Initial commit

cosyvoice/flow/flow.py

+# Copyright (c) 2024 Alibaba Inc (authors: Xiang Lyu, Zhihao Du)
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import logging
+import random
+from typing import Dict, Optional
+import torch
+import torch.nn as nn
+from torch.nn import functional as F
+from omegaconf import DictConfig
+from cosyvoice.utils.mask import make_pad_mask
+
+
+class MaskedDiffWithXvec(torch.nn.Module):
+    def __init__(self,
+                 input_size: int = 512,
+                 output_size: int = 80,
+                 spk_embed_dim: int = 192,
+                 output_type: str = "mel",
+                 vocab_size: int = 4096,
+                 input_frame_rate: int = 50,
+                 only_mask_loss: bool = True,
+                 encoder: torch.nn.Module = None,
+                 length_regulator: torch.nn.Module = None,
+                 decoder: torch.nn.Module = None,
+                 decoder_conf: Dict = {'in_channels': 240, 'out_channel': 80, 'spk_emb_dim': 80, 'n_spks': 1, 'cfm_params': DictConfig({'sigma_min': 1e-06, 'solver': 'euler', 't_scheduler': 'cosine', 'training_cfg_rate': 0.2, 'inference_cfg_rate': 0.7, 'reg_loss_type': 'l1'}), 'decoder_params': {'channels': [256, 256], 'dropout': 0.0, 'attention_head_dim': 64, 'n_blocks': 4, 'num_mid_blocks': 12, 'num_heads': 8, 'act_fn': 'gelu'}},
+                 mel_feat_conf: Dict = {'n_fft': 1024, 'num_mels': 80, 'sampling_rate': 22050, 'hop_size': 256, 'win_size': 1024, 'fmin': 0, 'fmax': 8000}):
+        super().__init__()
+        self.input_size = input_size
+        self.output_size = output_size
+        self.decoder_conf = decoder_conf
+        self.mel_feat_conf = mel_feat_conf
+        self.vocab_size = vocab_size
+        self.output_type = output_type
+        self.input_frame_rate = input_frame_rate
+        logging.info(f"input frame rate={self.input_frame_rate}")
+        self.input_embedding = nn.Embedding(vocab_size, input_size)
+        self.spk_embed_affine_layer = torch.nn.Linear(spk_embed_dim, output_size)
+        self.encoder = encoder
+        self.encoder_proj = torch.nn.Linear(self.encoder.output_size(), output_size)
+        self.decoder = decoder
+        self.length_regulator = length_regulator
+        self.only_mask_loss = only_mask_loss
+
+    def forward(
+            self,
+            batch: dict,
+            device: torch.device,
+    ) -> Dict[str, Optional[torch.Tensor]]:
+        token = batch['speech_token'].to(device)
+        token_len = batch['speech_token_len'].to(device)
+        feat = batch['speech_feat'].to(device)
+        feat_len = batch['speech_feat_len'].to(device)
+        embedding = batch['embedding'].to(device)
+
+        # xvec projection
+        embedding = F.normalize(embedding, dim=1)
+        embedding = self.spk_embed_affine_layer(embedding)
+
+        # concat text and prompt_text
+        mask = (~make_pad_mask(token_len)).float().unsqueeze(-1).to(device)
+        token = self.input_embedding(torch.clamp(token, min=0)) * mask
+
+        # text encode
+        h, h_lengths = self.encoder(token, token_len)
+        h = self.encoder_proj(h)
+        h, h_lengths = self.length_regulator(h, feat_len)
+
+        # get conditions
+        conds = torch.zeros(feat.shape, device=token.device)
+        for i, j in enumerate(feat_len):
+            if random.random() < 0.5:
+                continue
+            index = random.randint(0, int(0.3 * j))
+            conds[i, :index] = feat[i, :index]
+        conds = conds.transpose(1, 2)
+
+        mask = (~make_pad_mask(feat_len)).to(h)
+        feat = F.interpolate(feat.unsqueeze(dim=1), size=h.shape[1:], mode="nearest").squeeze(dim=1)
+        loss, _ = self.decoder.compute_loss(
+            feat.transpose(1, 2).contiguous(),
+            mask.unsqueeze(1),
+            h.transpose(1, 2).contiguous(),
+            embedding,
+            cond=conds
+        )
+        return {'loss': loss}
+
+    @torch.inference_mode()
+    def inference(self,
+                  token,
+                  token_len,
+                  prompt_token,
+                  prompt_token_len,
+                  prompt_feat,
+                  prompt_feat_len,
+                  embedding):
+        assert token.shape[0] == 1
+        # xvec projection
+        embedding = F.normalize(embedding, dim=1)
+        embedding = self.spk_embed_affine_layer(embedding)
+
+        # concat text and prompt_text
+        token, token_len = torch.concat([prompt_token, token], dim=1), prompt_token_len + token_len
+        mask = (~make_pad_mask(token_len)).float().unsqueeze(-1).to(embedding)
+        token = self.input_embedding(torch.clamp(token, min=0)) * mask
+
+        # text encode
+        h, h_lengths = self.encoder(token, token_len)
+        h = self.encoder_proj(h)
+        feat_len = (token_len / 50 * 22050 / 256).int()
+        h, h_lengths = self.length_regulator(h, feat_len)
+
+        # get conditions
+        conds = torch.zeros([1, feat_len.max().item(), self.output_size], device=token.device)
+        if prompt_feat.shape[1] != 0:
+            for i, j in enumerate(prompt_feat_len):
+                conds[i, :j] = prompt_feat[i]
+        conds = conds.transpose(1, 2)
+
+        mask = (~make_pad_mask(feat_len)).to(h)
+        feat = self.decoder(
+            mu=h.transpose(1, 2).contiguous(),
+            mask=mask.unsqueeze(1),
+            spks=embedding,
+            cond=conds,
+            n_timesteps=10
+        )
+        if prompt_feat.shape[1] != 0:
+            feat = feat[:, :, prompt_feat.shape[1]:]
+        return feat

