diff --git a/GPT_SoVITS/module/data_utils.py b/GPT_SoVITS/module/data_utils.py index bfa29e8..323bf1b 100644 --- a/GPT_SoVITS/module/data_utils.py +++ b/GPT_SoVITS/module/data_utils.py @@ -9,7 +9,7 @@ import torch.utils.data from tqdm import tqdm from module import commons -from module.mel_processing import spectrogram_torch +from module.mel_processing import spectrogram_torch,spec_to_mel_torch from text import cleaned_text_to_sequence from utils import load_wav_to_torch, load_filepaths_and_text import torch.nn.functional as F @@ -170,8 +170,6 @@ class TextAudioSpeakerLoader(torch.utils.data.Dataset): assert abs(ssl.shape[-1] - wav2.shape[-1] // self.hop_length) < 3, ( ssl.shape, wav.shape, wav2.shape, mel.shape, sep_point, self.hop_length, sep_point * self.hop_length, dir) return reference_mel, ssl, wav2, mel - - class TextAudioSpeakerCollate(): """ Zero-pads model inputs and targets """ @@ -232,7 +230,232 @@ class TextAudioSpeakerCollate(): text_lengths[i] = text.size(0) return ssl_padded, ssl_lengths, spec_padded, spec_lengths, wav_padded, wav_lengths, text_padded, text_lengths +class TextAudioSpeakerLoaderV3(torch.utils.data.Dataset): + """ + 1) loads audio, speaker_id, text pairs + 2) normalizes text and converts them to sequences of integers + 3) computes spectrograms from audio files. + """ + def __init__(self, hparams, val=False): + exp_dir = hparams.exp_dir + self.path2 = "%s/2-name2text.txt" % exp_dir + self.path4 = "%s/4-cnhubert" % exp_dir + self.path5 = "%s/5-wav32k" % exp_dir + assert os.path.exists(self.path2) + assert os.path.exists(self.path4) + assert os.path.exists(self.path5) + names4 = set([name[:-3] for name in list(os.listdir(self.path4))]) # 去除.pt后缀 + names5 = set(os.listdir(self.path5)) + self.phoneme_data = {} + with open(self.path2, "r", encoding="utf8") as f: + lines = f.read().strip("\n").split("\n") + + for line in lines: + tmp = line.split("\t") + if (len(tmp) != 4): + continue + self.phoneme_data[tmp[0]] = [tmp[1]] + + self.audiopaths_sid_text = list(set(self.phoneme_data) & names4 & names5) + tmp = self.audiopaths_sid_text + leng = len(tmp) + min_num = 100 + if (leng < min_num): + self.audiopaths_sid_text = [] + for _ in range(max(2, int(min_num / leng))): + self.audiopaths_sid_text += tmp + self.max_wav_value = hparams.max_wav_value + self.sampling_rate = hparams.sampling_rate + self.filter_length = hparams.filter_length + self.hop_length = hparams.hop_length + self.win_length = hparams.win_length + self.sampling_rate = hparams.sampling_rate + self.val = val + + random.seed(1234) + random.shuffle(self.audiopaths_sid_text) + + print("phoneme_data_len:", len(self.phoneme_data.keys())) + print("wav_data_len:", len(self.audiopaths_sid_text)) + + audiopaths_sid_text_new = [] + lengths = [] + skipped_phone = 0 + skipped_dur = 0 + for audiopath in tqdm(self.audiopaths_sid_text): + try: + phoneme = self.phoneme_data[audiopath][0] + phoneme = phoneme.split(' ') + phoneme_ids = cleaned_text_to_sequence(phoneme, version) + except Exception: + print(f"{audiopath} not in self.phoneme_data !") + skipped_phone += 1 + continue + + size = os.path.getsize("%s/%s" % (self.path5, audiopath)) + duration = size / self.sampling_rate / 2 + + if duration == 0: + print(f"Zero duration for {audiopath}, skipping...") + skipped_dur += 1 + continue + + if 54 > duration > 0.6 or self.val: + audiopaths_sid_text_new.append([audiopath, phoneme_ids]) + lengths.append(size // (2 * self.hop_length)) + else: + skipped_dur += 1 + continue + + print("skipped_phone: ", skipped_phone, ", skipped_dur: ", skipped_dur) + print("total left: ", len(audiopaths_sid_text_new)) + assert len(audiopaths_sid_text_new) > 1 # 至少能凑够batch size，这里todo + self.audiopaths_sid_text = audiopaths_sid_text_new + self.lengths = lengths + self.spec_min=-12 + self.spec_max=2 + + self.filter_length_mel=self.win_length_mel=1024 + self.hop_length_mel=256 + self.n_mel_channels=100 + self.sampling_rate_mel=24000 + self.mel_fmin=0 + self.mel_fmax=None + def norm_spec(self, x): + return (x - self.spec_min) / (self.spec_max - self.spec_min) * 2 - 1 + + def get_audio_text_speaker_pair(self, audiopath_sid_text): + audiopath, phoneme_ids = audiopath_sid_text + text = torch.FloatTensor(phoneme_ids) + try: + spec, mel = self.get_audio("%s/%s" % (self.path5, audiopath)) + with torch.no_grad(): + ssl = torch.load("%s/%s.pt" % (self.path4, audiopath), map_location="cpu") + if (ssl.shape[-1] != spec.shape[-1]): + typee = ssl.dtype + ssl = F.pad(ssl.float(), (0, 1), mode="replicate").to(typee) + ssl.requires_grad = False + except: + traceback.print_exc() + mel = torch.zeros(100, 180) + # wav = torch.zeros(1, 96 * self.hop_length) + spec = torch.zeros(1025, 96) + ssl = torch.zeros(1, 768, 96) + text = text[-1:] + print("load audio or ssl error!!!!!!", audiopath) + return (ssl, spec, mel, text) + + def get_audio(self, filename): + audio_array = load_audio(filename,self.sampling_rate)#load_audio的方法是已经归一化到-1~1之间的，不用再/32768 + audio=torch.FloatTensor(audio_array)#/32768 + audio_norm = audio + audio_norm = audio_norm.unsqueeze(0) + audio_array24 = load_audio(filename,24000)#load_audio的方法是已经归一化到-1~1之间的，不用再/32768######这里可以用GPU重采样加速 + audio24=torch.FloatTensor(audio_array24)#/32768 + audio_norm24 = audio24 + audio_norm24 = audio_norm24.unsqueeze(0) + + spec = spectrogram_torch(audio_norm, self.filter_length, + self.sampling_rate, self.hop_length, self.win_length, + center=False) + spec = torch.squeeze(spec, 0) + + + spec1 = spectrogram_torch(audio_norm24, self.filter_length_mel,self.sampling_rate_mel, self.hop_length_mel, self.win_length_mel,center=False) + mel = spec_to_mel_torch(spec1, self.filter_length_mel, self.n_mel_channels, self.sampling_rate_mel, self.mel_fmin, self.mel_fmax) + mel = torch.squeeze(mel, 0) + mel=self.norm_spec(mel) + # print(1111111,spec.shape,mel.shape) + return spec, mel + + def get_sid(self, sid): + sid = torch.LongTensor([int(sid)]) + return sid + + def __getitem__(self, index): + # with torch.no_grad(): + return self.get_audio_text_speaker_pair(self.audiopaths_sid_text[index]) + + def __len__(self): + return len(self.audiopaths_sid_text) +class TextAudioSpeakerCollateV3(): + """ Zero-pads model inputs and targets + """ + + def __init__(self, return_ids=False): + self.return_ids = return_ids + + def __call__(self, batch): + """Collate's training batch from normalized text, audio