File size: 12,008 Bytes
22257c4
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
075b64e
22257c4
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
import os
import numpy as np
import pickle
import torch
from math import ceil
from src.autovc.retrain_version.model_vc_37_1 import Generator
from pydub import AudioSegment
import pynormalize.pynormalize
from scipy.io import  wavfile as wav
from scipy.signal import stft


def match_target_amplitude(sound, target_dBFS):
    change_in_dBFS = target_dBFS - sound.dBFS
    return sound.apply_gain(change_in_dBFS)

class AutoVC_mel_Convertor():

    def __init__(self, src_dir, proportion=(0., 1.), seed=0):
        self.src_dir = src_dir
        if(not os.path.exists(os.path.join(src_dir, 'filename_index.txt'))):
            self.filenames = []
        else:
            with open(os.path.join(src_dir, 'filename_index.txt'), 'r') as f:
                lines = f.readlines()
                self.filenames = [(int(line.split(' ')[0]), line.split(' ')[1][:-1]) for line in lines]

        np.random.seed(seed)
        rand_perm = np.random.permutation(len(self.filenames))
        proportion_idx = (int(proportion[0] * len(rand_perm)), int(proportion[1] * len(rand_perm)))
        selected_index = rand_perm[proportion_idx[0] : proportion_idx[1]]
        self.selected_filenames = [self.filenames[i] for i in selected_index]

        print('{} out of {} are in this portion'.format(len(self.selected_filenames), len(self.filenames)))

    def __convert_single_only_au_AutoVC_format_to_dataset__(self, filename, build_train_dataset=True):
        """
        Convert a single file (only audio in AutoVC embedding format) to numpy arrays
        :param filename:
        :param is_map_to_std_face:
        :return:
        """

        global_clip_index, video_name = filename

        # audio_file = os.path.join(self.src_dir, 'raw_wav', '{}.wav'.
        #                           format(video_name[:-4]))
        audio_file = os.path.join(self.src_dir, 'raw_wav', '{:05d}_{}_audio.wav'.
                                  format(global_clip_index, video_name[:-4]))
        if(not build_train_dataset):
            import shutil
            audio_file = os.path.join(self.src_dir, 'raw_wav', '{:05d}_{}_audio.wav'.
                                      format(global_clip_index, video_name[:-4]))
            shutil.copy(os.path.join(self.src_dir, 'test_wav_files', video_name), audio_file)

        sound = AudioSegment.from_file(audio_file, "wav")
        normalized_sound = match_target_amplitude(sound, -20.0)
        normalized_sound.export(audio_file, format='wav')


        from src.autovc.retrain_version.vocoder_spec.extract_f0_func import extract_f0_func_audiofile
        S, f0_norm = extract_f0_func_audiofile(audio_file, 'M')

        from src.autovc.utils import quantize_f0_interp
        f0_onehot = quantize_f0_interp(f0_norm)

        from thirdparty.resemblyer_util.speaker_emb import get_spk_emb
        mean_emb, _ = get_spk_emb(audio_file)


        return S, mean_emb, f0_onehot

    def convert_wav_to_autovc_input(self, build_train_dataset=True, autovc_model_path=r'E:\Dataset\VCTK\stargan_vc\train_85_withpre1125000_local\360000-G.ckpt'):


        def pad_seq(x, base=32):
            len_out = int(base * ceil(float(x.shape[0]) / base))
            len_pad = len_out - x.shape[0]
            assert len_pad >= 0
            return np.pad(x, ((0, len_pad), (0, 0)), 'constant'), len_pad

        device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
        print(device)
        G = Generator(16, 256, 512, 16).eval().to(device)
        g_checkpoint = torch.load(autovc_model_path, map_location=device)
        G.load_state_dict(g_checkpoint['model'])

        emb = np.loadtxt('autovc/retrain_version/obama_emb.txt')
        emb_trg = torch.from_numpy(emb[np.newaxis, :].astype('float32')).to(device)

        aus = []

        for i, file in enumerate(self.selected_filenames):
            print(i, file)
            x_real_src, emb, f0_org_src = self.__convert_single_only_au_AutoVC_format_to_dataset__(filename=file, build_train_dataset=build_train_dataset)

            '''# normal length #'''
            # with torch.no_grad():
            #     x_identic, x_identic_psnt, code_real = G(x_real, emb_org, f0_org, emb_trg, f0_org)
            #     g_loss_id_psnt = F.mse_loss(x_real, x_identic_psnt, reduction='sum')
            #     print('loss:', g_loss_id_psnt / x_identic_psnt.shape[1] * 128)