cosyvoice/flow/flow_matching.py

+# Copyright (c) 2024 Alibaba Inc (authors: Xiang Lyu, Zhihao Du)
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import torch
+import torch.nn.functional as F
+from matcha.models.components.flow_matching import BASECFM
+
+class ConditionalCFM(BASECFM):
+    def __init__(self, in_channels, cfm_params, n_spks=1, spk_emb_dim=64, estimator: torch.nn.Module = None):
+        super().__init__(
+            n_feats=in_channels,
+            cfm_params=cfm_params,
+            n_spks=n_spks,
+            spk_emb_dim=spk_emb_dim,
+        )
+        self.t_scheduler = cfm_params.t_scheduler
+        self.training_cfg_rate = cfm_params.training_cfg_rate
+        self.inference_cfg_rate = cfm_params.inference_cfg_rate
+        in_channels = in_channels + (spk_emb_dim if n_spks > 0 else 0)
+        # Just change the architecture of the estimator here
+        self.estimator = estimator
+
+    @torch.inference_mode()
+    def forward(self, mu, mask, n_timesteps, temperature=1.0, spks=None, cond=None):
+        """Forward diffusion
+
+        Args:
+            mu (torch.Tensor): output of encoder
+                shape: (batch_size, n_feats, mel_timesteps)
+            mask (torch.Tensor): output_mask
+                shape: (batch_size, 1, mel_timesteps)
+            n_timesteps (int): number of diffusion steps
+            temperature (float, optional): temperature for scaling noise. Defaults to 1.0.
+            spks (torch.Tensor, optional): speaker ids. Defaults to None.
+                shape: (batch_size, spk_emb_dim)
+            cond: Not used but kept for future purposes
+
+        Returns:
+            sample: generated mel-spectrogram
+                shape: (batch_size, n_feats, mel_timesteps)
+        """
+        z = torch.randn_like(mu) * temperature
+        t_span = torch.linspace(0, 1, n_timesteps + 1, device=mu.device)
+        if self.t_scheduler == 'cosine':
+            t_span = 1 - torch.cos(t_span * 0.5 * torch.pi)
+        return self.solve_euler(z, t_span=t_span, mu=mu, mask=mask, spks=spks, cond=cond)
+
+    def solve_euler(self, x, t_span, mu, mask, spks, cond):
+        """
+        Fixed euler solver for ODEs.
+        Args:
+            x (torch.Tensor): random noise
+            t_span (torch.Tensor): n_timesteps interpolated
+                shape: (n_timesteps + 1,)
+            mu (torch.Tensor): output of encoder
+                shape: (batch_size, n_feats, mel_timesteps)
+            mask (torch.Tensor): output_mask
+                shape: (batch_size, 1, mel_timesteps)
+            spks (torch.Tensor, optional): speaker ids. Defaults to None.
+                shape: (batch_size, spk_emb_dim)
+            cond: Not used but kept for future purposes
+        """
+        t, _, dt = t_span[0], t_span[-1], t_span[1] - t_span[0]
+
+        # I am storing this because I can later plot it by putting a debugger here and saving it to a file
+        # Or in future might add like a return_all_steps flag
+        sol = []
+
+        for step in range(1, len(t_span)):
+            dphi_dt = self.estimator(x, mask, mu, t, spks, cond)
+            # Classifier-Free Guidance inference introduced in VoiceBox
+            if self.inference_cfg_rate > 0:
+                cfg_dphi_dt = self.estimator(
+                    x, mask,
+                    torch.zeros_like(mu), t,
+                    torch.zeros_like(spks) if spks is not None else None,
+                    torch.zeros_like(cond)
+                )
+                dphi_dt = ((1.0 + self.inference_cfg_rate) * dphi_dt -
+                           self.inference_cfg_rate * cfg_dphi_dt)
+            x = x + dt * dphi_dt
+            t = t + dt
+            sol.append(x)
+            if step < len(t_span) - 1:
+                dt = t_span[step + 1] - t
+
+        return sol[-1]
+
+    def compute_loss(self, x1, mask, mu, spks=None, cond=None):
+        """Computes diffusion loss
+
+        Args:
+            x1 (torch.Tensor): Target
+                shape: (batch_size, n_feats, mel_timesteps)
+            mask (torch.Tensor): target mask
+                shape: (batch_size, 1, mel_timesteps)
+            mu (torch.Tensor): output of encoder
+                shape: (batch_size, n_feats, mel_timesteps)
+            spks (torch.Tensor, optional): speaker embedding. Defaults to None.
+                shape: (batch_size, spk_emb_dim)
+
+        Returns:
+            loss: conditional flow matching loss
+            y: conditional flow
+                shape: (batch_size, n_feats, mel_timesteps)
+        """
+        b, _, t = mu.shape
+
+        # random timestep
+        t = torch.rand([b, 1, 1], device=mu.device, dtype=mu.dtype)
+        if self.t_scheduler == 'cosine':
+            t = 1 - torch.cos(t * 0.5 * torch.pi)
+        # sample noise p(x_0)
+        z = torch.randn_like(x1)
+
+        y = (1 - (1 - self.sigma_min) * t) * z + t * x1
+        u = x1 - (1 - self.sigma_min) * z
+
+        # during training, we randomly drop condition to trade off mode coverage and sample fidelity
+        if self.training_cfg_rate > 0:
+            cfg_mask = torch.rand(b, device=x1.device) > self.training_cfg_rate
+            mu = mu * cfg_mask.view(-1, 1, 1)
+            spks = spks * cfg_mask.view(-1, 1)
+            cond = cond * cfg_mask.view(-1, 1, 1)
+
+        pred = self.estimator(y, mask, mu, t.squeeze(), spks, cond)
+        loss = F.mse_loss(pred * mask, u * mask, reduction="sum") / (torch.sum(mask) * u.shape[1])
+        return loss, y

cosyvoice/flow/length_regulator.py

+# Copyright (c) 2024 Alibaba Inc (authors: Xiang Lyu, Zhihao Du)
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from typing import Tuple
+import torch.nn as nn
+from torch.nn import functional as F
+from cosyvoice.utils.mask import make_pad_mask
+
+
+class InterpolateRegulator(nn.Module):
+    def __init__(
+            self,
+            channels: int,
+            sampling_ratios: Tuple,
+            out_channels: int = None,
+            groups: int = 1,
+    ):
+        super().__init__()
+        self.sampling_ratios = sampling_ratios
+        out_channels = out_channels or channels
+        model = nn.ModuleList([])
+        if len(sampling_ratios) > 0:
+            for _ in sampling_ratios:
+                module = nn.Conv1d(channels, channels, 3, 1, 1)
+                norm = nn.GroupNorm(groups, channels)
+                act = nn.Mish()
+                model.extend([module, norm, act])
+        model.append(
+            nn.Conv1d(channels, out_channels, 1, 1)
+        )
+        self.model = nn.Sequential(*model)
+
+    def forward(self, x, ylens=None):
+        # x in (B, T, D)
+        mask = (~make_pad_mask(ylens)).to(x).unsqueeze(-1)
+        x = F.interpolate(x.transpose(1, 2).contiguous(), size=ylens.max(), mode='nearest')
+        out = self.model(x).transpose(1, 2).contiguous()
+        olens = ylens
+        return out * mask, olens

cosyvoice/hifigan/f0_predictor.py

+# Copyright (c) 2024 Alibaba Inc (authors: Xiang Lyu, Kai Hu)
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import torch
+import torch.nn as nn
+from torch.nn.utils import weight_norm
+
+
+class ConvRNNF0Predictor(nn.Module):
+    def __init__(self,
+                 num_class: int = 1,
+                 in_channels: int = 80,
+                 cond_channels: int = 512
+                 ):
+        super().__init__()
+
+        self.num_class = num_class
+        self.condnet = nn.Sequential(
+            weight_norm(
+                nn.Conv1d(in_channels, cond_channels, kernel_size=3, padding=1)
+            ),
+            nn.ELU(),
+            weight_norm(
+                nn.Conv1d(cond_channels, cond_channels, kernel_size=3, padding=1)
+            ),
+            nn.ELU(),
+            weight_norm(
+                nn.Conv1d(cond_channels, cond_channels, kernel_size=3, padding=1)
+            ),
+            nn.ELU(),
+            weight_norm(
+                nn.Conv1d(cond_channels, cond_channels, kernel_size=3, padding=1)
+            ),
+            nn.ELU(),
+            weight_norm(
+                nn.Conv1d(cond_channels, cond_channels, kernel_size=3, padding=1)
+            ),
+            nn.ELU(),
+        )
+        self.classifier = nn.Linear(in_features=cond_channels, out_features=self.num_class)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x = self.condnet(x)
+        x = x.transpose(1, 2)
+        return torch.abs(self.classifier(x).squeeze(-1))

