""" from https://github.com/jaywalnut310/glow-tts""" import math import torch from torch import nn import tqdm from diffusers import DiffusionPipeline from diffusers.configuration_utils import ConfigMixin from diffusers.modeling_utils import ModelMixin from .grad_tts_utils import GradTTSTokenizer # flake8: noqa def sequence_mask(length, max_length=None): if max_length is None: max_length = length.max() x = torch.arange(int(max_length), dtype=length.dtype, device=length.device) return x.unsqueeze(0) < length.unsqueeze(1) def fix_len_compatibility(length, num_downsamplings_in_unet=2): while True: if length % (2**num_downsamplings_in_unet) == 0: return length length += 1 def convert_pad_shape(pad_shape): l = pad_shape[::-1] pad_shape = [item for sublist in l for item in sublist] return pad_shape def generate_path(duration, mask): device = duration.device b, t_x, t_y = mask.shape cum_duration = torch.cumsum(duration, 1) path = torch.zeros(b, t_x, t_y, dtype=mask.dtype).to(device=device) cum_duration_flat = cum_duration.view(b * t_x) path = sequence_mask(cum_duration_flat, t_y).to(mask.dtype) path = path.view(b, t_x, t_y) path = path - torch.nn.functional.pad(path, convert_pad_shape([[0, 0], [1, 0], [0, 0]]))[:, :-1] path = path * mask return path def duration_loss(logw, logw_, lengths): loss = torch.sum((logw - logw_) ** 2) / torch.sum(lengths) return loss class LayerNorm(nn.Module): def __init__(self, channels, eps=1e-4): super(LayerNorm, self).__init__() self.channels = channels self.eps = eps self.gamma = torch.nn.Parameter(torch.ones(channels)) self.beta = torch.nn.Parameter(torch.zeros(channels)) def forward(self, x): n_dims = len(x.shape) mean = torch.mean(x, 1, keepdim=True) variance = torch.mean((x - mean) ** 2, 1, keepdim=True) x = (x - mean) * torch.rsqrt(variance + self.eps) shape = [1, -1] + [1] * (n_dims - 2) x = x * self.gamma.view(*shape) + self.beta.view(*shape) return x class ConvReluNorm(nn.Module): def __init__(self, in_channels, hidden_channels, out_channels, kernel_size, n_layers, p_dropout): super(ConvReluNorm, self).__init__() self.in_channels = in_channels self.hidden_channels = hidden_channels self.out_channels = out_channels self.kernel_size = kernel_size self.n_layers = n_layers self.p_dropout = p_dropout self.conv_layers = torch.nn.ModuleList() self.norm_layers = torch.nn.ModuleList() self.conv_layers.append(torch.nn.Conv1d(in_channels, hidden_channels, kernel_size, padding=kernel_size // 2)) self.norm_layers.append(LayerNorm(hidden_channels)) self.relu_drop = torch.nn.Sequential(torch.nn.ReLU(), torch.nn.Dropout(p_dropout)) for _ in range(n_layers - 1): self.conv_layers.append( torch.nn.Conv1d(hidden_channels, hidden_channels, kernel_size, padding=kernel_size // 2) ) self.norm_layers.append(LayerNorm(hidden_channels)) self.proj = torch.nn.Conv1d(hidden_channels, out_channels, 1) self.proj.weight.data.zero_() self.proj.bias.data.zero_() def forward(self, x, x_mask): x_org = x for i in range(self.n_layers): x = self.conv_layers[i](x * x_mask) x = self.norm_layers[i](x) x = self.relu_drop(x) x = x_org + self.proj(x) return x * x_mask class DurationPredictor(nn.Module): def __init__(self, in_channels, filter_channels, kernel_size, p_dropout): super(DurationPredictor, self).__init__() self.in_channels = in_channels self.filter_channels = filter_channels self.p_dropout = p_dropout self.drop = torch.nn.Dropout(p_dropout) self.conv_1 = torch.nn.Conv1d(in_channels, filter_channels, kernel_size, padding=kernel_size // 2) self.norm_1 = LayerNorm(filter_channels) self.conv_2 = torch.nn.Conv1d(filter_channels, filter_channels, kernel_size, padding=kernel_size // 2) self.norm_2 = LayerNorm(filter_channels) self.proj = torch.nn.Conv1d(filter_channels, 1, 1) def forward(self, x, x_mask): x = self.conv_1(x * x_mask) x = torch.relu(x) x = self.norm_1(x) x = self.drop(x) x = self.conv_2(x * x_mask) x = torch.relu(x) x = self.norm_2(x) x = self.drop(x) x = self.proj(x * x_mask) return x * x_mask class MultiHeadAttention(nn.Module): def __init__( self, channels, out_channels, n_heads, window_size=None, heads_share=True, p_dropout=0.0, proximal_bias=False, proximal_init=False, ): super(MultiHeadAttention, self).