app.py

import spaces
import gradio as gr
from gradio_client import Client
import cv2
from moviepy.editor import *

# 1. extract and store 1 image every 5 images from video input
# 2. extract audio
# 3. for each image from extracted_images, get caption from caption model and concatenate into list
# 4. for audio, ask audio questioning model to describe sound/scene
# 5. give all to LLM, and ask it to resume, according to image caption list combined to audio caption

import re
import torch
from transformers import pipeline

zephyr_model = "HuggingFaceH4/zephyr-7b-beta"
pipe = pipeline("text-generation", model=zephyr_model, torch_dtype=torch.bfloat16, device_map="auto")

standard_sys = f"""
You will be provided a list of visual details observed at regular intervals, along with an audio description. These pieces of information originate from a single video. The visual details are extracted from the video at fixed time intervals and represent consecutive frames. Typically, the video consists of a brief sequence showing one or more subjects...

Please note that the following list of image descriptions (visual details) was obtained by extracting individual frames from a continuous video featuring one or more subjects. Depending on the case, all depicted individuals may correspond to the same person(s), with minor variations due to changes in lighting, angle, and facial expressions over time. Regardless, assume temporal continuity among the frames unless otherwise specified.

Audio events are actual recordings from the video, representing sounds and spoken words independent of the visuals. While audio events offer rich context and background information, elucidating the environment and ambient noises, the visual representation tends to focus mainly on the primary subjects. Despite the high likelihood of alignment, there might be rare occasions where audio information doesn't precisely match the visual aspect. In such circumstances, prioritize visual evidence, and cautiously incorporate seemingly incongruous auditory clues into your summary. Exercise vigilance when reconciling conflicts and maintain a strong commitment to fidelity in generating a comprehensive overview. Your job is to integrate these multimodal inputs intelligently and provide a very short resume about what is happening in the origin video. Provide a succinct yet thorough overview of what you understood.   
"""

def extract_frames(video_in, interval=24, output_format='.jpg'):
    """Extract frames from a video at a specified interval and store them in a list.

    Args:
    - video_in: string or path-like object pointing to the video file
    - interval: integer specifying how many frames apart to extract images (default: 5)
    - output_format: string indicating desired format for saved images (default: '.jpg')

    Returns:
    A list of strings containing paths to saved images.
    """

    # Initialize variables
    vidcap = cv2.VideoCapture(video_in)
    frames = []
    count = 0

    # Loop through frames until there are no more
    while True:
        success, image = vidcap.read()

        # Check if successful read and not past end of video
        if success:
            print('Read a new frame:', success)

            # Save current frame if it meets criteria
            if count % interval == 0:
                filename = f'frame_{count // interval}{output_format}'
                frames.append(filename)
                cv2.imwrite(filename, image)
                print(f'Saved {filename}')

            # Increment counter
            count += 1

        # Break out of loop when done reading frames
        else:
            break

    # Close video capture
    vidcap.release()
    print('Done extracting frames!')

    return frames

from transformers import AutoModelForCausalLM, AutoTokenizer
from PIL import Image

model_id = "vikhyatk/moondream2"
revision = "2024-03-06"
model = AutoModelForCausalLM.from_pretrained(
    model_id, trust_remote_code=True, revision=revision
)
tokenizer = AutoTokenizer.from_pretrained(model_id, revision=revision)

@spaces.GPU()
def process_image(image_in):
    '''
    client = Client("https://vikhyatk-moondream1.hf.space/")
    result = client.predict(
		image_in,	# filepath  in 'image' Image component
		"Describe precisely the image in one sentence.",	# str  in 'Question' Textbox component
		api_name="/answer_question"
        #api_name="/predict"
    )
    print(result)
    return result
    '''
    image = Image.open(image_in)
    enc_image = model.encode_image(image)
    result = model.answer_question(enc_image, "Describe the image in one sentence.", tokenizer)
    print(result)
    return result

def extract_audio(video_path):
    video_clip = VideoFileClip(video_path)
    audio_clip = video_clip.audio
    audio_clip.write_audiofile("output_audio.mp3")
    return "output_audio.mp3"

def get_salmonn(audio_in):
    salmonn_prompt = "Please describe the audio"
    client = Client("fffiloni/SALMONN-7B-gradio")
    result = client.predict(
    		audio_in,	# filepath  in 'Audio' Audio component
    		salmonn_prompt,	# str  in 'User question' Textbox component
    		4,	# float (numeric value between 1 and 10) in 'beam search numbers' Slider component
    		1,	# float (numeric value between 0.8 and 2.0) in 'temperature' Slider component
    		0.9,	# float (numeric value between 0.1 and 1.0) in 'top p' Slider component
    		api_name="/gradio_answer"
    )
    print(result)
    return result

@spaces.GPU()
def llm_process(user_prompt):
    agent_maker_sys = standard_sys
    
    instruction = f"""
<|system|>
{agent_maker_sys}</s>
<|user|>
"""
    
    prompt = f"{instruction.strip()}\n{user_prompt}</s>"    
    outputs = pipe(prompt, max_new_tokens=256, do_sample=True, temperature=0.7, top_k=50, top_p=0.95)
    pattern = r'\<\|system\|\>(.*?)\<\|assistant\|\>'
    cleaned_text = re.sub(pattern, '', outputs[0]["generated_text"], flags=re.DOTALL)
    
    print(f"SUGGESTED video description: {cleaned_text}")
    return cleaned_text.lstrip("\n")

def infer(video_in):
    # Extract frames from a video
    frame_files = extract_frames(video_in)
    
    # Process each extracted frame and collect results in a list
    processed_texts = []
    for frame_file in frame_files:
        text = process_image(frame_file)
        processed_texts.append(text)
    print(processed_texts)

    # Convert processed_texts list to a string list with line breaks
    string_list = '\n'.join(processed_texts)

    # Extract audio from video
    extracted_audio = extract_audio(video_in)
    print(extracted_audio)

    # Get description of audio content
    audio_content_described = get_salmonn(extracted_audio)

    # Assemble captions
    formatted_captions = f"""
### Visual events:\n{string_list}\n ### Audio events:\n{audio_content_described}
"""
    print(formatted_captions)

    # Send formatted captions to LLM
    video_description_from_llm = llm_process(formatted_captions)
    
    return video_description_from_llm

with gr.Blocks() as demo :
    with gr.Column(elem_id="col-container"):
        gr.HTML("""
        <h2 style="text-align: center;">Video description</h2>
        """)
        video_in = gr.Video(label="Video input")
        submit_btn = gr.Button("Submit")
        video_description = gr.Textbox(label="Video description")
    submit_btn.click(
        fn = infer,
        inputs = [video_in],
        outputs = [video_description]
    )
demo.queue().launch()