            ''' too long split length '''
            l = x_real_src.shape[0]
            x_identic_psnt = []
            step = 4096
            for i in range(0, l, step):
                x_real = x_real_src[i:i+step]
                f0_org = f0_org_src[i:i+step]

                x_real, len_pad = pad_seq(x_real.astype('float32'))
                f0_org, _ = pad_seq(f0_org.astype('float32'))
                x_real = torch.from_numpy(x_real[np.newaxis, :].astype('float32')).to(device)
                emb_org = torch.from_numpy(emb[np.newaxis, :].astype('float32')).to(device)
                # emb_trg = torch.from_numpy(emb[np.newaxis, :].astype('float32')).to(device)
                f0_org = torch.from_numpy(f0_org[np.newaxis, :].astype('float32')).to(device)

                print('source shape:', x_real.shape, emb_org.shape, emb_trg.shape, f0_org.shape)

                with torch.no_grad():
                    x_identic, x_identic_psnt_i, code_real = G(x_real, emb_org, f0_org, emb_trg, f0_org)
                    x_identic_psnt.append(x_identic_psnt_i)

            x_identic_psnt = torch.cat(x_identic_psnt, dim=1)
            print('converted shape:', x_identic_psnt.shape, code_real.shape)
            if len_pad == 0:
                uttr_trg = x_identic_psnt[0, :, :].cpu().numpy()
            else:
                uttr_trg = x_identic_psnt[0, :-len_pad, :].cpu().numpy()

            # ''' plot source and converted mel-spec figures '''
            # import matplotlib.pyplot as plt
            # plt.subplot(1, 2, 1)
            # plt.imshow(x_real_src[0:200, :])
            # plt.subplot(1, 2, 2)
            # plt.imshow(uttr_trg[0:200, :])
            # plt.show()
            #
            # exit(0)

            file = (file[0], file[1], emb)
            aus.append((uttr_trg, file))

        return aus

    def convert_single_wav_to_input(self, audio_filename):
        aus = []
        audio_file = os.path.join(self.src_dir, 'demo_wav', audio_filename)

        # Default param
        TARGET_AUDIO_DBFS = -20.0
        WAV_STEP = int(0.2 * 16000)  # 0.2s = 5 frames
        STFT_WINDOW_SIZE = {'25': 320, '29.97': 356}
        STFT_WINDOW_STEP = {'25': 4, '29.97': 3}
        FPS = 25

        # Step 1 : Normalize the volume
        target_dbfs = TARGET_AUDIO_DBFS
        pynormalize.process_files(
            Files=[audio_file],
            target_dbfs=target_dbfs,
            directory=os.path.join(self.src_dir, 'raw_wav')
        )

        #  Step 2 : load wav file
        sample_rate, samples = wav.read(audio_file)
        assert (sample_rate == 16000)
        if (len(samples.shape) > 1):
            samples = samples[:, 0]  # pick mono

        # Step 3 : STFT,
        # 1 frame = 1/25 * 16k = 640 samples => windowsize=320,  overlap=160
        # 1 frame = 1/29.97 * 16k = 533.86 samples => windowsize=356, overlap=178, (mis-align = 4.2sample / 1s)
        f, t, Zxx = stft(samples, fs=sample_rate, nperseg=STFT_WINDOW_SIZE[str(FPS)])

        # stft_abs = np.abs(Zxx)
        stft_abs = np.log(np.abs(Zxx) ** 2 + 1e-10)
        stft_abs_max = np.max(stft_abs)
        stft_abs /= stft_abs_max

        # Step 4 : align AV (drop last 2 frames of V)
        fl_length = stft_abs.shape[1] // STFT_WINDOW_STEP[str(FPS)]
        audio_stft_length = (fl_length - 2) * STFT_WINDOW_STEP[str(FPS)]
        stft_signal = Zxx[:, 0:audio_stft_length]
        stft_abs = stft_abs[:, 0:audio_stft_length]

        audio_wav_length = int((fl_length - 2) * sample_rate / FPS)
        wav_signal = samples[0:audio_wav_length]