__init__() assert channels % n_heads == 0 self.channels = channels self.out_channels = out_channels self.n_heads = n_heads self.window_size = window_size self.heads_share = heads_share self.proximal_bias = proximal_bias self.p_dropout = p_dropout self.attn = None self.k_channels = channels // n_heads self.conv_q = torch.nn.Conv1d(channels, channels, 1) self.conv_k = torch.nn.Conv1d(channels, channels, 1) self.conv_v = torch.nn.Conv1d(channels, channels, 1) if window_size is not None: n_heads_rel = 1 if heads_share else n_heads rel_stddev = self.k_channels**-0.5 self.emb_rel_k = torch.nn.Parameter( torch.randn(n_heads_rel, window_size * 2 + 1, self.k_channels) * rel_stddev ) self.emb_rel_v = torch.nn.Parameter( torch.randn(n_heads_rel, window_size * 2 + 1, self.k_channels) * rel_stddev ) self.conv_o = torch.nn.Conv1d(channels, out_channels, 1) self.drop = torch.nn.Dropout(p_dropout) torch.nn.init.xavier_uniform_(self.conv_q.weight) torch.nn.init.xavier_uniform_(self.conv_k.weight) if proximal_init: self.conv_k.weight.data.copy_(self.conv_q.weight.data) self.conv_k.bias.data.copy_(self.conv_q.bias.data) torch.nn.init.xavier_uniform_(self.conv_v.weight) def forward(self, x, c, attn_mask=None): q = self.conv_q(x) k = self.conv_k(c) v = self.conv_v(c) x, self.attn = self.attention(q, k, v, mask=attn_mask) x = self.conv_o(x) return x def attention(self, query, key, value, mask=None): b, d, t_s, t_t = (*key.size(), query.size(2)) query = query.view(b, self.n_heads, self.k_channels, t_t).transpose(2, 3) key = key.view(b, self.n_heads, self.k_channels, t_s).transpose(2, 3) value = value.view(b, self.n_heads, self.k_channels, t_s).transpose(2, 3) scores = torch.matmul(query, key.transpose(-2, -1)) / math.sqrt(self.k_channels) if self.window_size is not None: assert t_s == t_t, "Relative attention is only available for self-attention." key_relative_embeddings = self._get_relative_embeddings(self.emb_rel_k, t_s) rel_logits = self._matmul_with_relative_keys(query, key_relative_embeddings) rel_logits = self._relative_position_to_absolute_position(rel_logits) scores_local = rel_logits / math.sqrt(self.k_channels) scores = scores + scores_local if self.proximal_bias: assert t_s == t_t, "Proximal bias is only available for self-attention." scores = scores + self._attention_bias_proximal(t_s).to(device=scores.device, dtype=scores.dtype) if mask is not None: scores = scores.masked_fill(mask == 0, -1e4) p_attn = torch.nn.functional.softmax(scores, dim=-1) p_attn = self.drop(p_attn) output = torch.matmul(p_attn, value) if self.window_size is not None: relative_weights = self._absolute_position_to_relative_position(p_attn) value_relative_embeddings = self._get_relative_embeddings(self.emb_rel_v, t_s) output = output + self._matmul_with_relative_values(relative_weights, value_relative_embeddings) output = output.transpose(2, 3).contiguous().view(b, d, t_t) return output, p_attn def _matmul_with_relative_values(self, x, y): ret = torch.matmul(x, y.unsqueeze(0)) return ret def _matmul_with_relative_keys(self, x, y): ret = torch.matmul(x, y.unsqueeze(0).transpose(-2, -1)) return ret def _get_relative_embeddings(self, relative_embeddings, length): pad_length = max(length - (self.window_size + 1), 0) slice_start_position = max((self.window_size + 1) - length, 