app_parallel.py

from flask import Flask, request, jsonify, stream_with_context
import torch
import shutil
import os
import sys
from time import strftime
from src.utils.preprocess import CropAndExtract
from src.test_audio2coeff import Audio2Coeff
from src.facerender.animate import AnimateFromCoeff
from src.generate_batch import get_data
from src.generate_facerender_batch import get_facerender_data
# from src.utils.init_path import init_path
import tempfile
from openai import OpenAI
import elevenlabs
from elevenlabs import set_api_key, generate, play, clone, Voice, VoiceSettings
import uuid
import time
from PIL import Image
import moviepy.editor as mp
import requests
import json
import pickle
# from dotenv import load_dotenv
from concurrent.futures import ProcessPoolExecutor, as_completed, ThreadPoolExecutor

# Load environment variables from .env file
# load_dotenv()

# Initialize ProcessPoolExecutor for parallel processing
executor = ThreadPoolExecutor(max_workers=3)
torch.cuda.empty_cache()

class AnimationConfig:
    def __init__(self, driven_audio_path, source_image_path, result_folder,pose_style,expression_scale,enhancer,still,preprocess,ref_pose_video_path, image_hardcoded):
        self.driven_audio = driven_audio_path
        self.source_image = source_image_path
        self.ref_eyeblink = None
        self.ref_pose = None
        self.checkpoint_dir = './checkpoints'
        self.result_dir = result_folder
        self.pose_style = pose_style
        self.batch_size = 2
        self.expression_scale = expression_scale
        self.input_yaw = None
        self.input_pitch = None
        self.input_roll = None
        self.enhancer = enhancer
        self.background_enhancer = None
        self.cpu = False
        self.face3dvis = False
        self.still = still  
        self.preprocess = preprocess
        self.verbose = False
        self.old_version = False
        self.net_recon = 'resnet50'
        self.init_path = None
        self.use_last_fc = False
        self.bfm_folder = './checkpoints/BFM_Fitting/'
        self.bfm_model = 'BFM_model_front.mat'
        self.focal = 1015.
        self.center = 112.
        self.camera_d = 10.
        self.z_near = 5.
        self.z_far = 15.
        self.device = 'cuda'
        self.image_hardcoded = image_hardcoded


app = Flask(__name__)
# CORS(app)

TEMP_DIR = None
start_time = None
future_to_chunk = {}

app.config['temp_response'] = None
app.config['generation_thread'] = None
app.config['text_prompt'] = None
app.config['final_video_path'] = None
app.config['final_video_duration'] = None

# Global paths
dir_path = os.path.dirname(os.path.realpath(__file__))
current_root_path = dir_path

path_of_lm_croper = os.path.join(current_root_path, 'checkpoints', 'shape_predictor_68_face_landmarks.dat')
path_of_net_recon_model = os.path.join(current_root_path, 'checkpoints', 'epoch_20.pth')
dir_of_BFM_fitting = os.path.join(current_root_path, 'checkpoints', 'BFM_Fitting')
wav2lip_checkpoint = os.path.join(current_root_path, 'checkpoints', 'wav2lip.pth')
audio2pose_checkpoint = os.path.join(current_root_path, 'checkpoints', 'auido2pose_00140-model.pth')
audio2pose_yaml_path = os.path.join(current_root_path, 'src', 'config', 'auido2pose.yaml')
audio2exp_checkpoint = os.path.join(current_root_path, 'checkpoints', 'auido2exp_00300-model.pth')
audio2exp_yaml_path = os.path.join(current_root_path, 'src', 'config', 'auido2exp.yaml')
free_view_checkpoint = os.path.join(current_root_path, 'checkpoints', 'facevid2vid_00189-model.pth.tar')


# Function for running the actual task (using preprocessed data)
def process_chunk(audio_chunk, preprocessed_data, args):
    print("Entered Process Chunk Function")
    global audio2pose_checkpoint, audio2pose_yaml_path, audio2exp_checkpoint, audio2exp_yaml_path, wav2lip_checkpoint
    global free_view_checkpoint
    if args.preprocess == 'full':
        mapping_checkpoint = os.path.join(current_root_path, 'checkpoints', 'mapping_00109-model.pth.tar')
        facerender_yaml_path = os.path.join(current_root_path, 'src', 'config', 'facerender_still.yaml')
    else:
        mapping_checkpoint = os.path.join(current_root_path, 'checkpoints', 'mapping_00229-model.pth.tar')
        facerender_yaml_path = os.path.join(current_root_path, 'src', 'config', 'facerender.yaml')

