app.py

import gradio as gr
import requests
from PIL import Image
from transformers import PaliGemmaForConditionalGeneration, PaliGemmaProcessor
import spaces

@spaces.GPU
def infer_diagram(image, question):
    model = PaliGemmaForConditionalGeneration.from_pretrained("google/paligemma-3b-ft-ai2d-448").to("cuda")
    processor = PaliGemmaProcessor.from_pretrained("google/paligemma-3b-ft-ai2d-448")

    inputs = processor(images=image, text=question, return_tensors="pt").to("cuda")

    predictions = model.generate(**inputs, max_new_tokens=100)
    return processor.decode(predictions[0], skip_special_tokens=True)[len(question):].lstrip("\n")

@spaces.GPU
def infer_ocrvqa(image, question):
    model = PaliGemmaForConditionalGeneration.from_pretrained("google/paligemma-3b-ft-ocrvqa-896").to("cuda")
    processor = PaliGemmaProcessor.from_pretrained("google/paligemma-3b-ft-ocrvqa-896")

    inputs = processor(images=image,text=question, return_tensors="pt").to("cuda")

    predictions = model.generate(**inputs, max_new_tokens=100)
    return processor.decode(predictions[0], skip_special_tokens=True)[len(question):].lstrip("\n")

@spaces.GPU
def infer_infographics(image, question):
  model = PaliGemmaForConditionalGeneration.from_pretrained("google/paligemma-3b-ft-infovqa-896").to("cuda")
  processor = PaliGemmaProcessor.from_pretrained("google/paligemma-3b-ft-infovqa-896")

  inputs = processor(images=image, text=question, return_tensors="pt").to("cuda")

  predictions = model.generate(**inputs, max_new_tokens=100)
  return processor.decode(predictions[0], skip_special_tokens=True)[len(question):].lstrip("\n")
@spaces.GPU
def infer_doc(image, question):
  model = PaliGemmaForConditionalGeneration.from_pretrained("google/paligemma-3b-ft-docvqa-896").to("cuda")

  processor = PaliGemmaProcessor.from_pretrained("google/paligemma-3b-ft-docvqa-896")
  inputs = processor(images=image, text=question, return_tensors="pt").to("cuda")
  predictions = model.generate(**inputs, max_new_tokens=100)
  return processor.decode(predictions[0], skip_special_tokens=True)[len(question):].lstrip("\n")

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

with gr.Blocks(css=css) as demo:
  gr.HTML("<h1><center>PaliGemma Fine-tuned on Documents 📄<center><h1>")
  gr.HTML("<h3><center>This Space is built for you to compare different PaliGemma models fine-tuned on document tasks. ⚡</h3>")
  gr.HTML("<h3><center>Each tab in this app demonstrates PaliGemma models fine-tuned on document question answering, infographics question answering, diagram understanding, and reading comprehension from images. 📄📕📊<h3>")
  gr.HTML("<h3><center>Models are downloaded on the go, so first inference in each tab might take time if it's not already downloaded.<h3>")

  with gr.Tab(label="Visual Question Answering over Documents"):
    with gr.Row():
      with gr.Column():
        input_img = gr.Image(label="Input Document")
        question = gr.Text(label="Question")
        submit_btn = gr.Button(value="Submit")
      output = gr.Text(label="Answer")
    gr.Examples(
    [["assets/docvqa_example.png", "How many items are sold?"]],
    inputs = [input_img, question],
    outputs = [output],
    fn=infer_doc,
    label='Click on any Examples below to get Document Question Answering results quickly 👇'
    )

    submit_btn.click(infer_doc, [input_img, question], [output])

  with gr.Tab(label="Visual Question Answering over Infographics"):
    with gr.Row():
      with gr.Column():
        input_img = gr.Image(label="Input Image")
        question = gr.Text(label="Question")
        submit_btn = gr.Button(value="Submit")
      output = gr.Text(label="Answer")
    gr.Examples(
    [["assets/infographics_example (1).jpeg", "What is this infographic about?"]],
    inputs = [input_img, question],
    outputs = [output],
    fn=infer_infographics,
    label='Click on any Examples below to get Infographics QA results quickly 👇'
    )

    submit_btn.click(infer_infographics, [input_img, question], [output])
  with gr.Tab(label="Reading from Images"):
    with gr.Row():
      with gr.Column():
        input_img = gr.Image(label="Input Document")
        question = gr.Text(label="Question")
        submit_btn = gr.Button(value="Submit")
      output = gr.Text(label="Infer")
    submit_btn.click(infer_ocrvqa, [input_img, question], [output])
    gr.Examples(
    [["assets/ocrvqa.jpg", "Who is the author of this book?"]],
    inputs = [input_img, question],
    outputs = [output],
    fn=infer_doc,
    label='Click on any Examples below to get image reading comprehension results quickly 👇'
    )
  with gr.Tab(label="Diagram Understanding"):
    with gr.Row():
      with gr.Column():
        input_img = gr.Image(label="Input Diagram")
        question = gr.Text(label="Question")
        submit_btn = gr.Button(value="Submit")
      output = gr.Text(label="Infer")
    submit_btn.click(infer_diagram, [input_img, question], [output])
    gr.Examples(
    [["assets/diagram.png", "What is the diagram showing?"]],
    inputs = [input_img, question],
    outputs = [output],
    fn=infer_doc,
    label='Click on any Examples below to get diagram understanding results quickly 👇'
    )

demo.launch(debug=True)