app.py

import gradio as gr
import ffmpeg
from pathlib import Path
import os
import ast
import json
import base64
import requests
import moviepy.editor as mp
from PIL import Image, ImageSequence
import cv2


API_URL = "https://api-inference.huggingface.co/models/facebook/wav2vec2-base-960h"
headers = {"Authorization": "Bearer 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"}
#HF_TOKEN = os.environ["HF_TOKEN"]
#headers = {"Authorization": f"Bearer {HF_TOKEN}"}


def generate_transcripts(in_video): #generate_gifs(in_video, gif_transcript):
    print("********* Inside generate_transcripts() **********")
    #convert video to audio
    print(f" input video is : {in_video}")
    
    video_path = Path("./ShiaLaBeouf.mp4")
    audio_memory, _ = ffmpeg.input(video_path).output('-', format="wav", ac=1, ar='16k').overwrite_output().global_args('-loglevel', 'quiet').run(capture_stdout=True)
    
    #Getting transcripts using wav2Vec2 huggingface hosted accelerated inference
    #sending audio file in request along with stride and chunk length information
    model_response = query_api(audio_memory)
    
    #model response has both - transcripts as well as character timestamps or chunks
    print(f"model_response is : {model_response}")
    transcription = model_response["text"].lower()
    chnk = model_response["chunks"]
    
    #creating lists from chunks to consume downstream easily
    timestamps = [[chunk["text"].lower(), chunk["timestamp"][0], chunk["timestamp"][1]]
              for chunk in chnk]
    
    #getting words and word timestamps
    words, words_timestamp = get_word_timestamps(timestamps)
    print(f"Total words in the audio transcript is:{len(words)}, transcript word list is :{words}, type of words is :{type(words)} ") 
    print(f"Total Word timestamps derived fromcharacter timestamp are :{len(words_timestamp)}, Word timestamps are :{words_timestamp}")
    
    return transcription, words, words_timestamp
    
    
def generate_gifs(gif_transcript, words, words_timestamp):
    print("********* Inside generate_gifs() **********")
    
    #creating list from input gif transcript 
    #gif = "don't let your dreams be dreams"
    gif = gif_transcript
    #gif = gif_transcript
    giflist = gif.split()
    
    #getting gif indexes from the generator
    # Converting string to list
    words = ast.literal_eval(words)
    words_timestamp = ast.literal_eval(words_timestamp)
    print(f"words is :{words}")
    print(f"type of words is :{type(words)}")
    print(f"length of words is :{len(words)}")
    print(f"giflist is :{giflist}")

    giflist_indxs = list(list(get_gif_word_indexes(words, giflist))[0])
    print(f"giflist_indxs is : {giflist_indxs}")
    #getting start and end timestamps for a gif video
    start_seconds, end_seconds = get_gif_timestamps(giflist_indxs, words_timestamp)
    print(f"start_seconds, end_seconds  are : ({start_seconds}, {end_seconds})")
    #generated .gif image
    gif_out, vid_out = gen_moviepy_gif(start_seconds, end_seconds)

    return gif_out 

    
#calling the hosted model
def query_api(audio_bytes: bytes):
    """
    Query for Huggingface Inference API for Automatic Speech Recognition task
    """
    print("********* Inside query_api() **********")
    payload = json.dumps({
        "inputs": base64.b64encode(audio_bytes).decode("utf-8"),
        "parameters": {
            "return_timestamps": "char",
            "chunk_length_s": 10,
            "stride_length_s": [4, 2]
        },
        "options": {"use_gpu": False}
    }).encode("utf-8")

    response = requests.request(
        "POST", API_URL, headers=headers, data=payload)
    json_reponse = json.loads(response.content.decode("utf-8"))
    print(f"json_reponse is :{json_reponse}")
    return json_reponse


#getting word timestamps from character timestamps
def get_word_timestamps(timestamps): 
  words, word = [], []
  letter_timestamp, word_timestamp, words_timestamp = [], [], []
  for idx,entry in enumerate(timestamps):
    word.append(entry[0])
    letter_timestamp.append(entry[1])
    if entry[0] == ' ':
      words.append(''.join(word))
      word_timestamp.append(letter_timestamp[0])
      word_timestamp.append(timestamps[idx-1][2])
      words_timestamp.append(word_timestamp)
      word, word_timestamp, letter_timestamp = [], [], []

  words = [word.strip() for word in words]
  return words, words_timestamp


#getting index of gif words in main transcript
def get_gif_word_indexes(total_words_list, gif_words_list):  
    if not gif_words_list:
        return
    # just optimization
    COUNT=0
    lengthgif_words_list = len(gif_words_list)
    firstgif_words_list = gif_words_list[0]
    
    print(f"total_words_list is :{total_words_list}")
    print(f"length of total_words_list is :{len(total_words_list)}")
    print(f"gif_words_list is :{gif_words_list}")
    print(f"length of gif_words_list is :{len(gif_words_list)}")
    
    for idx, item in enumerate(total_words_list):
        COUNT+=1
        if item == firstgif_words_list:
            if total_words_list[idx:idx+lengthgif_words_list] == gif_words_list:
                print(f"value of tuple is : {tuple(range(idx, idx+lengthgif_words_list))}")
                yield tuple(range(idx, idx+lengthgif_words_list))