    first_coeff_path = preprocessed_data["first_coeff_path"]
    crop_pic_path = preprocessed_data["crop_pic_path"]
    crop_info_path = preprocessed_data["crop_info"]
    with open(crop_info_path , "rb") as f:
            crop_info = pickle.load(f)

    print(f"Loaded existing preprocessed data")
    print("first_coeff_path",first_coeff_path)
    print("crop_pic_path",crop_pic_path)
    print("crop_info",crop_info)
    torch.cuda.empty_cache()
    batch = get_data(first_coeff_path, audio_chunk, args.device, ref_eyeblink_coeff_path=None, still=args.still)
    audio_to_coeff = Audio2Coeff(audio2pose_checkpoint, audio2pose_yaml_path, 
                                audio2exp_checkpoint, audio2exp_yaml_path, 
                                wav2lip_checkpoint, args.device)
    coeff_path = audio_to_coeff.generate(batch, args.result_dir, args.pose_style, ref_pose_coeff_path=None)
    
    # Further processing with animate_from_coeff using the coeff_path
    animate_from_coeff = AnimateFromCoeff(free_view_checkpoint, mapping_checkpoint, 
                                            facerender_yaml_path, args.device)

    torch.cuda.empty_cache()
    data = get_facerender_data(coeff_path, crop_pic_path, first_coeff_path, audio_chunk, 
                                args.batch_size, args.input_yaw, args.input_pitch, args.input_roll, 
                                expression_scale=args.expression_scale, still_mode=args.still, preprocess=args.preprocess)
    torch.cuda.empty_cache()
    print("Will Enter Animation")
    result, base64_video, temp_file_path, _ = animate_from_coeff.generate(data, args.result_dir, args.source_image, crop_info, 
                                    enhancer=args.enhancer, background_enhancer=args.background_enhancer, preprocess=args.preprocess)

    # video_clip = mp.VideoFileClip(temp_file_path)
    # duration = video_clip.duration
    
    app.config['temp_response'] = base64_video
    app.config['final_video_path'] = temp_file_path
    # app.config['final_video_duration'] = duration
    torch.cuda.empty_cache()
    return base64_video, temp_file_path


def create_temp_dir():
    return tempfile.TemporaryDirectory()

def save_uploaded_file(file, filename,TEMP_DIR):
    print("Entered save_uploaded_file")
    unique_filename = str(uuid.uuid4()) + "_" + filename
    file_path = os.path.join(TEMP_DIR.name, unique_filename)
    file.save(file_path)
    return file_path


def custom_cleanup(temp_dir):
    # Iterate over the files and directories in TEMP_DIR
    for filename in os.listdir(temp_dir):
        file_path = os.path.join(temp_dir, filename)
        if os.path.isdir(file_path):
            shutil.rmtree(file_path)
        else:
            os.remove(file_path)
        print(f"Deleted: {file_path}")
    
    torch.cuda.empty_cache()
    import gc
    gc.collect()


def generate_audio(voice_cloning, voice_gender, text_prompt):
    print("generate_audio")
    if voice_cloning == 'no':
        if voice_gender == 'male':
            voice = 'echo'
            print('Entering Audio creation using elevenlabs')
            set_api_key('92e149985ea2732b4359c74346c3daee')

            audio = generate(text = text_prompt, voice = "Daniel", model = "eleven_multilingual_v2",stream=True, latency=4)
            with tempfile.NamedTemporaryFile(suffix=".mp3", prefix="text_to_speech_",dir=TEMP_DIR.name, delete=False) as temp_file:
                for chunk in audio:
                    temp_file.write(chunk)
                driven_audio_path = temp_file.name
                print('driven_audio_path',driven_audio_path)
                print('Audio file saved using elevenlabs')
                    
        else:
            voice = 'nova'

            print('Entering Audio creation using whisper')
            response = client.audio.speech.create(model="tts-1-hd",
                                            voice=voice,
                                            input = text_prompt)

            print('Audio created using whisper')
            with tempfile.NamedTemporaryFile(suffix=".wav", prefix="text_to_speech_",dir=TEMP_DIR.name, delete=False) as temp_file:
                driven_audio_path = temp_file.name
            
            response.write_to_file(driven_audio_path)
            print('Audio file saved using whisper')
    