cosyvoice/hifigan/generator.py

+# Copyright (c) 2024 Alibaba Inc (authors: Xiang Lyu, Kai Hu)
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""HIFI-GAN"""
+
+import typing as tp
+import numpy as np
+from scipy.signal import get_window
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torch.nn import Conv1d
+from torch.nn import ConvTranspose1d
+from torch.nn.utils import remove_weight_norm
+from torch.nn.utils import weight_norm
+from torch.distributions.uniform import Uniform
+
+from cosyvoice.transformer.activation import Snake
+from cosyvoice.utils.common import get_padding
+from cosyvoice.utils.common import init_weights
+
+
+"""hifigan based generator implementation.
+
+This code is modified from https://github.com/jik876/hifi-gan
+ ,https://github.com/kan-bayashi/ParallelWaveGAN and
+ https://github.com/NVIDIA/BigVGAN
+
+"""
+class ResBlock(torch.nn.Module):
+    """Residual block module in HiFiGAN/BigVGAN."""
+    def __init__(
+        self,
+        channels: int = 512,
+        kernel_size: int = 3,
+        dilations: tp.List[int] = [1, 3, 5],
+    ):
+        super(ResBlock, self).__init__()
+        self.convs1 = nn.ModuleList()
+        self.convs2 = nn.ModuleList()
+
+        for dilation in dilations:
+            self.convs1.append(
+                weight_norm(
+                    Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=dilation,
+                        padding=get_padding(kernel_size, dilation)
+                    )
+                )
+            )
+            self.convs2.append(
+                weight_norm(
+                    Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=1,
+                        padding=get_padding(kernel_size, 1)
+                    )
+                )
+            )
+        self.convs1.apply(init_weights)
+        self.convs2.apply(init_weights)
+        self.activations1 = nn.ModuleList([
+            Snake(channels, alpha_logscale=False)
+            for _ in range(len(self.convs1))
+        ])
+        self.activations2 = nn.ModuleList([
+            Snake(channels, alpha_logscale=False)
+            for _ in range(len(self.convs2))
+        ])
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        for idx in range(len(self.convs1)):
+            xt = self.activations1[idx](x)
+            xt = self.convs1[idx](xt)
+            xt = self.activations2[idx](xt)
+            xt = self.convs2[idx](xt)
+            x = xt + x
+        return x
+
+    def remove_weight_norm(self):
+        for idx in range(len(self.convs1)):
+            remove_weight_norm(self.convs1[idx])
+            remove_weight_norm(self.convs2[idx])
+
+class SineGen(torch.nn.Module):
+    """ Definition of sine generator
+    SineGen(samp_rate, harmonic_num = 0,
+            sine_amp = 0.1, noise_std = 0.003,
+            voiced_threshold = 0,
+            flag_for_pulse=False)
+    samp_rate: sampling rate in Hz
+    harmonic_num: number of harmonic overtones (default 0)
+    sine_amp: amplitude of sine-wavefrom (default 0.1)
+    noise_std: std of Gaussian noise (default 0.003)
+    voiced_thoreshold: F0 threshold for U/V classification (default 0)
+    flag_for_pulse: this SinGen is used inside PulseGen (default False)
+    Note: when flag_for_pulse is True, the first time step of a voiced
+        segment is always sin(np.pi) or cos(0)
+    """
+
+    def __init__(self, samp_rate, harmonic_num=0,
+                 sine_amp=0.1, noise_std=0.003,
+                 voiced_threshold=0):
+        super(SineGen, self).__init__()
+        self.sine_amp = sine_amp
+        self.noise_std = noise_std
+        self.harmonic_num = harmonic_num
+        self.sampling_rate = samp_rate
+        self.voiced_threshold = voiced_threshold
+
+    def _f02uv(self, f0):
+        # generate uv signal
+        uv = (f0 > self.voiced_threshold).type(torch.float32)
+        return uv
+
+    @torch.no_grad()
+    def forward(self, f0):
+        """
+        :param f0: [B, 1, sample_len], Hz
+        :return: [B, 1, sample_len]
+        """
+
+        F_mat = torch.zeros((f0.size(0), self.harmonic_num + 1, f0.size(-1))).to(f0.device)
+        for i in range(self.harmonic_num + 1):
+            F_mat[:, i: i + 1, :] = f0 * (i + 1) / self.sampling_rate
+
+        theta_mat = 2 * np.pi * (torch.cumsum(F_mat, dim=-1) % 1)
+        u_dist = Uniform(low=-np.pi, high=np.pi)
+        phase_vec = u_dist.sample(sample_shape=(f0.size(0), self.harmonic_num + 1, 1)).to(F_mat.device)
+        phase_vec[:, 0, :] = 0
+
+        # generate sine waveforms
+        sine_waves = self.sine_amp * torch.sin(theta_mat + phase_vec)
+
+        # generate uv signal
+        uv = self._f02uv(f0)
+
+        # noise: for unvoiced should be similar to sine_amp
+        #        std = self.sine_amp/3 -> max value ~ self.sine_amp
+        # .       for voiced regions is self.noise_std
+        noise_amp = uv * self.noise_std + (1 - uv) * self.sine_amp / 3
+        noise = noise_amp * torch.randn_like(sine_waves)
+
+        # first: set the unvoiced part to 0 by uv
+        # then: additive noise
+        sine_waves = sine_waves * uv + noise
+        return sine_waves, uv, noise
+
+
+class SourceModuleHnNSF(torch.nn.Module):
+    """ SourceModule for hn-nsf
+    SourceModule(sampling_rate, harmonic_num=0, sine_amp=0.1,
+                 add_noise_std=0.003, voiced_threshod=0)
+    sampling_rate: sampling_rate in Hz
+    harmonic_num: number of harmonic above F0 (default: 0)
+    sine_amp: amplitude of sine source signal (default: 0.1)
+    add_noise_std: std of additive Gaussian noise (default: 0.003)
+        note that amplitude of noise in unvoiced is decided
+        by sine_amp
+    voiced_threshold: threhold to set U/V given F0 (default: 0)
+    Sine_source, noise_source = SourceModuleHnNSF(F0_sampled)
+    F0_sampled (batchsize, length, 1)
+    Sine_source (batchsize, length, 1)
+    noise_source (batchsize, length 1)
+    uv (batchsize, length, 1)
+    """
+
+    def __init__(self, sampling_rate, upsample_scale, harmonic_num=0, sine_amp=0.1,
+                 add_noise_std=0.003, voiced_threshod=0):
+        super(SourceModuleHnNSF, self).__init__()
+
+        self.sine_amp = sine_amp
+        self.noise_std = add_noise_std
+
+        # to produce sine waveforms
+        self.l_sin_gen = SineGen(sampling_rate, harmonic_num,
+                                 sine_amp, add_noise_std, voiced_threshod)
+
+        # to merge source harmonics into a single excitation
+        self.l_linear = torch.nn.Linear(harmonic_num + 1, 1)
+        self.l_tanh = torch.nn.Tanh()
+
+    def forward(self, x):
+        """
+        Sine_source, noise_source = SourceModuleHnNSF(F0_sampled)
+        F0_sampled (batchsize, length, 1)
+        Sine_source (batchsize, length, 1)
+        noise_source (batchsize, length 1)
+        """
+        # source for harmonic branch
+        with torch.no_grad():
+            sine_wavs, uv, _ = self.l_sin_gen(x.transpose(1, 2))
+            sine_wavs = sine_wavs.transpose(1, 2)
+            uv = uv.transpose(1, 2)
+        sine_merge = self.l_tanh(self.l_linear(sine_wavs))
+
+        # source for noise branch, in the same shape as uv
