import torch import torch.nn as nn import torch.optim as optim import pytorch_lightning.core.module as pl import matplotlib.pyplot as plt from module import commons from transformer.embedding import SinePositionalEmbedding from transformer.transformer import LayerNorm, TransformerEncoder, TransformerEncoderLayer plt.switch_backend('agg') from module.models import SynthesizerTrn from module.modules import MelStyleEncoder def init_vits_s2(config_path, cpt_path): import utils hps = utils.get_hparams_from_file(config_path) vits_s2 = SynthesizerTrn( hps.data.filter_length // 2 + 1, hps.train.segment_size // hps.data.hop_length, n_speakers=hps.data.n_speakers, **hps.model) ckpt = torch.load(cpt_path, map_location='cpu')['model'] vits_s2.load_state_dict(ckpt) vits_s2.eval() return vits_s2 import math import torch.nn.functional as F from tqdm import tqdm class SinusoidalPosEmb(nn.Module): def __init__(self, dim): super().__init__() self.dim = dim def forward(self, x): device = x.device half_dim = self.dim // 2 emb = math.log(10000) / (half_dim - 1) emb = torch.exp(torch.arange(half_dim, device=device) * -emb) emb = x[:, None] * emb[None, :] emb = torch.cat((emb.sin(), emb.cos()), dim=-1) return emb class DiT(nn.Module): def __init__(self, io_channels, cond_channels): super().__init__() self.in_proj = nn.Linear(io_channels+cond_channels, 512) norm_first = False self.position = SinePositionalEmbedding( 512, dropout=0.1, scale=False, alpha=True) self.encoder = TransformerEncoder( TransformerEncoderLayer( d_model=512, nhead=16, dim_feedforward=2048, dropout=0.1, batch_first=True, norm_first=False, ), num_layers=12, norm=LayerNorm(512) if norm_first else None, ) self.proj = nn.Linear(512, io_channels) self.diffusion_embedding = SinusoidalPosEmb(512) def forward(self, x, t, cond, mask=None): if mask is None: mask = torch.ones(x.shape[0], x.shape[-1], device=x.device).long() mask_bool = ~mask.bool() diffusion_step = self.diffusion_embedding(t).unsqueeze(1) x = torch.cat([x, cond], dim=1).transpose(1, 2) x = self.in_proj(x) x = torch.cat([diffusion_step, x], dim=1) mask_bool = torch.cat([torch.zeros([mask_bool.shape[0], 1], device=mask_bool.device).bool(), mask_bool], dim=1) x = self.position(x) x, _ = self.encoder( (x, None), src_key_padding_mask=mask_bool) x = x[:, 1:, :] x = self.proj(x).transpose(1, 2) return x class ReFlow(nn.Module): def __init__(self, io_channels, cond_channels, ): super().__init__() self.model = DiT(io_channels, cond_channels) self.io_channels = io_channels self.rng = torch.quasirandom.SobolEngine(1, scramble=True) def forward(self, reals, cond, mask): # reals : [B, 128, T] # cond : [B, 192, T] # mask : [B, 1, T] t = self.rng.draw(reals.shape[0])[:, 0].to(reals.device) alphas, sigmas = 1 - t, t alphas = alphas[:, None, None] sigmas = sigmas[:, None, None] noise = torch.randn_like(reals) noised_inputs = reals * alphas + noise * sigmas targets = noise - reals output = self.model(noised_inputs, t, cond, mask) # mse loss with mask loss = F.mse_loss(output, targets, reduction='none') loss = loss * mask.unsqueeze(1) return loss.mean() def sample(self, cond, uncond, steps, guidance, sigma_max=1): x = torch.randn(cond.shape[0], self.io_channels, cond.shape[2]).to(cond.device) t = torch.linspace(sigma_max, 0, steps + 1) for t_curr, t_prev in tqdm(zip(t[:-1], t[1:])): t_curr_tensor = t_curr * torch.ones( (x.shape[0],), dtype=x.dtype, device=x.device ) dt = t_prev - t_curr v_cond = self.model(x, t_curr_tensor, cond) v_uncond = self.model(x, t_curr_tensor, uncond) v = guidance * v_cond + (1 - guidance) * v_uncond x = x + dt * v return x class NARPredictor(nn.Module): def __init__(self, latent_dim=16): super().