and speaker identities + PARAMS + ------ + batch: [text_normalized, spec_normalized, wav_normalized, sid] + """ + #ssl, spec, wav,mel, text + # Right zero-pad all one-hot text sequences to max input length + _, ids_sorted_decreasing = torch.sort( + torch.LongTensor([x[1].size(1) for x in batch]), + dim=0, descending=True) +#(ssl, spec,mel, text) + max_ssl_len = max([x[0].size(2) for x in batch]) + + max_ssl_len1 = int(8 * ((max_ssl_len // 8) + 1)) + max_ssl_len = int(2 * ((max_ssl_len // 2) + 1)) + + # max_ssl_len = int(8 * ((max_ssl_len // 8) + 1)) + # max_ssl_len1=max_ssl_len + + max_spec_len = max([x[1].size(1) for x in batch]) + max_spec_len = int(2 * ((max_spec_len // 2) + 1)) + # max_wav_len = max([x[2].size(1) for x in batch]) + + max_text_len = max([x[3].size(0) for x in batch]) + max_mel_len=int(max_ssl_len1*1.25*1.5)###24000/256,32000/640=16000/320 + + ssl_lengths = torch.LongTensor(len(batch)) + spec_lengths = torch.LongTensor(len(batch)) + text_lengths = torch.LongTensor(len(batch)) + # wav_lengths = torch.LongTensor(len(batch)) + mel_lengths = torch.LongTensor(len(batch)) + + spec_padded = torch.FloatTensor(len(batch), batch[0][1].size(0), max_spec_len) + mel_padded = torch.FloatTensor(len(batch), batch[0][2].size(0), max_mel_len) + ssl_padded = torch.FloatTensor(len(batch), batch[0][0].size(1), max_ssl_len) + text_padded = torch.LongTensor(len(batch), max_text_len) + # wav_padded = torch.FloatTensor(len(batch), 1, max_wav_len) + + spec_padded.zero_() + mel_padded.zero_() + ssl_padded.zero_() + text_padded.zero_() + # wav_padded.zero_() + + for i in range(len(ids_sorted_decreasing)): + row = batch[ids_sorted_decreasing[i]] + # ssl, spec, wav,mel, text + ssl = row[0] + ssl_padded[i, :, :ssl.size(2)] = ssl[0, :, :] + ssl_lengths[i] = ssl.size(2) + + spec = row[1] + spec_padded[i, :, :spec.size(1)] = spec + spec_lengths[i] = spec.size(1) + + # wav = row[2] + # wav_padded[i, :, :wav.size(1)] = wav + # wav_lengths[i] = wav.size(1) + + mel = row[2] + mel_padded[i, :, :mel.size(1)] = mel + mel_lengths[i] = mel.size(1) + + text = row[3] + text_padded[i, :text.size(0)] = text + text_lengths[i] = text.size(0) + + # return ssl_padded, spec_padded,mel_padded, ssl_lengths, spec_lengths, text_padded, text_lengths, wav_padded, wav_lengths,mel_lengths + return ssl_padded, spec_padded,mel_padded, ssl_lengths, spec_lengths, text_padded, text_lengths,mel_lengths class DistributedBucketSampler(torch.utils.data.distributed.DistributedSampler): """ diff --git a/GPT_SoVITS/module/models.py b/GPT_SoVITS/module/models.py index 968c4cb..b6a6d09 100644 --- a/GPT_SoVITS/module/models.py +++ b/GPT_SoVITS/module/models.py @@ -12,7 +12,7 @@ from torch.nn import functional as F from module import commons from module import modules from module import attentions - +from f5_tts.model import DiT from torch.nn import Conv1d, ConvTranspose1d, AvgPool1d, Conv2d from torch.nn.utils import weight_norm, remove_weight_norm, spectral_norm from module.commons import init_weights, get_padding @@ -22,7 +22,7 @@ from module.quantize import ResidualVectorQuantizer from text import symbols as symbols_v1 from text import symbols2 as symbols_v2 from torch.cuda.amp import autocast -import contextlib +import contextlib,random class StochasticDurationPredictor(nn.Module): @@ -371,6 +371,37 @@ class PosteriorEncoder(nn.Module): return z, m, logs, x_mask +class Encoder(nn.Module): + def __init__(self, + in_channels, + out_channels, + hidden_channels, + kernel_size, + dilation_rate, + n_layers, + gin_channels=0): + super().