#getting start and end timestamps for gif transcript
def get_gif_timestamps(giflist_indxs, words_timestamp):    
  print(f"******** Inside get_gif_timestamps() **********")
  min_idx = min(giflist_indxs)
  max_idx = max(giflist_indxs)
  print(f"min_idx is :{min_idx}")
  print(f"max_idx is :{max_idx}")
  
  gif_words_timestamp = words_timestamp[min_idx : max_idx+1]
  print(f"words_timestamp is :{words_timestamp}")
  print(f"gif_words_timestamp is :{gif_words_timestamp}")
  
  start_seconds, end_seconds = gif_words_timestamp[0][0], gif_words_timestamp[-1][-1]
  print(f"start_seconds, end_seconds are :{start_seconds},{end_seconds}")
  
  return start_seconds, end_seconds


#extracting the  video and building and serving a .gif image
def gen_moviepy_gif(start_seconds, end_seconds):
  print("******** inside moviepy_gif () ***************")
  video_path = "./ShiaLaBeouf.mp4"
  video = mp.VideoFileClip(video_path) 
  final_clip = video.subclip(start_seconds, end_seconds)
  
  #writing to RAM
  final_clip.write_gif("gifimage.gif") #, program='ffmpeg', tempfiles=True, fps=15, fuzz=3)
  final_clip.write_videofile("gifimage.mp4")
  final_clip.close()
  #reading in a variable
  gif_img = mp.VideoFileClip("gifimage.gif")
  #gif_vid = mp.VideoFileClip("gifimage.mp4")
  #im = Image.open("gifimage.gif")
  #vid_cap = cv2.VideoCapture('gifimage.mp4')
  return "gifimage.gif", "gifimage.mp4" #im, gif_img, gif_vid, vid_cap,  #"gifimage.mp4"


sample_video = ['./ShiaLaBeouf.mp4']
sample_vid = gr.Video(label='Video file')  #for displaying the example
examples = gr.components.Dataset(components=[sample_vid], samples=[sample_video], type='values')


demo = gr.Blocks()

with demo:
    gr.Markdown("""# **Create Any GIF From Your Favorite Videos!** """)
    gr.Markdown("""
    ### Now you ca get your own unlimited supply of cool GIFs and rections from the videos you most like..
    
    **Motivation and background:** In this Gradio-Space Blog I will be taking you through my efforts in reproducing the brilliant app [Edit Video By Editing Text](https://huggingface.co/spaces/radames/edit-video-by-editing-text) by [@radames](https://huggingface.co/radames). My valule-add are - 
    - A permanent supply for your own new GIFs  
    - This Space written in the form of a Notebook or a Blog if I may, to help someone understand how they can too build this kind of an app.
    
    **How To Use:** 1. Upload a video or simply click on the Shia LaBeouf's sample provided here. 
    2. Then click on 'Generate transcripts' button and first textbox will display the extract Transcript from the audio associated with your sample.
    3. Clip the text from transcript or type manually in the second Textbox provided.
    4. A .Gif image will get generated on the right hand side of animated Shia Labeouf! 
    
    Hopee you have fun using this 😀
    """)
    
    with gr.Row():
        #for incoming video
        input_video = gr.Video(label="Upload a Video", visible=True)  
        #to generate and display transcriptions for input video
        text_transcript = gr.Textbox(label="Transcripts", lines = 10, interactive = True )
        
        #Just to move dgata between function hence keeping visible false
        text_words = gr.Textbox(visible=False)
        text_wordstimestamps = gr.Textbox(visible=False)
        
        #to copy paste required gif transcript / or to populate by itslef on pressing the button
        text_gif_transcript = gr.Textbox(label="Transcripts", placeholder="Copy paste transcripts here to create GIF image" , lines = 3, interactive = True ) 
        
        def load_gif_text(text):
            print("****** inside load_gif_text() ******")
            print("text for gif is : ", text)
            return text
             
        text_transcript.change(load_gif_text, text_transcript, text_gif_transcript )
        
        out_gif = gr.Image(label="Generated GIF image") 
                
        
    with gr.Row():
        button_transcript = gr.Button("Generate transcripts")
        button_gifs = gr.Button("Create Gif")
        
    with gr.Row():
        #to render video example on mouse hover/click        
        examples.render()
        #to load sample video into input_video upon clicking on it
        def load_examples(video):  
            print("****** inside load_example() ******")
            print("in_video is : ", video[0])
            return video[0]
        
        examples.click(load_examples, examples, input_video) 
    
    with gr.Row():
        gr.Markdown(""" I will start with a short note on my understanding of Radames's app and tools used in it - 
    
          - His is a supercool and handy proof of concept of a simple video editor where you can edit a video by playing with its audio transcriptions (ASR pipeline output). 
          - Both of our apps uses **Huggingface's [Automatic Speech Recognition Pipeline]**(https://huggingface.co/tasks/automatic-speech-recognition) build over **Wav2Vec2** model which internally uses CTC to improve predictions. The pipeline allows you to predict text transcriptions along with the timestamps for every characters and pauses that are there in the audio text. 
          - His app uses FFmpeg library to a good extent to clip and merge videos. FFmpeg is an open-source library for video handling consisting of a suite of functions for handling video, audio, and other multimedia files. My app uses FFmpeg as well as Moviepy to do the bulk of video+audio processing. 
          