+        noise = torch.randn_like(uv) * self.sine_amp / 3
+        return sine_merge, noise, uv
+
+
+class HiFTGenerator(nn.Module):
+    """
+    HiFTNet Generator: Neural Source Filter + ISTFTNet
+    https://arxiv.org/abs/2309.09493
+    """
+    def __init__(
+            self,
+            in_channels: int = 80,
+            base_channels: int = 512,
+            nb_harmonics: int = 8,
+            sampling_rate: int = 22050,
+            nsf_alpha: float = 0.1,
+            nsf_sigma: float = 0.003,
+            nsf_voiced_threshold: float = 10,
+            upsample_rates: tp.List[int] = [8, 8],
+            upsample_kernel_sizes: tp.List[int] = [16, 16],
+            istft_params: tp.Dict[str, int] = {"n_fft": 16, "hop_len": 4},
+            resblock_kernel_sizes: tp.List[int] = [3, 7, 11],
+            resblock_dilation_sizes: tp.List[tp.List[int]] = [[1, 3, 5], [1, 3, 5], [1, 3, 5]],
+            source_resblock_kernel_sizes: tp.List[int] = [7, 11],
+            source_resblock_dilation_sizes: tp.List[tp.List[int]] = [[1, 3, 5], [1, 3, 5]],
+            lrelu_slope: float = 0.1,
+            audio_limit: float = 0.99,
+            f0_predictor: torch.nn.Module = None,
+    ):
+        super(HiFTGenerator, self).__init__()
+
+        self.out_channels = 1
+        self.nb_harmonics = nb_harmonics
+        self.sampling_rate = sampling_rate
+        self.istft_params = istft_params
+        self.lrelu_slope = lrelu_slope
+        self.audio_limit = audio_limit
+
+        self.num_kernels = len(resblock_kernel_sizes)
+        self.num_upsamples = len(upsample_rates)
+        self.m_source = SourceModuleHnNSF(
+            sampling_rate=sampling_rate,
+            upsample_scale=np.prod(upsample_rates) * istft_params["hop_len"],
+            harmonic_num=nb_harmonics,
+            sine_amp=nsf_alpha,
+            add_noise_std=nsf_sigma,
+            voiced_threshod=nsf_voiced_threshold)
+        self.f0_upsamp = torch.nn.Upsample(scale_factor=np.prod(upsample_rates) * istft_params["hop_len"])
+
+        self.conv_pre = weight_norm(
+            Conv1d(in_channels, base_channels, 7, 1, padding=3)
+        )
+
+        # Up
+        self.ups = nn.ModuleList()
+        for i, (u, k) in enumerate(zip(upsample_rates, upsample_kernel_sizes)):
+            self.ups.append(
+                weight_norm(
+                    ConvTranspose1d(
+                        base_channels // (2**i),
+                        base_channels // (2**(i + 1)),
+                        k,
+                        u,
+                        padding=(k - u) // 2,
+                    )
+                )
+            )
+
+        # Down
+        self.source_downs = nn.ModuleList()
+        self.source_resblocks = nn.ModuleList()
+        downsample_rates = [1] + upsample_rates[::-1][:-1]
+        downsample_cum_rates = np.cumprod(downsample_rates)
+        for i, (u, k, d) in enumerate(zip(downsample_cum_rates[::-1], source_resblock_kernel_sizes,
+                                          source_resblock_dilation_sizes)):
+            if u == 1:
+                self.source_downs.append(
+                    Conv1d(istft_params["n_fft"] + 2, base_channels // (2 ** (i + 1)), 1, 1)
+                )
+            else:
+                self.source_downs.append(
+                    Conv1d(istft_params["n_fft"] + 2, base_channels // (2 ** (i + 1)), u * 2, u, padding=(u // 2))
+                )
+
+            self.source_resblocks.append(
+                ResBlock(base_channels // (2 ** (i + 1)), k, d)
+            )
+
+        self.resblocks = nn.ModuleList()
+        for i in range(len(self.ups)):
+            ch = base_channels // (2**(i + 1))
+            for j, (k, d) in enumerate(zip(resblock_kernel_sizes, resblock_dilation_sizes)):
+                self.resblocks.append(ResBlock(ch, k, d))
+
+        self.conv_post = weight_norm(Conv1d(ch, istft_params["n_fft"] + 2, 7, 1, padding=3))
+        self.ups.apply(init_weights)
+        self.conv_post.apply(init_weights)
+        self.reflection_pad = nn.ReflectionPad1d((1, 0))
+        self.stft_window = torch.from_numpy(get_window("hann", istft_params["n_fft"], fftbins=True).astype(np.float32))
+        self.f0_predictor = f0_predictor
+
+    def _f02source(self, f0: torch.Tensor) -> torch.Tensor:
+        f0 = self.f0_upsamp(f0[:, None]).transpose(1, 2)  # bs,n,t
+
+        har_source, _, _ = self.m_source(f0)
+        return har_source.transpose(1, 2)
+
+    def _stft(self, x):
+        spec = torch.stft(
+            x,
+            self.istft_params["n_fft"], self.istft_params["hop_len"], self.istft_params["n_fft"], window=self.stft_window.to(x.device),
+            return_complex=True)
+        spec = torch.view_as_real(spec)  # [B, F, TT, 2]
+        return spec[..., 0], spec[..., 1]
+
+    def _istft(self, magnitude, phase):
+        magnitude = torch.clip(magnitude, max=1e2)
+        real = magnitude * torch.cos(phase)
+        img = magnitude * torch.sin(phase)
+        inverse_transform = torch.istft(torch.complex(real, img), self.istft_params["n_fft"], self.istft_params["hop_len"], self.istft_params["n_fft"], window=self.stft_window.to(magnitude.device))
+        return inverse_transform
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        f0 = self.f0_predictor(x)
+        s = self._f02source(f0)
+
+        s_stft_real, s_stft_imag = self._stft(s.squeeze(1))
+        s_stft = torch.cat([s_stft_real, s_stft_imag], dim=1)
+
+        x = self.conv_pre(x)
+        for i in range(self.num_upsamples):
+            x = F.leaky_relu(x, self.lrelu_slope)
+            x = self.ups[i](x)
+
+            if i == self.num_upsamples - 1:
+                x = self.reflection_pad(x)
+
+            # fusion
+            si = self.source_downs[i](s_stft)
+            si = self.source_resblocks[i](si)
+            x = x + si
+
+            xs = None
+            for j in range(self.num_kernels):
+                if xs is None:
+                    xs = self.resblocks[i * self.num_kernels + j](x)
+                else:
+                    xs += self.resblocks[i * self.num_kernels + j](x)
+            x = xs / self.num_kernels
+
+        x = F.leaky_relu(x)
+        x = self.conv_post(x)
+        magnitude = torch.exp(x[:, :self.istft_params["n_fft"] // 2 + 1, :])
+        phase = torch.sin(x[:, self.istft_params["n_fft"] // 2 + 1:, :])  # actually, sin is redundancy
+
+        x = self._istft(magnitude, phase)
+        x = torch.clamp(x, -self.audio_limit, self.audio_limit)
+        return x
+
+    def remove_weight_norm(self):
+        print('Removing weight norm...')
+        for l in self.ups:
+            remove_weight_norm(l)
+        for l in self.resblocks:
+            l.remove_weight_norm()
+        remove_weight_norm(self.conv_pre)
+        remove_weight_norm(self.conv_post)
+        self.source_module.remove_weight_norm()
+        for l in self.source_downs:
+            remove_weight_norm(l)
+        for l in self.source_resblocks:
+            l.remove_weight_norm()
+
+    @torch.inference_mode()
+    def inference(self, mel: torch.Tensor) -> torch.Tensor:
+        return self.forward(x=mel)