__init__() + self.in_channels = in_channels + self.out_channels = out_channels + self.hidden_channels = hidden_channels + self.kernel_size = kernel_size + self.dilation_rate = dilation_rate + self.n_layers = n_layers + self.gin_channels = gin_channels + + self.pre = nn.Conv1d(in_channels, hidden_channels, 1) + self.enc = modules.WN(hidden_channels, kernel_size, dilation_rate, n_layers, gin_channels=gin_channels) + self.proj = nn.Conv1d(hidden_channels, out_channels, 1) + + def forward(self, x, x_lengths, g=None): + if(g!=None): + g = g.detach() + x_mask = torch.unsqueeze(commons.sequence_mask(x_lengths, x.size(2)), 1).to(x.dtype) + x = self.pre(x) * x_mask + x = self.enc(x, x_mask, g=g) + stats = self.proj(x) * x_mask + return stats, x_mask + class WNEncoder(nn.Module): def __init__( self, @@ -1028,3 +1059,218 @@ class SynthesizerTrn(nn.Module): ssl = self.ssl_proj(x) quantized, codes, commit_loss, quantized_list = self.quantizer(ssl) return codes.transpose(0, 1) +class CFM(torch.nn.Module): + def __init__( + self, + in_channels,dit + ): + super().__init__() + self.sigma_min = 1e-6 + + self.estimator = dit + + self.in_channels = in_channels + + self.criterion = torch.nn.MSELoss() + + @torch.inference_mode() + def inference(self, mu, x_lens, prompt, n_timesteps, temperature=1.0, inference_cfg_rate=0): + """Forward diffusion""" + B, T = mu.size(0), mu.size(1) + x = torch.randn([B, self.in_channels, T], device=mu.device,dtype=mu.dtype) * temperature + prompt_len = prompt.size(-1) + prompt_x = torch.zeros_like(x,dtype=mu.dtype) + prompt_x[..., :prompt_len] = prompt[..., :prompt_len] + x[..., :prompt_len] = 0 + mu=mu.transpose(2,1) + t = 0 + d = 1 / n_timesteps + for j in range(n_timesteps): + t_tensor = torch.ones(x.shape[0], device=x.device,dtype=mu.dtype) * t + d_tensor = torch.ones(x.shape[0], device=x.device,dtype=mu.dtype) * d + # v_pred = model(x, t_tensor, d_tensor, **extra_args) + v_pred = self.estimator(x, prompt_x, x_lens, t_tensor,d_tensor, mu,drop_audio_cond=False,drop_text=False).transpose(2, 1) + if inference_cfg_rate>1e-5: + neg = self.estimator(x, prompt_x, x_lens, t_tensor, d_tensor, mu, drop_audio_cond=True, drop_text=True).transpose(2, 1) + v_pred=v_pred+(v_pred-neg)*inference_cfg_rate + x = x + d * v_pred + t = t + d + x[:, :, :prompt_len] = 0 + return x + def forward(self, x1, x_lens, prompt_lens, mu): + b, _, t = x1.shape + + # random timestep + t = torch.rand([b], device=mu.device, dtype=x1.dtype) + x0 = torch.randn_like(x1,device=mu.device) + vt = x1 - x0 + xt = x0 + t[:, None, None] * vt + dt = torch.zeros_like(t,device=mu.device) + prompt = torch.zeros_like(x1) + for bib in range(b): + prompt[bib, :, :prompt_lens[bib]] = x1[bib, :, :prompt_lens[bib]] + xt[bib, :, :prompt_lens[bib]] = 0 + gailv=0.3# if ttime()>1736250488 else 0.1 + if