    elif voice_cloning == 'yes':
        set_api_key('92e149985ea2732b4359c74346c3daee')
        # voice = clone(name = "User Cloned Voice",
        #             files = [user_voice_path] )
        voice = Voice(voice_id="CEii8R8RxmB0zhAiloZg",name="Marc",settings=VoiceSettings(
                        stability=0.71, similarity_boost=0.5, style=0.0, use_speaker_boost=True),)

        audio = generate(text = text_prompt, voice = voice, model = "eleven_multilingual_v2",stream=True, latency=4)
        with tempfile.NamedTemporaryFile(suffix=".mp3", prefix="cloned_audio_",dir=TEMP_DIR.name, delete=False) as temp_file:
            for chunk in audio:
                temp_file.write(chunk)
            driven_audio_path = temp_file.name
            print('driven_audio_path',driven_audio_path)

    return driven_audio_path

# Preprocessing step that runs only once
def run_preprocessing(args):
    global path_of_lm_croper, path_of_net_recon_model, dir_of_BFM_fitting
    first_frame_dir = os.path.join(args.result_dir, 'first_frame_dir')
    os.makedirs(first_frame_dir, exist_ok=True)

    # Check if preprocessed data already exists
    fixed_temp_dir = "C:/Users/fd01076/Downloads/preprocess_data"
    os.makedirs(fixed_temp_dir, exist_ok=True)
    preprocessed_data_path = os.path.join(fixed_temp_dir, "preprocessed_data.pkl")

    if os.path.exists(preprocessed_data_path) and args.image_hardcoded == "yes":
        with open(preprocessed_data_path, "rb") as f:
            preprocessed_data = pickle.load(f)
        print("Loaded existing preprocessed data from:", preprocessed_data_path)
    else:
        preprocess_model = CropAndExtract(path_of_lm_croper, path_of_net_recon_model, dir_of_BFM_fitting, args.device)
        first_coeff_path, crop_pic_path, crop_info = preprocess_model.generate(args.source_image, first_frame_dir, args.preprocess, source_image_flag=True)

        if not first_coeff_path:
            raise Exception("Failed to get coefficients")

        # Save the preprocessed data
        preprocessed_data = {
            "first_coeff_path": first_coeff_path,
            "crop_pic_path": crop_pic_path,
            "crop_info": crop_info
        }
        with open(preprocessed_data_path, "wb") as f:
            pickle.dump(preprocessed_data, f)

    return preprocessed_data

client = OpenAI(api_key="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")

def openai_chat_avatar(text_prompt):
    response = client.chat.completions.create(
        model="gpt-4o-mini",
        messages=[{"role": "system", "content": "Ensure answers are concise, human-like, and clear while maintaining quality. Use the fewest possible words, avoiding unnecessary articles, prepositions, and adjectives. Responses should be short but still address the question thoroughly without being verbose.Keep them to one sentence only"},
            {"role": "user", "content": f"Hi! I need help with something. {text_prompt}"},
        ],
        max_tokens = len(text_prompt) + 300 # Use the length of the input text
        # temperature=0.3,
        # stop=["Translate:", "Text:"]
    )
    return response
    
def split_audio(audio_path, chunk_duration):
    audio_clip = mp.AudioFileClip(audio_path)
    total_duration = audio_clip.duration

    audio_chunks = []
    for start_time in range(0, int(total_duration), chunk_duration):
        end_time = min(start_time + chunk_duration, total_duration)
        chunk = audio_clip.subclip(start_time, end_time)
        with tempfile.NamedTemporaryFile(suffix=f"_chunk_{start_time}-{end_time}.wav", prefix="audio_chunk_", dir=TEMP_DIR.name, delete=False) as temp_file:
            chunk_path = temp_file.name
            chunk.write_audiofile(chunk_path)
            audio_chunks.append((start_time, chunk_path))
    
    return audio_chunks

# Generator function to yield chunk results as they are processed
def generate_chunks(audio_chunks, preprocessed_data, args):
    global TEMP_DIR
    future_to_chunk = {executor.submit(process_chunk, chunk[1], preprocessed_data, args): chunk[0] for chunk in audio_chunks}
    
    try:
        for future in as_completed(future_to_chunk):
            idx = future_to_chunk[future]  # Get the original chunk that was processed
            try:
                base64_video, temp_file_path = future.result()  # Get the result of the completed task
                yield json.dumps({'start_time': idx, 'base64_video': base64_video}).encode('utf-8')
            except Exception as e:
                yield f"Task for chunk {idx} failed: {e}\n"
    finally:
        if TEMP_DIR:
            custom_cleanup(TEMP_DIR.name)

