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-AI"
# model_id = "selamw/bird-Identifier"

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
    # title, content = map(str.strip, input_text.split(":", 1))
    # # Bold the keywords 
    # for word in bold_words:
    #     content = content.replace(word, f'\n**{word}**\n')
    #     content = content.replace(f'** ', f' ')
    
    # # Construct the Markdown output with headings
    # formatted_output = f"**{title}{content}" 
    input_text = """**ABBOTT'S BABBLER (Malacocincla abbotti)** \n\n **Look:** \n  Robin-sized detective! This bird has a sandy-brown body with rusty flanks, a short tail, and a heavy hooked bill. Check for a pale gray eyebrow in Southeast Asia. \n\n **Cool Fact!:** \n  Works in pairs, hopping on the forest floor like little detectives searching for clues (their prey!). \n\n **Habitat:** \n  Prefers the shady undergrowth of permanent forests, especially near streams and tangled vegetation. \n\n **Food:** \n  Not picky eaters! They enjoy a varied menu of insects, worms, and even seeds. \n\n **Birdie Behaviors:** \n  Secretive by nature, but listen for their surprising calls – a mix of harsh churrs and melodic whistles."""


    return input_text

@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)
    # formatted_output = (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: left;
    }
    span.gray-text {
        color: gray;
    }
"""

with gr.Blocks(css=css) as demo:
    gr.HTML("<h1>🦩 BirdWatcher AI 🦜</h1>")
    gr.HTML("<h3>Upload an image of a bird, and the model will generate a detailed description of its species.</h3>")
    
    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")
                with gr.Row():
                    submit_btn = gr.Button(value="Run")
                with gr.Row():
                    output = gr.Markdown(label="Response")  # Use Markdown component to display output
                    # output = gr.Text(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.jpeg", "Describe this bird species"]],
            inputs=[input_img, question],
            outputs=[output],
            fn=infer_fin_pali,
            label='Examples 👇'
        )

demo.launch(debug=True)