cosyvoice/llm/llm.py

+# Copyright (c) 2024 Alibaba Inc (authors: Xiang Lyu, Zhihao Du)
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from typing import Dict, Optional, Union
+import torch
+from torch import nn
+import torch.nn.functional as F
+from torch.nn.utils.rnn import pad_sequence, unpad_sequence
+from cosyvoice.utils.common import IGNORE_ID
+from cosyvoice.transformer.label_smoothing_loss import LabelSmoothingLoss
+from cosyvoice.utils.common import th_accuracy
+
+
+class TransformerLM(torch.nn.Module):
+    def __init__(
+            self,
+            text_encoder_input_size: int,
+            llm_input_size: int,
+            llm_output_size: int,
+            text_token_size: int,
+            speech_token_size: int,
+            text_encoder: torch.nn.Module,
+            llm: torch.nn.Module,
+            length_normalized_loss: bool = True,
+            lsm_weight: float = 0.0,
+            spk_embed_dim: int = 192,
+    ):
+        super().__init__()
+        self.llm_input_size = llm_input_size
+        self.speech_token_size = speech_token_size
+        # 1. build text token inputs related modules
+        self.text_embedding = torch.nn.Embedding(text_token_size, text_encoder_input_size)
+        self.text_encoder = text_encoder
+        self.text_encoder_affine_layer = nn.Linear(
+            self.text_encoder.output_size(),
+            llm_input_size
+        )
+
+        # 2. build speech token language model related modules
+        self.sos_eos = 0
+        self.task_id = 1
+        self.llm_embedding = torch.nn.Embedding(2, llm_input_size)
+        self.llm = llm
+        self.llm_decoder = nn.Linear(llm_output_size, speech_token_size + 1)
+        self.criterion_ce = LabelSmoothingLoss(
+            size=speech_token_size + 1,
+            padding_idx=IGNORE_ID,
+            smoothing=lsm_weight,
+            normalize_length=length_normalized_loss,
+        )
+
+        # 3. [Optional] build speech token related modules
+        self.speech_embedding = torch.nn.Embedding(speech_token_size, llm_input_size)
+        self.spk_embed_affine_layer = torch.nn.Linear(spk_embed_dim, llm_input_size)
+
+    def encode(
+            self,
+            text: torch.Tensor,
+            text_lengths: torch.Tensor,
+    ):
+        encoder_out, encoder_mask = self.text_encoder(text, text_lengths, decoding_chunk_size=1, num_decoding_left_chunks=-1)
+        encoder_out_lens = encoder_mask.squeeze(1).sum(1)
+        encoder_out = self.text_encoder_affine_layer(encoder_out)
+        return encoder_out, encoder_out_lens
+
+    def pad_unpad_sequence(self, sos_eos_emb, embedding, text_token, text_token_len, task_id_emb, speech_token, speech_token_len):
+        text_token = unpad_sequence(text_token, text_token_len.cpu(), batch_first=True)
+        speech_token = unpad_sequence(speech_token, speech_token_len.cpu(), batch_first=True)
+        lm_input = [torch.concat([sos_eos_emb.squeeze(dim=0), embedding[i], text_token[i], task_id_emb.squeeze(dim=0), speech_token[i]], dim=0) for i in range(len(text_token))]
+        lm_input_len = torch.tensor([i.size(0) for i in lm_input], dtype=torch.int32)
+        lm_input = pad_sequence(lm_input, batch_first=True, padding_value=IGNORE_ID)
+        return lm_input, lm_input_len
+
+    def forward(
+            self,
+            batch: dict,
+            device: torch.device,
+    ) -> Dict[str, Optional[torch.Tensor]]:
+        """
+        Args:
+            text: (B, L, D)
+            text_lengths: (B,)
+            audio: (B, T, N) or (B, T)
+            audio_lengths: (B,)
+        """
+        text_token = batch['text_token'].to(device)
+        text_token_len = batch['text_token_len'].to(device)
+        speech_token = batch['speech_token'].to(device)
+        speech_token_len = batch['speech_token_len'].to(device)
+        embedding = batch['embedding'].to(device)
+
+        # 1. prepare llm_target
+        lm_target = [torch.tensor([IGNORE_ID] * (2 + text_token_len[i]) + speech_token[i, :speech_token_len[i]].tolist() + [self.speech_token_size]) for i in range(text_token.size(0))]
+        lm_target = pad_sequence(lm_target, batch_first=True, padding_value=IGNORE_ID).to(device)
+
+        # 1. encode text_token
+        text_token = self.text_embedding(text_token)
+        text_token, text_token_len = self.encode(text_token, text_token_len)
+
+        # 2. embedding projection
+        embedding = F.normalize(embedding, dim=1)
+        embedding = self.spk_embed_affine_layer(embedding)
+        embedding = embedding.unsqueeze(1)
+
+        # 3. eos and task_id
+        sos_eos_emb = self.llm_embedding.weight[self.sos_eos].reshape(1, 1, -1)
+        task_id_emb = self.llm_embedding.weight[self.task_id].reshape(1, 1, -1)
+
+        # 4. encode speech_token
+        speech_token = self.speech_embedding(speech_token)
+
+        # 5. unpad and pad
+        lm_input, lm_input_len = self.pad_unpad_sequence(sos_eos_emb, embedding, text_token, text_token_len, task_id_emb, speech_token, speech_token_len)
+
+        # 6. run lm forward
+        lm_output, lm_output_mask = self.llm(lm_input, lm_input_len.to(device))
+        logits = self.llm_decoder(lm_output)
+        loss = self.criterion_ce(logits, lm_target)
+        acc = th_accuracy(logits.view(-1, self.speech_token_size + 1), lm_target, ignore_label=IGNORE_ID)
+        return {'loss': loss, 'acc': acc}
+
+    def sampling_ids(
+            self,
+            weighted_scores: torch.Tensor,
+            sampling: Union[bool, int, float] = True,
+            beam_size: int = 1,
+            ignore_eos: bool = True,
+    ):
+        while True:
+            prob, indices = weighted_scores.softmax(dim=-1).topk(sampling)
+            top_ids = prob.multinomial(beam_size, replacement=True)
+            top_ids = indices[top_ids]
+            if (not ignore_eos) or (self.speech_token_size not in top_ids):
+                break
+        return top_ids
+
+    @torch.inference_mode()
+    def inference(
+            self,
+            text: torch.Tensor,
+            text_len: torch.Tensor,
+            prompt_text: torch.Tensor,
+            prompt_text_len: torch.Tensor,
+            prompt_speech_token: torch.Tensor,
+            prompt_speech_token_len: torch.Tensor,
+            embedding: torch.Tensor,
+            beam_size: int = 1,
+            sampling: int = 25,
+            max_token_text_ratio: float = 20,
+            min_token_text_ratio: float = 2,
+    ) -> torch.Tensor:
+        device = text.device
+        text = torch.concat([prompt_text, text], dim=1)
+        text_len += prompt_text_len
+        text = self.text_embedding(text)
+
+        # 1. encode text
+        text, text_len = self.encode(text, text_len)
+
+        # 2. encode embedding
+        if embedding.shape[0] != 0:
+            embedding = F.normalize(embedding, dim=1)
+            embedding = self.spk_embed_affine_layer(embedding)
+            embedding = embedding.unsqueeze(dim=1)
+        else:
+            embedding = torch.zeros(1, 0, self.llm_input_size).to(device)
+
+        # 3. concat llm_input
+        sos_eos_emb = self.llm_embedding.weight[self.sos_eos].reshape(1, 1, -1)
+        task_id_emb = self.llm_embedding.weight[self.task_id].reshape(1, 1, -1)
+        if prompt_speech_token_len != 0:
+            prompt_speech_token_emb = self.speech_embedding(prompt_speech_token)
+        else:
+            prompt_speech_token_emb = torch.zeros(1, 0, self.llm_input_size).to(device)
+        lm_input = torch.concat([sos_eos_emb, embedding, text, task_id_emb, prompt_speech_token_emb], dim=1)
+
+        # 4. cal min/max_length
+        min_len = int((text_len - prompt_text_len) * min_token_text_ratio)
+        max_len = int((text_len - prompt_text_len) * max_token_text_ratio)
+
+        # 5. step by step decode
+        out_tokens = []
+        offset = 0
+        att_cache, cnn_cache = torch.zeros((0, 0, 0, 0), device=lm_input.device), torch.zeros((0, 0, 0, 0), device=lm_input.device)
+        for i in range(max_len):
+            y_pred, att_cache, cnn_cache = self.llm.forward_chunk(lm_input, offset=0, required_cache_size=-1, att_cache=att_cache, cnn_cache=cnn_cache,
+                                                                  att_mask=torch.tril(torch.ones((1, lm_input.shape[1], lm_input.shape[1]), device=lm_input.device)).to(torch.bool))
+            logp = self.llm_decoder(y_pred[:, -1]).log_softmax(dim=-1)
+            top_ids = self.sampling_ids(logp.squeeze(dim=0), sampling, beam_size, ignore_eos=True if i < min_len else False).item()
+            if top_ids == self.speech_token_size:
+                break
+            out_tokens.append(top_ids)
+            offset += lm_input.size(1)
+            lm_input = self.speech_embedding.weight[top_ids].reshape(1, 1, -1)
+
+        return torch.tensor([out_tokens], dtype=torch.int64, device=device)