random.random() < gailv: + base = torch.randint(2, 8, (t.shape[0],), device=mu.device) + d = 1/torch.pow(2, base) + d_input = d.clone() + d_input[d_input < 1e-2] = 0 + # with torch.no_grad(): + v_pred_1 = self.estimator(xt, prompt, x_lens, t, d_input, mu).transpose(2, 1).detach() + # v_pred_1 = self.diffusion(xt, t, d_input, cond=conditioning).detach() + x_mid = xt + d[:, None, None] * v_pred_1 + # v_pred_2 = self.diffusion(x_mid, t+d, d_input, cond=conditioning).detach() + v_pred_2 = self.estimator(x_mid, prompt, x_lens, t+d, d_input, mu).transpose(2, 1).detach() + vt = (v_pred_1 + v_pred_2) / 2 + vt = vt.detach() + dt = 2*d + + vt_pred = self.estimator(xt, prompt, x_lens, t,dt, mu).transpose(2,1) + loss = 0 + + # print(45555555,estimator_out.shape,u.shape,x_lens,prompt_lens)#45555555 torch.Size([7, 465, 100]) torch.Size([7, 100, 465]) tensor([461, 461, 451, 451, 442, 442, 442], device='cuda:0') tensor([ 96, 93, 185, 59, 244, 262, 294], device='cuda:0') + for bib in range(b): + loss += self.criterion(vt_pred[bib, :, prompt_lens[bib]:x_lens[bib]], vt[bib, :, prompt_lens[bib]:x_lens[bib]]) + loss /= b + + return loss#, estimator_out + (1 - self.sigma_min) * z + + +class SynthesizerTrnV3(nn.Module): + """ + Synthesizer for Training + """ + + def __init__(self, + spec_channels, + segment_size, + inter_channels, + hidden_channels, + filter_channels, + n_heads, + n_layers, + kernel_size, + p_dropout, + resblock, + resblock_kernel_sizes, + resblock_dilation_sizes, + upsample_rates, + upsample_initial_channel, + upsample_kernel_sizes, + n_speakers=0, + gin_channels=0, + use_sdp=True, + semantic_frame_rate=None, + freeze_quantizer=None, + **kwargs): + + super().__init__() + self.spec_channels = spec_channels + self.inter_channels = inter_channels + self.hidden_channels = hidden_channels + self.filter_channels = filter_channels + self.n_heads = n_heads + self.n_layers = n_layers + self.kernel_size = kernel_size + self.p_dropout = p_dropout + self.resblock = resblock + self.resblock_kernel_sizes = resblock_kernel_sizes + self.resblock_dilation_sizes = resblock_dilation_sizes + self.upsample_rates = upsample_rates + self.upsample_initial_channel = upsample_initial_channel + self.upsample_kernel_sizes = upsample_kernel_sizes + self.segment_size = segment_size + self.n_speakers = n_speakers + self.gin_channels = gin_channels + + self.model_dim=512 + self.use_sdp = use_sdp + self.enc_p = TextEncoder(inter_channels,hidden_channels,filter_channels,n_heads,n_layers,kernel_size,p_dropout) + # self.ref_enc = modules.MelStyleEncoder(spec_channels, style_vector_dim=gin_channels)###ع + self.ref_enc = modules.MelStyleEncoder(704, style_vector_dim=gin_channels)###ع + # self.dec = Generator(inter_channels, resblock, resblock_kernel_sizes, resblock_dilation_sizes, upsample_rates, + # upsample_initial_channel, upsample_kernel_sizes, gin_channels=gin_channels) + # self.enc_q = PosteriorEncoder(spec_channels, inter_channels, hidden_channels, 5, 1, 16, + # gin_channels=gin_channels) + # self.flow = ResidualCouplingBlock(inter_channels, hidden_channels, 5, 1, 4, gin_channels=gin_channels) + + + ssl_dim = 768 + assert semantic_frame_rate in ['25hz', "50hz"] + self.semantic_frame_rate = semantic_frame_rate + if semantic_frame_rate == '25hz': + self.ssl_proj = nn.Conv1d(ssl_dim, ssl_dim, 2, stride=2) + else: + self.ssl_proj = nn.Conv1d(ssl_dim, ssl_dim, 1, stride=1) + + self.quantizer = ResidualVectorQuantizer( + dimension=ssl_dim, + n_q=1, + bins=1024 + ) + self.freeze_quantizer=freeze_quantizer + + inter_channels2=512 + self.bridge=nn.Sequential( + nn.Conv1d(inter_channels, inter_channels2, 1, stride=1), + nn.LeakyReLU() + ) + self.wns1=Encoder(inter_channels2, inter_channels2, inter_channels2, 5, 1, 8,gin_channels=gin_channels) + self.linear_mel=nn.Conv1d(inter_channels2,100,1,stride=1) + self.cfm = CFM(100,DiT(**dict(dim=1024, depth=22, heads=16, ff_mult=2, text_dim=inter_channels2, conv_layers=4)),)#text_dim is condition feature dim + + def forward(self, ssl, y, mel,ssl_lengths,y_lengths, text, text_lengths,mel_lengths):#ssl_lengths no need now + with autocast(enabled=False): + y_mask = torch.unsqueeze(commons.sequence_mask(y_lengths, y.size(2)), 1).to(y.dtype) + ge = self.ref_enc(y[:,:704] * y_mask, y_mask) + maybe_no_grad = torch.no_grad() if self.freeze_quantizer else contextlib.nullcontext() + with maybe_no_grad: + if self.freeze_quantizer: + self.ssl_proj.eval()# + self.quantizer.eval() + ssl = self.ssl_proj(ssl) + quantized, codes, commit_loss, quantized_list = self.quantizer( + ssl, layers=[0] + ) + with maybe_no_grad: + quantized = F.interpolate(quantized, scale_factor=2, mode="nearest")##BCT + x, m_p, logs_p, y_mask = self.enc_p(quantized, y_lengths, text, text_lengths, ge) + fea=self.bridge(x) + fea = F.interpolate(fea, scale_factor=1.875, mode="nearest")##BCT + fea, y_mask_ = self.wns1(fea, mel_lengths, ge)###1min΢ûͲҪ΢ѧϰ + B=ssl.shape[0] + prompt_len_max = mel_lengths*2/3 + prompt_len = (torch.rand([B], device=fea.device) * prompt_len_max).floor().to(dtype=torch.long) + minn=min(mel.shape[-1],fea.shape[-1]) + mel=mel[:,:,:minn] + fea=fea[:,:,:minn] + cfm_loss= self.cfm(mel, mel_lengths, prompt_len, fea) + return cfm_loss + + @torch.no_grad() + def decode_encp(self, codes,text, refer,ge=None): + # print(2333333,refer.shape) + # ge=None + if(ge==None): + refer_lengths = torch.LongTensor([refer.size(2)]).to(refer.device) + refer_mask = torch.unsqueeze(commons.sequence_mask(refer_lengths, refer.size(2)), 1).to(refer.dtype) + ge = self.ref_enc(refer[:,:704] * refer_mask, refer_mask) + y_lengths = torch.LongTensor([int(codes.size(2)*2)]).to(codes.device) + y_lengths1 = torch.LongTensor([int(codes.size(2)*2.5*1.5)]).to(codes.device) + text_lengths = torch.LongTensor([text.size(-1)]).to(text.device) + + quantized = self.quantizer.decode(codes) + if self.semantic_frame_rate == '25hz': + quantized = F.interpolate(quantized, scale_factor=2, mode="nearest")##BCT + x, m_p, logs_p, y_mask = self.enc_p(quantized, y_lengths, text, text_lengths, ge) + fea=self.bridge(x) + fea = F.interpolate(fea, scale_factor=1.875, mode="nearest")##BCT + ####more wn paramter to learn mel + fea, y_mask_ = self.wns1(fea, y_lengths1, ge) + return fea,ge + + def extract_latent(self, x): + ssl = self.ssl_proj(x) + quantized, codes, commit_loss, quantized_list = self.quantizer(ssl) + return codes.transpose(0,1)