app.py

import warnings
warnings.filterwarnings("ignore") # 경고 무시
#!pip install pyannote.audio
#!pip install moviepy
import librosa
import numpy as np
import os

import torch
import torch.nn as nn
import torch.optim as optim
from torch.utils.data import DataLoader, TensorDataset
import torch.functional as F

from pyannote.audio import Pipeline

from pyannote.audio import Audio
import torchaudio


import torch.nn.functional as F
import os

from moviepy.editor import VideoFileClip

from transformers import pipeline

from huggingface_hub import hf_hub_download

#!pip install gradio
import gradio as gr

from moviepy.editor import VideoFileClip


# 오디오 변환 mp4 --> wav
def extract_audio_from_video(video_file_path, audio_file_path):
    # mp4 파일 불러오기
    video = VideoFileClip(video_file_path)

    # 오디오를 추출하여 wav 파일로 저장
    video.audio.write_audiofile(audio_file_path, codec='pcm_s16le')

# 전체 오디오 파일 불러오기
def seprate_speaker(audio_file, pipeline):
    audio = Audio()
    waveform, sample_rate = torchaudio.load(audio_file)
    diarization = pipeline(audio_file)

    # 화자별로 발화 구간을 저장할 딕셔너리 초기화
    speaker_segments = {}

    # diarization 결과를 순회하며 각 화자의 발화를 딕셔너리에 추가
    for segment, _, speaker in diarization.itertracks(yield_label=True):
        start_time = segment.start
        end_time = segment.end

        # 해당 화자가 처음 등장하면 리스트를 초기화
        if speaker not in speaker_segments:
            speaker_segments[speaker] = []

        # 발화 구간을 해당 화자의 리스트에 추가
        segment_waveform = waveform[:, int(start_time * sample_rate):int(end_time * sample_rate)]
        speaker_segments[speaker].append(segment_waveform)

    # 각 화자별로 모든 발화 구간을 하나의 파일로 이어붙여 저장
    for speaker, segments in speaker_segments.items():
        # 화자의 모든 발화 구간을 이어붙임
        combined_waveform = torch.cat(segments, dim=1)
        # output_path = "/content/wav"    # 경로
        output_path = './output'
        os.makedirs(output_path, exist_ok=True) # 경로가 없으면 생성
        output_filename = os.path.join(output_path,f"{speaker}.wav")

        torchaudio.save(output_filename, combined_waveform, sample_rate) #오디오 파일 저장
        #print(f"Saved {output_filename} for speaker {speaker}")


# 간단한 DeepVoice 스타일 모델 정의
class DeepVoiceModel(nn.Module):
    def __init__(self, input_dim, hidden_dim, num_classes, dropout_rate=0.3, l2_reg=0.01):
        super(DeepVoiceModel, self).__init__()
        self.conv1 = nn.Conv1d(input_dim, hidden_dim, kernel_size=5, padding=2)
        self.bn1 = nn.BatchNorm1d(hidden_dim)
        self.conv2 = nn.Conv1d(hidden_dim, hidden_dim, kernel_size=5, padding=2)
        self.bn2 = nn.BatchNorm1d(hidden_dim)
        self.dropout = nn.Dropout(dropout_rate)
        self.fc = nn.Linear(hidden_dim, num_classes)

    def forward(self, x):
        x = self.bn1(torch.relu(self.conv1(x)))
        x = self.dropout(x)
        x = self.bn2(torch.relu(self.conv2(x)))
        x = self.dropout(x)
        x = torch.mean(x, dim=2)  # Temporal pooling
        x = self.fc(x)
        return x


def extract_mfcc_path(file_path, n_mfcc=13, max_len=100):
    # 음성 파일
    audio, sample_rate = librosa.load(file_path, sr=None)
    # mfcc 특성 추출
    mfcc = librosa.feature.mfcc(y=audio, sr=sample_rate, n_mfcc=n_mfcc)