cosyvoice/transformer/activation.py

+# Copyright (c) 2020 Johns Hopkins University (Shinji Watanabe)
+#               2020 Northwestern Polytechnical University (Pengcheng Guo)
+#               2020 Mobvoi Inc (Binbin Zhang)
+#               2024 Alibaba Inc (Xiang Lyu)
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#   http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Swish() activation function for Conformer."""
+
+import torch
+from torch import nn, sin, pow
+from torch.nn import Parameter
+
+
+class Swish(torch.nn.Module):
+    """Construct an Swish object."""
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """Return Swish activation function."""
+        return x * torch.sigmoid(x)
+
+
+# Implementation adapted from https://github.com/EdwardDixon/snake under the MIT license.
+#   LICENSE is in incl_licenses directory.
+class Snake(nn.Module):
+    '''
+    Implementation of a sine-based periodic activation function
+    Shape:
+        - Input: (B, C, T)
+        - Output: (B, C, T), same shape as the input
+    Parameters:
+        - alpha - trainable parameter
+    References:
+        - This activation function is from this paper by Liu Ziyin, Tilman Hartwig, Masahito Ueda:
+        https://arxiv.org/abs/2006.08195
+    Examples:
+        >>> a1 = snake(256)
+        >>> x = torch.randn(256)
+        >>> x = a1(x)
+    '''
+    def __init__(self, in_features, alpha=1.0, alpha_trainable=True, alpha_logscale=False):
+        '''
+        Initialization.
+        INPUT:
+            - in_features: shape of the input
+            - alpha: trainable parameter
+            alpha is initialized to 1 by default, higher values = higher-frequency.
+            alpha will be trained along with the rest of your model.
+        '''
+        super(Snake, self).__init__()
+        self.in_features = in_features
+
+        # initialize alpha
+        self.alpha_logscale = alpha_logscale
+        if self.alpha_logscale:  # log scale alphas initialized to zeros
+            self.alpha = Parameter(torch.zeros(in_features) * alpha)
+        else:  # linear scale alphas initialized to ones
+            self.alpha = Parameter(torch.ones(in_features) * alpha)
+
+        self.alpha.requires_grad = alpha_trainable
+
+        self.no_div_by_zero = 0.000000001
+
+    def forward(self, x):
+        '''
+        Forward pass of the function.
+        Applies the function to the input elementwise.
+        Snake ∶= x + 1/a * sin^2 (xa)
+        '''
+        alpha = self.alpha.unsqueeze(0).unsqueeze(-1)  # line up with x to [B, C, T]
+        if self.alpha_logscale:
+            alpha = torch.exp(alpha)
+        x = x + (1.0 / (alpha + self.no_div_by_zero)) * pow(sin(x * alpha), 2)
+
+        return x

