app.py

import gradio as gr
from PIL import Image
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
import cv2
import numpy as np
import ast
from collections import Counter


# # Ensure GPU usage if available
device = "cuda" if torch.cuda.is_available() else "cpu"

# Initialize the model and tokenizer
model = AutoModelForCausalLM.from_pretrained("ManishThota/SparrowVQE",
                                             torch_dtype=torch.float16, 
                                             device_map="auto",
                                             trust_remote_code=True)
tokenizer = AutoTokenizer.from_pretrained("ManishThota/SparrowVQE", trust_remote_code=True)




def video_to_frames(video, fps=1):
    """Converts a video file into frames and stores them as PNG images in a list."""
    frames_png = []
    cap = cv2.VideoCapture(video)
    
    if not cap.isOpened():
        print("Error opening video file")
        return frames_png
    
    frame_count = 0
    frame_interval = int(cap.get(cv2.CAP_PROP_FPS)) // fps  # Calculate frame interval
    
    while cap.isOpened():
        ret, frame = cap.read()
        if not ret:
            print("Can't receive frame (stream end?). Exiting ...")
            break
        
        if frame_count % frame_interval == 0:
            is_success, buffer = cv2.imencode(".png", frame)
            if is_success:
                frames_png.append(np.array(buffer).tobytes())
        
        frame_count += 1
    
    cap.release()
    return frames_png

def extract_frames(frame):

    # Convert binary data to a numpy array
    frame_np = np.frombuffer(frame, dtype=np.uint8)

    # Decode the PNG image
    image_rgb = cv2.imdecode(frame_np, flags=cv2.IMREAD_COLOR)  # Assuming it's in RGB format

    # Convert RGB to BGR
    image_bgr = cv2.cvtColor(image_rgb, cv2.COLOR_RGB2BGR)

    return image_bgr

def predict_answer(video, image, question):

    text = f"A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions. USER: <image>\n{question}? ASSISTANT:"
    input_ids = tokenizer(text, return_tensors='pt').input_ids.to(device)

    
    if image:
        # Process as an image
        image = image.convert("RGB")
        image_tensor = model.image_preprocess(image)
        
        #Generate the answer
        output_ids = model.generate(
            input_ids,
            max_new_tokens=25,
            images=image_tensor,
            use_cache=True)[0]
        
        return tokenizer.decode(output_ids[input_ids.shape[1]:], skip_special_tokens=True).strip()
        
    elif video:
        frames = video_to_frames(video)
        image = extract_frames(frames[2])
        image_tensor = model.image_preprocess([image])
        # Generate the answer
        output_ids = model.generate(
                input_ids,
                max_new_tokens=25,
                images=image_tensor,
                use_cache=True)[0]
        answer = tokenizer.decode(output_ids[input_ids.shape[1]:], skip_special_tokens=True).strip()

        return answer
        # # Process as a video
        # frames = video_to_frames(video)
        # answers = []
        # for frame in frames:
        #     image = extract_frames(frame)
        #     image_tensor = model.image_preprocess([image])
            
        #     # Generate the answer
        #     output_ids = model.generate(
        #         input_ids,
        #         max_new_tokens=25,
        #         images=image_tensor,
        #         use_cache=True)[0]
            
        #     answer = tokenizer.decode(output_ids[input_ids.shape[1]:], skip_special_tokens=True).strip()
        #     answers.append(answer)

        # # Modify this logic based on your specific needs
        # most_common_answer = Counter(answers).most_common(1)[0][0]
        
        # # Safely evaluate the most common answer assuming it's a string representation of a Python literal
        # try:
        #     evaluated_answer = ast.literal_eval(most_common_answer)
        # except (ValueError, SyntaxError):
        #     # Handle malformed answer string
        #     evaluated_answer = f"Error evaluating answer: {most_common_answer}"
        
        # return evaluated_answer
    
    #     return ast.literal_eval(answers[0])
        
    # else:
    #     return "Unsupported file type. Please upload an image or video."
        
promt_cat_dog = """
    Annotate this image with this schema:
    {
        “description”: “Is there a cat in the image?”,
        “value”: “Cat”
    },
    {
        “description”: “Is there a dog in the image?”,
        “value”: “Dog”,
    },
    {
        “description”: “Is there a horse in the image?”,
        “value”: “Horse”,
    },
provide me the answers as a dictionary with key as the string value of the variable value on top and its value should be boolean value        
"""
promt_bus_people = """
    Annotate this image with this schema:
    {
        “description”: “Is there a bus in the image?”,
        “value”: “Bus”,
    },
    {
        “description”: “Is there a bike in the image?”,
        “value”: “Bike”,
    },
provide me the answers as a dictionary with key as the string value of the variable value on top and its value should be boolean value        
"""

# promt_video = """
#     Annotate this image with this schema:
#     {
#         “description”: “Is the person standing?”,
#         “value”: “standing”,
#     },
#     {
#         “description”: “Is the person's hands free?”,
#         “value”: “Hands-Free”,
#     },
# provide me the answers as a dictionary with key as the string value of the variable value on top and its value should be boolean value
# """

promt_video = """
    Annotate this image by indicating the presence or absence of specific objects.
    {
        “description”: “Is the person standing?”,
        “value”: “standing”,
    },
    {
        “description”: “Is the person's hands free?”,
        “value”: “Hands-Free”,
    },
Provide your answers as a dictionary with the object type as the key and a boolean value indicating its presence in the image, Use 'true' for objects present in the image and 'false' for objects not present.

"""


test_examples = [[None, "Images/cat_dog.jpeg", promt_cat_dog], 
            [None,"Images/bus_people.jpeg", promt_bus_people], 
            ["videos/v2.mp4",None,promt_video],
            ["videos/v3.mp4",None,promt_video]]


def gradio_predict(video,image, question):
    answer = predict_answer(video,image, question)
    return answer

css = """
#container{
    display: block;
    margin-left: auto;
    margin-right: auto;
    width: 60%;
}
#intro{
    max-width: 100%;
    margin: 0 auto;
    text-align: center;
}

"""
with gr.Blocks(css = css) as app:
    with gr.Row(elem_id="container"):
        gr.Image("gsoc_redhen.png",min_width=60, label="GSOC 2024")
    gr.Markdown("""  
    ## This Gradio app serves as four folds:
    ### 1. My ability and experience to design a customizable Gradio application with Interface/Blocks structure.
    ### 2. One of my Multimodel Vision-Language model's capabilities with the LLaVA framework.
    ### 3. Demo for annotating random images and 4 second videos provided at Notion (https://shorturl.at/givyC) 
    ### 4. Ability to integrate a Large Language Model and Vision Encoder        
                 """)
    with gr.Row():
        video = gr.Video(label="Video")
        image = gr.Image(type="pil", label="Image")
    with gr.Row():
        with gr.Column():
            question = gr.Textbox(label="Question", placeholder="Annotate prompt", lines=4.3)
            btn = gr.Button("Annotate")
        with gr.Column():
            answer = gr.TextArea(label="Answer")
    
    
    btn.click(gradio_predict, inputs=[video,image, question], outputs=answer)

    gr.Examples(
        examples=test_examples,
        inputs=[video,image, question],
        outputs= answer,
        fn=gradio_predict,
        cache_examples=True,
    )

app.launch(debug=True)