    # 일정한 길이로 맞춤
    if mfcc.shape[1] < max_len:
        pad_width = max_len - mfcc.shape[1]
        mfcc = np.pad(mfcc, ((0, 0), (0, pad_width)), mode='constant')
    else:
        mfcc = mfcc[:, :max_len]

    return torch.Tensor(mfcc)

# 폴더에 있는 데이터 한번에 접근해서 한번에 체크
def real_fake_check(list_dir, path, model):
    THRESHOLD = 0.4     #딥페이크 기준을 0.4로 설정
    r_cnt = 0
    f_cnt = 0
    prob = {}
    for i in list_dir:      # real / fake 선택
        #print('------',i)
        input_data = extract_mfcc_path(os.path.join(path, i))
        input_data = torch.tensor(input_data).unsqueeze(0).to(device)  # 배치 차원을 추가하여 (1, input_dim, sequence_length)로 맞춤
        result = model(input_data.float())
        # predicted_class = torch.argmax(result, dim=1).item()
        probabilities = F.softmax(result, dim=1)
        prob[i]='%.2f'%probabilities[0][1].item()

        predicted_class = 0 if probabilities[0][0] >= THRESHOLD else 1  # 확률값이 기준치보다 크다면 real, 아니면 fake
        # print('-- %.2f'%probabilities[0][0].item()) #확률 값 출력
        if predicted_class == 0:
            # print("REAL")
            r_cnt += 1
        else:
            # print("FAKE")
            f_cnt += 1
    #print()
    #print('real: ',r_cnt,'/',len(list_dir))
    #print('fake: ',f_cnt,'/',len(list_dir))
    return {'real: ':f'{r_cnt}/{len(list_dir)}', 'fake: ':f'{f_cnt}/{len(list_dir)}', 'prob: ': prob}


def main(file_name):
    my_key = os.getenv("my_key")
    pipeline = Pipeline.from_pretrained("pyannote/speaker-diarization-3.1",
                                        use_auth_token=my_key)
    # pipeline = Pipeline.from_pretrained("pyannote/speaker-diarization-3.1")
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

    video_file = file_name #deepfake #meganfox.mp4'
    audio_file = './output_audio.wav'  # 저장할 오디오 파일의 경로, 이름 지정

    extract_audio_from_video(video_file, audio_file)

    seprate_speaker(audio_file,pipeline) # 발화자 분리해서 파일로 만들기

    mel_dim = 13 # Mel-spectrogram 차원
    num_classes = 2  # 분류할 클래스 수
    input_dim = mel_dim
    hidden_dim = 128
    dropout_rate = 0.2
    l2_reg = 0.01

    # 모델
    model_name = './deepvoice_model_girl.pth'
    model = DeepVoiceModel(input_dim, hidden_dim, num_classes, dropout_rate, l2_reg).to(device)
    model.load_state_dict(torch.load(model_name, map_location=torch.device(device)))#("/content/drive/MyDrive/캡스톤 1조/model/deepvoice_model_girl.pth"))
    model.eval()  # 평가 모드로 설정


    #real,fake 폴더
    #real_path = '/content/drive/MyDrive/캡스톤 1조/data/deepvoice/real'
    #real_path = '/content/drive/MyDrive/Celeb-DF-v2/Celeb-real'

    #real = os.listdir(real_path)
    fake_path = './output'#'/content/drive/MyDrive/캡스톤 1조/data/deepvoice/fake'
    fake = os.listdir(fake_path)

    #print("\n-------real data---------")
    #real_fake_check(real, real_path, model) #real dataset
    #print("\n-------fake data---------")
    rf_check = real_fake_check(fake, fake_path,model)       #fake dataset\
    return rf_check

#Gradio 메인 함수
def deepvoice_check(video_file):
    results = main(video_file)
    return results

# Gradio 인터페이스 생성
iface = gr.Interface(
    fn=main,
    inputs=gr.Video(label="Upload mp4 File"),
    outputs=gr.Textbox(label="Deepfake Detection Result"),
    title="DeepVoice Check",
    description="Upload an mp4 file to check for DeepVoice indicators."
)

# Gradio 인터페이스 실행
iface.launch(share=True, debug=True)