cosyvoice/transformer/attention.py

+# Copyright (c) 2019 Shigeki Karita
+#               2020 Mobvoi Inc (Binbin Zhang)
+#               2022 Xingchen Song (sxc19@mails.tsinghua.edu.cn)
+#               2024 Alibaba Inc (Xiang Lyu)
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#   http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Multi-Head Attention layer definition."""
+
+import math
+from typing import Tuple
+
+import torch
+from torch import nn
+
+
+class MultiHeadedAttention(nn.Module):
+    """Multi-Head Attention layer.
+
+    Args:
+        n_head (int): The number of heads.
+        n_feat (int): The number of features.
+        dropout_rate (float): Dropout rate.
+
+    """
+
+    def __init__(self,
+                 n_head: int,
+                 n_feat: int,
+                 dropout_rate: float,
+                 key_bias: bool = True):
+        """Construct an MultiHeadedAttention object."""
+        super().__init__()
+        assert n_feat % n_head == 0
+        # We assume d_v always equals d_k
+        self.d_k = n_feat // n_head
+        self.h = n_head
+        self.linear_q = nn.Linear(n_feat, n_feat)
+        self.linear_k = nn.Linear(n_feat, n_feat, bias=key_bias)
+        self.linear_v = nn.Linear(n_feat, n_feat)
+        self.linear_out = nn.Linear(n_feat, n_feat)
+        self.dropout = nn.Dropout(p=dropout_rate)
+
+    def forward_qkv(
+        self, query: torch.Tensor, key: torch.Tensor, value: torch.Tensor
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        """Transform query, key and value.
+
+        Args:
+            query (torch.Tensor): Query tensor (#batch, time1, size).
+            key (torch.Tensor): Key tensor (#batch, time2, size).
+            value (torch.Tensor): Value tensor (#batch, time2, size).
+
+        Returns:
+            torch.Tensor: Transformed query tensor, size
+                (#batch, n_head, time1, d_k).
+            torch.Tensor: Transformed key tensor, size
+                (#batch, n_head, time2, d_k).
+            torch.Tensor: Transformed value tensor, size
+                (#batch, n_head, time2, d_k).
+
+        """
+        n_batch = query.size(0)
+        q = self.linear_q(query).view(n_batch, -1, self.h, self.d_k)
+        k = self.linear_k(key).view(n_batch, -1, self.h, self.d_k)
+        v = self.linear_v(value).view(n_batch, -1, self.h, self.d_k)
+        q = q.transpose(1, 2)  # (batch, head, time1, d_k)
+        k = k.transpose(1, 2)  # (batch, head, time2, d_k)
+        v = v.transpose(1, 2)  # (batch, head, time2, d_k)
+
+        return q, k, v
+
+    def forward_attention(
+        self,
+        value: torch.Tensor,
+        scores: torch.Tensor,
+        mask: torch.Tensor = torch.ones((0, 0, 0), dtype=torch.bool)
+    ) -> torch.Tensor:
+        """Compute attention context vector.
+
+        Args:
+            value (torch.Tensor): Transformed value, size
+                (#batch, n_head, time2, d_k).
+            scores (torch.Tensor): Attention score, size
+                (#batch, n_head, time1, time2).
+            mask (torch.Tensor): Mask, size (#batch, 1, time2) or
+                (#batch, time1, time2), (0, 0, 0) means fake mask.
+
+        Returns:
+            torch.Tensor: Transformed value (#batch, time1, d_model)
+                weighted by the attention score (#batch, time1, time2).
+
+        """
+        n_batch = value.size(0)
+        # NOTE(xcsong): When will `if mask.size(2) > 0` be True?
+        #   1. onnx(16/4) [WHY? Because we feed real cache & real mask for the
+        #           1st chunk to ease the onnx export.]
+        #   2. pytorch training
+        if mask.size(2) > 0:  # time2 > 0
+            mask = mask.unsqueeze(1).eq(0)  # (batch, 1, *, time2)
+            # For last chunk, time2 might be larger than scores.size(-1)
+            mask = mask[:, :, :, :scores.size(-1)]  # (batch, 1, *, time2)
+            scores = scores.masked_fill(mask, -float('inf'))
+            attn = torch.softmax(scores, dim=-1).masked_fill(
+                mask, 0.0)  # (batch, head, time1, time2)
+        # NOTE(xcsong): When will `if mask.size(2) > 0` be False?
+        #   1. onnx(16/-1, -1/-1, 16/0)
+        #   2. jit (16/-1, -1/-1, 16/0, 16/4)
+        else:
+            attn = torch.softmax(scores, dim=-1)  # (batch, head, time1, time2)
+
+        p_attn = self.dropout(attn)
+        x = torch.matmul(p_attn, value)  # (batch, head, time1, d_k)
+        x = (x.transpose(1, 2).contiguous().view(n_batch, -1,
+                                                 self.h * self.d_k)
+             )  # (batch, time1, d_model)
+
+        return self.linear_out(x)  # (batch, time1, d_model)
+
+    def forward(
+        self,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        mask: torch.Tensor = torch.ones((0, 0, 0), dtype=torch.bool),
+        pos_emb: torch.Tensor = torch.empty(0),
+        cache: torch.Tensor = torch.zeros((0, 0, 0, 0))
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Compute scaled dot product attention.
+
+        Args:
+            query (torch.Tensor): Query tensor (#batch, time1, size).
+            key (torch.Tensor): Key tensor (#batch, time2, size).
+            value (torch.Tensor): Value tensor (#batch, time2, size).
+            mask (torch.Tensor): Mask tensor (#batch, 1, time2) or
+                (#batch, time1, time2).
+                1.When applying cross attention between decoder and encoder,
+                the batch padding mask for input is in (#batch, 1, T) shape.
+                2.When applying self attention of encoder,
+                the mask is in (#batch, T, T)  shape.
+                3.When applying self attention of decoder,
+                the mask is in (#batch, L, L)  shape.
+                4.If the different position in decoder see different block
+                of the encoder, such as Mocha, the passed in mask could be
+                in (#batch, L, T) shape. But there is no such case in current
+                CosyVoice.
+            cache (torch.Tensor): Cache tensor (1, head, cache_t, d_k * 2),
+                where `cache_t == chunk_size * num_decoding_left_chunks`
+                and `head * d_k == size`
+
+
+        Returns:
+            torch.Tensor: Output tensor (#batch, time1, d_model).
+            torch.Tensor: Cache tensor (1, head, cache_t + time1, d_k * 2)
+                where `cache_t == chunk_size * num_decoding_left_chunks`
+                and `head * d_k == size`
+
+        """
+        q, k, v = self.forward_qkv(query, key, value)
+
+        # NOTE(xcsong):
+        #   when export onnx model, for 1st chunk, we feed
+        #       cache(1, head, 0, d_k * 2) (16/-1, -1/-1, 16/0 mode)
+        #       or cache(1, head, real_cache_t, d_k * 2) (16/4 mode).
+        #       In all modes, `if cache.size(0) > 0` will alwayse be `True`
+        #       and we will always do splitting and
+        #       concatnation(this will simplify onnx export). Note that
+        #       it's OK to concat & split zero-shaped tensors(see code below).
+        #   when export jit  model, for 1st chunk, we always feed
+        #       cache(0, 0, 0, 0) since jit supports dynamic if-branch.
+        # >>> a = torch.ones((1, 2, 0, 4))
+        # >>> b = torch.ones((1, 2, 3, 4))
+        # >>> c = torch.cat((a, b), dim=2)
+        # >>> torch.equal(b, c)        # True
+        # >>> d = torch.split(a, 2, dim=-1)
+        # >>> torch.equal(d[0], d[1])  # True
+        if cache.size(0) > 0:
+            key_cache, value_cache = torch.split(cache,
+                                                 cache.size(-1) // 2,
+                                                 dim=-1)
+            k = torch.cat([key_cache, k], dim=2)
+            v = torch.cat([value_cache, v], dim=2)
+        # NOTE(xcsong): We do cache slicing in encoder.forward_chunk, since it's
+        #   non-trivial to calculate `next_cache_start` here.
+        new_cache = torch.cat((k, v), dim=-1)
+
+        scores = torch.matmul(q, k.transpose(-2, -1)) / math.sqrt(self.d_k)
+        return self.forward_attention(v, scores, mask), new_cache
+
+
+class RelPositionMultiHeadedAttention(MultiHeadedAttention):
+    """Multi-Head Attention layer with relative position encoding.
+    Paper: https://arxiv.org/abs/1901.02860
+    Args:
+        n_head (int): The number of heads.
+        n_feat (int): The number of features.
+        dropout_rate (float): Dropout rate.
+    """
+
+    def __init__(self,
+                 n_head: int,
+                 n_feat: int,
+                 dropout_rate: float,
+                 key_bias: bool = True):
+        """Construct an RelPositionMultiHeadedAttention object."""
+        super().__init__(n_head, n_feat, dropout_rate, key_bias)
+        # linear transformation for positional encoding
+        self.linear_pos = nn.Linear(n_feat, n_feat, bias=False)
+        # these two learnable bias are used in matrix c and matrix d
+        # as described in https://arxiv.org/abs/1901.02860 Section 3.3
+        self.pos_bias_u = nn.Parameter(torch.Tensor(self.h, self.d_k))
+        self.pos_bias_v = nn.Parameter(torch.Tensor(self.h, self.d_k))
+        torch.nn.init.xavier_uniform_(self.pos_bias_u)
+        torch.nn.init.xavier_uniform_(self.pos_bias_v)
+
+    def rel_shift(self, x):
+        """Compute relative positional encoding.
+
+        Args:
+            x (torch.Tensor): Input tensor (batch, head, time1, 2*time1-1).
+            time1 means the length of query vector.
+
+        Returns:
+            torch.Tensor: Output tensor.
+
+        """
+        zero_pad = torch.zeros((*x.size()[:3], 1), device=x.device, dtype=x.dtype)
+        x_padded = torch.cat([zero_pad, x], dim=-1)
+
+        x_padded = x_padded.view(*x.size()[:2], x.size(3) + 1, x.size(2))
+        x = x_padded[:, :, 1:].view_as(x)[
+            :, :, :, : x.size(-1) // 2 + 1
+        ]  # only keep the positions from 0 to time2
+        return x
+
+    def forward(
+        self,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        mask: torch.Tensor = torch.ones((0, 0, 0), dtype=torch.bool),
+        pos_emb: torch.Tensor = torch.empty(0),
+        cache: torch.Tensor = torch.zeros((0, 0, 0, 0))
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Compute 'Scaled Dot Product Attention' with rel. positional encoding.
+        Args:
+            query (torch.Tensor): Query tensor (#batch, time1, size).
+            key (torch.Tensor): Key tensor (#batch, time2, size).
+            value (torch.Tensor): Value tensor (#batch, time2, size).
+            mask (torch.Tensor): Mask tensor (#batch, 1, time2) or
+                (#batch, time1, time2), (0, 0, 0) means fake mask.
+            pos_emb (torch.Tensor): Positional embedding tensor
+                (#batch, time2, size).
+            cache (torch.Tensor): Cache tensor (1, head, cache_t, d_k * 2),
+                where `cache_t == chunk_size * num_decoding_left_chunks`
+                and `head * d_k == size`
+        Returns:
+            torch.Tensor: Output tensor (#batch, time1, d_model).
+            torch.Tensor: Cache tensor (1, head, cache_t + time1, d_k * 2)
+                where `cache_t == chunk_size * num_decoding_left_chunks`
+                and `head * d_k == size`
+        """
+        q, k, v = self.forward_qkv(query, key, value)
+        q = q.transpose(1, 2)  # (batch, time1, head, d_k)
+
+        # NOTE(xcsong):
+        #   when export onnx model, for 1st chunk, we feed
+        #       cache(1, head, 0, d_k * 2) (16/-1, -1/-1, 16/0 mode)
+        #       or cache(1, head, real_cache_t, d_k * 2) (16/4 mode).
+        #       In all modes, `if cache.size(0) > 0` will alwayse be `True`
+        #       and we will always do splitting and
+        #       concatnation(this will simplify onnx export). Note that
+        #       it's OK to concat & split zero-shaped tensors(see code below).
+        #   when export jit  model, for 1st chunk, we always feed
+        #       cache(0, 0, 0, 0) since jit supports dynamic if-branch.
+        # >>> a = torch.ones((1, 2, 0, 4))
+        # >>> b = torch.ones((1, 2, 3, 4))
+        # >>> c = torch.cat((a, b), dim=2)
+        # >>> torch.equal(b, c)        # True
+        # >>> d = torch.split(a, 2, dim=-1)
+        # >>> torch.equal(d[0], d[1])  # True
+        if cache.size(0) > 0:
+            key_cache, value_cache = torch.split(cache,
+                                                 cache.size(-1) // 2,
+                                                 dim=-1)
+            k = torch.cat([key_cache, k], dim=2)
+            v = torch.cat([value_cache, v], dim=2)
+        # NOTE(xcsong): We do cache slicing in encoder.forward_chunk, since it's
+        #   non-trivial to calculate `next_cache_start` here.
+        new_cache = torch.cat((k, v), dim=-1)
+
+        n_batch_pos = pos_emb.size(0)
+        p = self.linear_pos(pos_emb).view(n_batch_pos, -1, self.h, self.d_k)
+        p = p.transpose(1, 2)  # (batch, head, time1, d_k)
+
+        # (batch, head, time1, d_k)
+        q_with_bias_u = (q + self.pos_bias_u).transpose(1, 2)
+        # (batch, head, time1, d_k)
+        q_with_bias_v = (q + self.pos_bias_v).transpose(1, 2)
+
+        # compute attention score
+        # first compute matrix a and matrix c
+        # as described in https://arxiv.org/abs/1901.02860 Section 3.3
+        # (batch, head, time1, time2)
+        matrix_ac = torch.matmul(q_with_bias_u, k.transpose(-2, -1))
+
+        # compute matrix b and matrix d
+        # (batch, head, time1, time2)
+        matrix_bd = torch.matmul(q_with_bias_v, p.transpose(-2, -1))
+        # NOTE(Xiang Lyu): Keep rel_shift since espnet rel_pos_emb is used
+        if matrix_ac.shape != matrix_bd.shape:
+            matrix_bd = self.rel_shift(matrix_bd)
+
+        scores = (matrix_ac + matrix_bd) / math.sqrt(
+            self.d_k)  # (batch, head, time1, time2)
+
+        return self.forward_attention(v, scores, mask), new_cache