          Let me now briefly take you through the code and process involved in building this app *step by step* 😉 lol -
          - Firstly, I have used ffmpeg to extract audio from video (this code line is directly from Radames's above app) -
         
         ``` 
         audio_memory, _ = ffmpeg.input(video_path).output('-', format="wav", ac=1, ar='16k').overwrite_output().global_args('-loglevel', 'quiet').run(capture_stdout=True)
         ```
        - Then I am calling the ASR model as a service, using the Accelerated Inference API. Below is the code snippet for doing so -
        
        ```
         def query(in_audio):
            payload = json.dumps({ "inputs": base64.b64encode(in_audio).decode("utf-8"),  
                "parameters": {
                    "return_timestamps": "char",
                    "chunk_length_s": 10,
                    "stride_length_s": [4, 2]
                },
                "options": {"use_gpu": False}
            }).encode("utf-8")
        
            response = requests.request("POST", API_URL, data=payload) 
            
            json_response = json.loads(response.content.decode("utf-8"))
            
            return json_response
         ```
         - The transcript thus generated might have some words which are not correctly interpreted, for example, *tomorrow* is translated as 'to morrow', *hard at it* is translated as 'hot ati' and so on. However this won't hinder in the use-case I am demoing here, so we let's move on. 
         
         > do it just do it don't let your dreams be dreams yesterday you said to morrow so just do it make you dreams can't yro just do it some people dream of success while you're going to wake up and work hot ati nothing is impossible you should get to the point where any one else would quit and you're luck in a stop there no what are you waiting for do et jot do it just you can just do it if you're tired is starting over stop giving up
         
         - The other output generated by this ASR pipeline is a list of character timestamps dictionaries, look at the below sample to get an idea -
         
         ```
         {'text': 'D', 'timestamp': [2.36, 2.38]},
         {'text': 'O', 'timestamp': [2.52, 2.56]},
         {'text': ' ', 'timestamp': [2.68, 2.72]},
         {'text': 'I', 'timestamp': [2.84, 2.86]},
         {'text': 'T', 'timestamp': [2.88, 2.92]},
         {'text': ' ', 'timestamp': [2.94, 2.98]},
         {'text': 'J', 'timestamp': [4.48, 4.52]},
         ```
         
         - Next, using these character timestamps I have extracted word timestamps (by taking the start_timestamp of the first letters and the en_timestamp of the last letter in any give word.
             - Further when a *sub-transcript* is provided for the producing the GIF, I calculated the start and end timestamp for the whole group of words.
         
         - I have then used *moviepy* library to extract / concat videos into smaller clips and also to save the final processed videofile as a.GIF image.
         ```
          import moviepy.editor as mp
          
          video = mp.VideoFileClip(video_path) 
          final_clip = video.subclip(start_seconds, end_seconds)
          
          #writing to RAM
          final_clip.write_gif("gifimage.gif") #, program='ffmpeg', tempfiles=True, fps=15, fuzz=3)
          final_clip.write_videofile("gifimage.mp4")
          final_clip.close()

         ```
         
         While working on apps for [Gradio Blocks Party](https://huggingface.co/Gradio-Blocks) I have gained a tremendous amount of knowledge about Gradio and Huggingface APIs and infrastructure. I was also able to polish my understanding and learn new things on some of the key and most interesting ML aspects like Question Answering, Sentence Trnasformers, Summarization, Image Generation, LLMs, Prompt Engineering, and now ASR and Video processing. 
         I absolutely love Spaces, I believe Spaces is much more than a platform to showcase your ML demos. I suppose it can act like an ML Product Sandbox with the benefits whole of Huggingface might and infra behind it. I believe Spaces can become sort of a playground for future ML products and ideas. ALl of this is extremely exciting.
         
         Thanks for reading so far, I will see you at my next submission. Keep learning and sharing.
         
         My last two Gradio Blocks Party apps can be found here - 
         
         - [Gradio-Blocks/GPTJ6B_Poetry_LatentDiff_Illustration](https://huggingface.co/spaces/Gradio-Blocks/GPTJ6B_Poetry_LatentDiff_Illustration), and 
         - [Gradio-Blocks/Ask_Questions_To_YouTube_Videos](https://huggingface.co/spaces/Gradio-Blocks/Ask_Questions_To_YouTube_Videos)
        
          
        """)
        
    button_transcript.click(generate_transcripts, input_video, [text_transcript, text_words, text_wordstimestamps ])
    button_gifs.click(generate_gifs, [text_gif_transcript, text_words, text_wordstimestamps], out_gif )
    
   
demo.launch(debug=True)