answer Visual Question")
        
        gr.Examples(
            [["assets/vqa_example_1.png", "How is this made?"], ["assets/vqa_example_2.png", "What is this flower and where is it's origin?"]],
            inputs = [image_input, text_input],
            outputs = [vqa_output],
            fn=predict,
            cache_examples=True,
            label='Click on any Examples below to get VQA results quickly 👇'
            )

        
    with gr.Tab("Image Captioning"):
        with gr.Row():
            captioning_input = gr.Image(label="Upload your Image")
            captioning_output = gr.Textbox(label="Output")
        captioning_btn = gr.Button("Generate Caption")

        gr.Examples(
            [["assets/captioning_example_1.png"], ["assets/captioning_example_2.png"]],
            inputs = [captioning_input],
            outputs = [captioning_output],
            fn=caption,
            cache_examples=True,
            label='Click on any Examples below to get captioning results quickly 👇'
            )
        
    captioning_btn.click(fn=caption, inputs=captioning_input, outputs=captioning_output)
    vqa_btn.click(fn=predict, inputs=[image_input, text_input], outputs=vqa_output)

    
demo.launch(server_name="0.0.0.0")