__init__() self.latent_dim = latent_dim self.bert_proj = nn.Linear(1024, 192) self.x_emb = nn.Linear(192, 192) self.ref_enc = MelStyleEncoder(latent_dim, style_vector_dim=192) self.null_emb = nn.Parameter(torch.randn(192)) self.reflow = ReFlow(latent_dim, 192) def forward(self, x, bert, spk_emb, diffusion_input, mask): g = self.ref_enc(diffusion_input, mask.float().unsqueeze(1)).transpose(1, 2) bert = self.bert_proj(bert.transpose(1, 2)) x = self.x_emb(x.transpose(1,2)) cond = (x + bert + g).transpose(1, 2) # [B, 192, T] # dropout prob 0.1 for b in range(cond.shape[0]): if torch.rand(1) < 0.1: cond[b, :, :] = self.null_emb.unsqueeze(1).expand(-1, cond.shape[2]) loss = self.reflow(diffusion_input, cond, mask) return loss @torch.inference_mode() @torch.no_grad() def infer(self, x, bert, prompt, step, guidance): g = self.ref_enc(prompt).transpose(1, 2) bert = self.bert_proj(bert.transpose(1, 2)) x = self.x_emb(x.transpose(1,2)) cond = (x + bert + g).transpose(1, 2) uncond = self.null_emb.unsqueeze(1).expand(-1, cond.shape[2]).unsqueeze(0) fakes = self.reflow.sample(cond,uncond, step, guidance) return fakes class Wav2TTS(pl.LightningModule): def __init__(self, hp, vits_s2=None): super().__init__() self.model = NARPredictor(latent_dim=4) self.lr = hp.lr if vits_s2 is None: vits_s2 = init_vits_s2('configs/vits.json', 'logs/vits/G_38100.pth') self.s2 = [vits_s2] self.inited = False def configure_optimizers(self): print('train with lr ', self.lr) optimizer = optim.AdamW([*self.parameters()], lr=self.lr, betas=(0.9, 0.95), weight_decay=0.1) return optimizer def training_step(self, batch, batch_idx): loss = self.step(batch) self.log("train/loss", loss, on_step=True, prog_bar=True) # self.log("train/quantize_loss", quantize_loss, on_step=True, prog_bar=True) return loss def validation_step(self, batch, batch_idx): loss = self.step(batch) self.log("val/loss", loss, on_step=True, prog_bar=True) def step(self, batch): # speaker_embedding = self.spk_proj(batch['spk_emb']) # speaker_embedding, commitment_loss, codebook_loss, indices, z_e = self.spk_vq(speaker_embedding.unsqueeze(-1)) # quantize_loss = codebook_loss + 0.25*commitment_loss # speaker_embedding = speaker_embedding.squeeze(-1) # speaker_embedding = self.spk_emb(batch['spk_emb']) # Deal with phone segments phone = batch['phone'] lengths = batch['length'] bert_feature = batch['bert_feature'] speaker_embedding = batch['spk_emb'] latent = batch['latent'] mask = commons.sequence_mask(lengths, latent.size(-1)).long() # [1,1,1,1,0...] if not self.inited: self.s2 = [self.s2[0].to(phone.device)] self.inited = True s2 = self.s2[0] phone_features, _ = s2.enc_p(phone, mask.unsqueeze(1)) loss = self.model(phone_features, bert_feature, speaker_embedding, latent, mask) return loss @torch.inference_mode() @torch.no_grad() def inference(self, phonemes, bert_feature, prompt, n_steps=32, guidance=4): # speaker_embedding = self.spk_proj(spk_emb) # speaker_embedding, commitment_loss, codebook_loss, indices, z_e = self.spk_vq(speaker_embedding.unsqueeze(-1)) # speaker_embedding = speaker_embedding.squeeze(-1) # print('quantized spk code:', indices) mask = torch.ones(phonemes.shape[0],1, phonemes.shape[1]).to(phonemes.device).float() s2 = self.s2[0] phone_features, _ = s2.enc_p(phonemes, mask) output = self.model.infer(phone_features, bert_feature, prompt, n_steps, guidance) return output