app.py

import gradio as gr
from PIL import Image
from transformers import BitsAndBytesConfig, PaliGemmaForConditionalGeneration, PaliGemmaProcessor
import spaces
import torch
import os


access_token = os.getenv('HF_token')
model_id = "selamw/BirdWatcher"
bnb_config = BitsAndBytesConfig(load_in_8bit=True)


def convert_to_markdown(input_text):
    """Converts bird information text to Markdown format, 
       making specific keywords bold and adding headings.
    Args:
        input_text (str): The input text containing bird information.
    Returns:
        str: The formatted Markdown text.
    """
    
    bold_words = ['Look:', 'Cool Fact!:', 'Habitat:', 'Food:', 'Birdie Behaviors:']

    # Split into title and content based on the first ":", handling extra whitespace
    if ":" in input_text:
        title, content = map(str.strip, input_text.split(":", 1))
    else:
        title = input_text
        content = ""

    # Bold the keywords 
    for word in bold_words:
        content = content.replace(word, f'\n\n**{word}')
    
    # Construct the Markdown output with headings
    formatted_output = f"**{title}**{content}" 

    return formatted_output.strip()

    
@spaces.GPU
def infer_fin_pali(image, question):
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    
    model = PaliGemmaForConditionalGeneration.from_pretrained(model_id, quantization_config=bnb_config, token=access_token)
    processor = PaliGemmaProcessor.from_pretrained(model_id, token=access_token)

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

    predictions = model.generate(**inputs, max_new_tokens=512)
    decoded_output = processor.decode(predictions[0], skip_special_tokens=True)[len(question):].lstrip("\n")

    # Ensure proper Markdown formatting
    formatted_output = convert_to_markdown(decoded_output)

    return formatted_output


css = """
    #mkd {
        height: 500px; 
        overflow: auto; 
        border: 1px solid #ccc; 
    }
    h1 {
        text-align: center;
    }
    h3 {
        text-align: center;
    }
    h2 {
        text-align: center;
    }
    span.gray-text {
        color: gray;
    }
"""

with gr.Blocks(css=css) as demo:
    gr.HTML("<h1>🦩  BirdWatcher  🦜</h1>")
    gr.HTML("<h3>[Powered by Fine-tuned PaliGemma]</h3>")
    gr.HTML("<h3>Upload an image of a bird, and the model will generate a detailed description of its species.</h3>")
    gr.HTML("<p style='text-align: center;'>(There are over 11,000 bird species in the world, and this model was fine-tuned with over 500)</p>")
    
    with gr.Tab(label="Bird Identification"):
        with gr.Row():
            input_img = gr.Image(label="Input Bird Image") 
            with gr.Column():
                with gr.Row():
                    question = gr.Text(label="Default Prompt", value="Describe this bird species", elem_id="default-prompt", interactive=True)
                with gr.Row():
                    submit_btn = gr.Button(value="Run")
                with gr.Row():
                    output = gr.Markdown(label="Response")  # Use Markdown component to display output
        
        submit_btn.click(infer_fin_pali, [input_img, question], [output])
        
        gr.Examples(
            [["01.jpg", "Describe this bird species"],
             ["02.jpg", "Describe this bird species"],
             ["03.jpg", "Describe this bird species"],
             ["04.jpg", "Describe this bird species"],
             ["05.jpg", "Describe this bird species"],
             ["06.jpg", "Describe this bird species"]],
            inputs=[input_img, question],
            outputs=[output],
            fn=infer_fin_pali,
            label='Examples 👇'
        )

demo.launch(debug=True, share=True)