@app.route("/run", methods=['POST'])
def parallel_processing():
    global start_time, future_to_chunk
    start_time = time.time()
    global TEMP_DIR
    global audio_chunks
    TEMP_DIR = create_temp_dir()
    print('request:',request.method)
    try:
        if request.method == 'POST':
            # source_image = request.files['source_image']
            image_path = '/home/user/app/images/out.jpg'
            source_image = Image.open(image_path)
            text_prompt = request.form['text_prompt']
            
            print('Input text prompt: ',text_prompt)
            text_prompt = text_prompt.strip()
            if not text_prompt:
                return jsonify({'error': 'Input text prompt cannot be blank'}), 400
                
            voice_cloning = request.form.get('voice_cloning', 'yes')
            image_hardcoded = request.form.get('image_hardcoded', 'no')
            chat_model_used = request.form.get('chat_model_used', 'openai')
            target_language = request.form.get('target_language', 'original_text')
            print('target_language',target_language)
            pose_style = int(request.form.get('pose_style', 1))
            expression_scale = float(request.form.get('expression_scale', 1))
            enhancer = request.form.get('enhancer', None)
            voice_gender = request.form.get('voice_gender', 'male')
            still_str = request.form.get('still', 'False')
            still = still_str.lower() == 'false'
            print('still', still)
            preprocess = request.form.get('preprocess', 'crop')
            print('preprocess selected: ',preprocess)
            # ref_pose_video = request.files.get('ref_pose', None)

            # response = openai_chat_avatar(text_prompt)
            # text_prompt = response.choices[0].message.content.strip()
            app.config['text_prompt'] = text_prompt
            print('Final output text prompt using openai: ',text_prompt)
    
            source_image_path = save_uploaded_file(source_image, 'source_image.png',TEMP_DIR)
            print(source_image_path)
    
            driven_audio_path = generate_audio(voice_cloning, voice_gender, text_prompt)
        
            save_dir = tempfile.mkdtemp(dir=TEMP_DIR.name)
            result_folder = os.path.join(save_dir, "results")
            os.makedirs(result_folder, exist_ok=True)
    
            ref_pose_video_path = None
            # if ref_pose_video:
            #     with tempfile.NamedTemporaryFile(suffix=".mp4", prefix="ref_pose_",dir=TEMP_DIR.name, delete=False) as temp_file:
            #         ref_pose_video_path = temp_file.name
            #         ref_pose_video.save(ref_pose_video_path)
            #         print('ref_pose_video_path',ref_pose_video_path)
                    
    except Exception as e:
        app.logger.error(f"An error occurred: {e}")
        return jsonify({'status': 'error', 'message': str(e)}), 500
    
    args = AnimationConfig(driven_audio_path=driven_audio_path, source_image_path=source_image_path, result_folder=result_folder, pose_style=pose_style, expression_scale=expression_scale,enhancer=enhancer,still=still,preprocess=preprocess,ref_pose_video_path=ref_pose_video_path, image_hardcoded=image_hardcoded)
    
    preprocessed_data = run_preprocessing(args)
    chunk_duration = 3
    print(f"Splitting the audio into {chunk_duration}-second chunks...")
    audio_chunks = split_audio(driven_audio_path, chunk_duration=chunk_duration)
    print(f"Audio has been split into {len(audio_chunks)} chunks: {audio_chunks}")

    # future_to_chunk = {executor.submit(process_chunk, chunk[1], preprocessed_data, args): chunk[0] for chunk in audio_chunks}
    # return jsonify({"status": "processing started"}), 200
    try:
        return stream_with_context(generate_chunks(audio_chunks, preprocessed_data, args))
        # base64_video, temp_file_path, duration = process_chunk(driven_audio_path, preprocessed_data, args)
    except Exception as e:
        return jsonify({'status': 'error', 'message': str(e)}), 500

# @app.route("/stream", methods=["GET"])
# def stream_results():
#     global future_to_chunk
#     def generate():
#         for future in as_completed(future_to_chunk):
#             idx = future_to_chunk[future]
#             try:
#                 base64_video, temp_file_path = future.result()
#                 yield json.dumps({'start_time': idx, 'path': temp_file_path}).encode('utf-8')
#             except Exception as e:
#                 yield json.dumps({'start_time': idx, 'error': str(e)}).encode('utf-8')

#     return stream_with_context(generate())

@app.route("/health", methods=["GET"])
def health_status():
    response = {"online": "true"}
    return jsonify(response)

if __name__ == '__main__':
    app.run(debug=True)