        # # Step 6 : Save audio
        # info_audio = (0, stft_signal, fl_length - 2, audio_stft_length, audio_wav_length)
        # au_data = (stft_abs, wav_signal, info_audio)

        aus.append((stft_abs.T, None, (0, audio_filename, 0)))

        return aus


    def convert_single_wav_to_autovc_input(self, audio_filename, autovc_model_path):


        def pad_seq(x, base=32):
            len_out = int(base * ceil(float(x.shape[0]) / base))
            len_pad = len_out - x.shape[0]
            assert len_pad >= 0
            return np.pad(x, ((0, len_pad), (0, 0)), 'constant'), len_pad

        device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
        G = Generator(16, 256, 512, 16).eval().to(device)

        g_checkpoint = torch.load(autovc_model_path, map_location=device)
        G.load_state_dict(g_checkpoint['model'])

        emb = np.loadtxt('MakeItTalk/src/autovc/retrain_version/obama_emb.txt')
        emb_trg = torch.from_numpy(emb[np.newaxis, :].astype('float32')).to(device)

        aus = []
        audio_file = audio_filename

        sound = AudioSegment.from_file(audio_file, "wav")
        normalized_sound = match_target_amplitude(sound, -20.0)
        normalized_sound.export(audio_file, format='wav')

        from src.autovc.retrain_version.vocoder_spec.extract_f0_func import extract_f0_func_audiofile
        x_real_src, f0_norm = extract_f0_func_audiofile(audio_file, 'F')
        from src.autovc.utils import quantize_f0_interp
        f0_org_src = quantize_f0_interp(f0_norm)
        from thirdparty.resemblyer_util.speaker_emb import get_spk_emb
        emb, _ = get_spk_emb(audio_file)

        ''' normal length version '''
        # x_real, len_pad = pad_seq(x_real_src.astype('float32'))
        # f0_org, _ = pad_seq(f0_org_src.astype('float32'))
        # x_real = torch.from_numpy(x_real[np.newaxis, :].astype('float32')).to(device)
        # emb_org = torch.from_numpy(emb[np.newaxis, :].astype('float32')).to(device)
        # f0_org = torch.from_numpy(f0_org[np.newaxis, :].astype('float32')).to(device)
        # print('source shape:', x_real.shape, emb_org.shape, emb_trg.shape, f0_org.shape)
        #
        # with torch.no_grad():
        #     x_identic, x_identic_psnt, code_real = G(x_real, emb_org, f0_org, emb_trg, f0_org)
        # print('converted shape:', x_identic_psnt.shape, code_real.shape)

        ''' long split version '''
        l = x_real_src.shape[0]
        x_identic_psnt = []
        step = 4096
        for i in range(0, l, step):
            x_real = x_real_src[i:i + step]
            f0_org = f0_org_src[i:i + step]

            x_real, len_pad = pad_seq(x_real.astype('float32'))
            f0_org, _ = pad_seq(f0_org.astype('float32'))
            x_real = torch.from_numpy(x_real[np.newaxis, :].astype('float32')).to(device)
            emb_org = torch.from_numpy(emb[np.newaxis, :].astype('float32')).to(device)
            # emb_trg = torch.from_numpy(emb[np.newaxis, :].astype('float32')).to(device)
            f0_org = torch.from_numpy(f0_org[np.newaxis, :].astype('float32')).to(device)
            print('source shape:', x_real.shape, emb_org.shape, emb_trg.shape, f0_org.shape)

            with torch.no_grad():
                x_identic, x_identic_psnt_i, code_real = G(x_real, emb_org, f0_org, emb_trg, f0_org)
                x_identic_psnt.append(x_identic_psnt_i)

        x_identic_psnt = torch.cat(x_identic_psnt, dim=1)
        print('converted shape:', x_identic_psnt.shape, code_real.shape)

        if len_pad == 0:
            uttr_trg = x_identic_psnt[0, :, :].cpu().numpy()
        else:
            uttr_trg = x_identic_psnt[0, :-len_pad, :].cpu().numpy()

        aus.append((uttr_trg, (0, audio_filename, emb)))

        return aus



if __name__ == '__main__':
    c = AutoVC_mel_Convertor(r'E:\Dataset\TalkingToon\Obama_for_train', proportion=(0.0, 1.0))
    aus = c.convert_wav_to_autovc_input()

    with open(os.path.join(r'E:\Dataset\TalkingToon\Obama_for_train', 'dump', 'autovc_retrain_mel_au.pickle'), 'wb') as fp:
        pickle.dump(aus, fp)