0) slice_end_position = slice_start_position + 2 * length - 1 if pad_length > 0: padded_relative_embeddings = torch.nn.functional.pad( relative_embeddings, convert_pad_shape([[0, 0], [pad_length, pad_length], [0, 0]]) ) else: padded_relative_embeddings = relative_embeddings used_relative_embeddings = padded_relative_embeddings[:, slice_start_position:slice_end_position] return used_relative_embeddings def _relative_position_to_absolute_position(self, x): batch, heads, length, _ = x.size() x = torch.nn.functional.pad(x, convert_pad_shape([[0, 0], [0, 0], [0, 0], [0, 1]])) x_flat = x.view([batch, heads, length * 2 * length]) x_flat = torch.nn.functional.pad(x_flat, convert_pad_shape([[0, 0], [0, 0], [0, length - 1]])) x_final = x_flat.view([batch, heads, length + 1, 2 * length - 1])[:, :, :length, length - 1 :] return x_final def _absolute_position_to_relative_position(self, x): batch, heads, length, _ = x.size() x = torch.nn.functional.pad(x, convert_pad_shape([[0, 0], [0, 0], [0, 0], [0, length - 1]])) x_flat = x.view([batch, heads, length**2 + length * (length - 1)]) x_flat = torch.nn.functional.pad(x_flat, convert_pad_shape([[0, 0], [0, 0], [length, 0]])) x_final = x_flat.view([batch, heads, length, 2 * length])[:, :, :, 1:] return x_final def _attention_bias_proximal(self, length): r = torch.arange(length, dtype=torch.float32) diff = torch.unsqueeze(r, 0) - torch.unsqueeze(r, 1) return torch.unsqueeze(torch.unsqueeze(-torch.log1p(torch.abs(diff)), 0), 0) class FFN(nn.Module): def __init__(self, in_channels, out_channels, filter_channels, kernel_size, p_dropout=0.0): super(FFN, self).__init__() self.in_channels = in_channels self.out_channels = out_channels self.filter_channels = filter_channels self.kernel_size = kernel_size self.p_dropout = p_dropout self.conv_1 = torch.nn.Conv1d(in_channels, filter_channels, kernel_size, padding=kernel_size // 2) self.conv_2 = torch.nn.Conv1d(filter_channels, out_channels, kernel_size, padding=kernel_size // 2) self.drop = torch.nn.Dropout(p_dropout) def forward(self, x, x_mask): x = self.conv_1(x * x_mask) x = torch.relu(x) x = self.drop(x) x = self.conv_2(x * x_mask) return x * x_mask class Encoder(nn.Module): def __init__( self, hidden_channels, filter_channels, n_heads, n_layers, kernel_size=1, p_dropout=0.0, window_size=None, **kwargs, ): super(Encoder, self).__init__() 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.window_size = window_size self.drop = torch.nn.Dropout(p_dropout) self.attn_layers = torch.nn.ModuleList() self.norm_layers_1 = torch.nn.ModuleList() self.ffn_layers = torch.nn.ModuleList() self.norm_layers_2 = torch.nn.ModuleList() for _ in range(self.n_layers): self.attn_layers.append( MultiHeadAttention( hidden_channels, hidden_channels, n_heads, window_size=window_size, p_dropout=p_dropout ) ) self.norm_layers_1.append(LayerNorm(hidden_channels)) self.ffn_layers.append( FFN(hidden_channels, hidden_channels, filter_channels, kernel_size, p_dropout=p_dropout) ) self.norm_layers_2.append(LayerNorm(hidden_channels)) def forward(self, x, x_mask): attn_mask = x_mask.unsqueeze(2) * x_mask.unsqueeze(-1) for i in range(self.n_layers): x = x * x_mask y = self.attn_layers[i](x, x, attn_mask) y = self.drop(y) x = self.norm_layers_1[i](x + y) y = self.ffn_layers[i](x, x_mask) y = self.drop(y) x = self.norm_layers_2[i](x + y) x = x * x_mask return x class TextEncoder(ModelMixin, ConfigMixin): def __init__( self, n_vocab, n_feats, n_channels, filter_channels, filter_channels_dp, n_heads, n_layers, kernel_size, p_dropout, window_size=None, spk_emb_dim=64, n_spks=1, ): super(TextEncoder, self).__init__() self.register_to_config( n_vocab=n_vocab, n_feats=n_feats, n_channels=n_channels, filter_channels=filter_channels, filter_channels_dp=filter_channels_dp, n_heads=n_heads, n_layers=n_layers, kernel_size=kernel_size, p_dropout=p_dropout, window_size=window_size, spk_emb_dim=spk_emb_dim, n_spks=n_spks, ) self.n_vocab = n_vocab self.n_feats = n_feats self.n_channels = n_channels self.filter_channels = filter_channels self.filter_channels_dp = filter_channels_dp self.n_heads = n_heads self.n_layers = n_layers self.kernel_size = kernel_size self.p_dropout = p_dropout self.window_size = window_size self.spk_emb_dim = spk_emb_dim self.n_spks = n_spks self.emb = torch.nn.Embedding(n_vocab, n_channels) torch.nn.init.normal_(self.emb.weight, 0.0, n_channels**-0.5) self.prenet = ConvReluNorm(n_channels, n_channels, n_channels, kernel_size=5, n_layers=3, p_dropout=0.5) self.encoder = Encoder( n_channels + (spk_emb_dim if n_spks > 1 else 0), filter_channels, n_heads, n_layers, kernel_size, p_dropout, window_size=window_size, ) self.proj_m = torch.nn.Conv1d(n_channels + (spk_emb_dim if n_spks > 1 else 0), n_feats, 1) self.proj_w = DurationPredictor( n_channels + (spk_emb_dim if n_spks > 1 else 0), filter_channels_dp, kernel_size, p_dropout ) def forward(self, x, x_lengths, spk=None): x = self.emb(x) * math.sqrt(self.n_channels) x = torch.transpose(x, 1, -1) x_mask = torch.unsqueeze(sequence_mask(x_lengths, x.size(2)), 1).to(x.dtype) x = self.prenet(x, x_mask) if self.n_spks > 1: x = torch.cat([x, spk.unsqueeze(-1).repeat(1, 1, x.shape[-1])], dim=1) x = self.encoder(x, x_mask) mu = self.proj_m(x) * x_mask x_dp = torch.detach(x) logw = self.proj_w(x_dp, x_mask) return mu, logw, x_mask class GradTTSPipeline(DiffusionPipeline): def __init__(self, unet, text_encoder, noise_scheduler, tokenizer): super().__init__() noise_scheduler = noise_scheduler.set_format("pt") self.register_modules( unet=unet, text_encoder=text_encoder, noise_scheduler=noise_scheduler, tokenizer=tokenizer ) @torch.no_grad() def __call__( self, text, num_inference_steps=50, temperature=1.3, length_scale=0.91, speaker_id=15, torch_device=None, generator=None, ): if torch_device is None: torch_device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu") self.unet.to(torch_device) self.text_encoder.to(torch_device) x, x_lengths = self.tokenizer(text) x = x.to(torch_device) x_lengths = x_lengths.to(torch_device) if speaker_id is not None: speaker_id = torch.LongTensor([speaker_id]).to(torch_device) # Get encoder_outputs `mu_x` and log-scaled token durations `logw` mu_x, logw, x_mask = self.text_encoder(x, x_lengths) w = torch.exp(logw) * x_mask w_ceil = torch.ceil(w) * length_scale y_lengths = torch.clamp_min(torch.sum(w_ceil, [1, 2]), 1).long() y_max_length = int(y_lengths.max()) y_max_length_ = fix_len_compatibility(y_max_length) # Using obtained durations `w` construct alignment map `attn` y_mask = sequence_mask(y_lengths, y_max_length_).unsqueeze(1).to(x_mask.dtype) attn_mask = x_mask.unsqueeze(-1) * y_mask.unsqueeze(2) attn = generate_path(w_ceil.squeeze(1), attn_mask.squeeze(1)).unsqueeze(1) # Align encoded text and get mu_y mu_y = torch.matmul(attn.squeeze(1).transpose(1, 2), mu_x.transpose(1, 2)) mu_y = mu_y.transpose(1, 2) # Sample latent representation from terminal distribution N(mu_y, I) z = mu_y + torch.randn(mu_y.shape, generator=generator).to(mu_y.device) xt = z * y_mask h = 1.0 / num_inference_steps # (Patrick: TODO) for t in tqdm.tqdm(range(num_inference_steps), total=num_inference_steps): t_new = num_inference_steps - t - 1 t = (1.0 - (t + 0.5) * h) * torch.ones(z.shape[0], dtype=z.dtype, device=z.device) residual = self.unet(xt, t, mu_y, y_mask, speaker_id) scheduler_residual = residual - mu_y + xt xt = self.noise_scheduler.step(scheduler_residual, xt, t_new, num_inference_steps) xt = xt * y_mask return xt[:, :, :y_max_length]