cosyvoice/transformer/convolution.py

+# Copyright (c) 2020 Mobvoi Inc. (authors: Binbin Zhang, Di Wu)
+#               2024 Alibaba Inc (Xiang Lyu)
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# Modified from ESPnet(https://github.com/espnet/espnet)
+"""ConvolutionModule definition."""
+
+from typing import Tuple
+
+import torch
+from torch import nn
+
+
+class ConvolutionModule(nn.Module):
+    """ConvolutionModule in Conformer model."""
+
+    def __init__(self,
+                 channels: int,
+                 kernel_size: int = 15,
+                 activation: nn.Module = nn.ReLU(),
+                 norm: str = "batch_norm",
+                 causal: bool = False,
+                 bias: bool = True):
+        """Construct an ConvolutionModule object.
+        Args:
+            channels (int): The number of channels of conv layers.
+            kernel_size (int): Kernel size of conv layers.
+            causal (int): Whether use causal convolution or not
+        """
+        super().__init__()
+
+        self.pointwise_conv1 = nn.Conv1d(
+            channels,
+            2 * channels,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            bias=bias,
+        )
+        # self.lorder is used to distinguish if it's a causal convolution,
+        # if self.lorder > 0: it's a causal convolution, the input will be
+        #    padded with self.lorder frames on the left in forward.
+        # else: it's a symmetrical convolution
+        if causal:
+            padding = 0
+            self.lorder = kernel_size - 1
+        else:
+            # kernel_size should be an odd number for none causal convolution
+            assert (kernel_size - 1) % 2 == 0
+            padding = (kernel_size - 1) // 2
+            self.lorder = 0
+        self.depthwise_conv = nn.Conv1d(
+            channels,
+            channels,
+            kernel_size,
+            stride=1,
+            padding=padding,
+            groups=channels,
+            bias=bias,
+        )
+
+        assert norm in ['batch_norm', 'layer_norm']
+        if norm == "batch_norm":
+            self.use_layer_norm = False
+            self.norm = nn.BatchNorm1d(channels)
+        else:
+            self.use_layer_norm = True
+            self.norm = nn.LayerNorm(channels)
+
+        self.pointwise_conv2 = nn.Conv1d(
+            channels,
+            channels,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            bias=bias,
+        )
+        self.activation = activation
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        mask_pad: torch.Tensor = torch.ones((0, 0, 0), dtype=torch.bool),
+        cache: torch.Tensor = torch.zeros((0, 0, 0)),
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Compute convolution module.
+        Args:
+            x (torch.Tensor): Input tensor (#batch, time, channels).
+            mask_pad (torch.Tensor): used for batch padding (#batch, 1, time),
+                (0, 0, 0) means fake mask.
+            cache (torch.Tensor): left context cache, it is only
+                used in causal convolution (#batch, channels, cache_t),
+                (0, 0, 0) meas fake cache.
+        Returns:
+            torch.Tensor: Output tensor (#batch, time, channels).
+        """
+        # exchange the temporal dimension and the feature dimension
+        x = x.transpose(1, 2)  # (#batch, channels, time)
+
+        # mask batch padding
+        if mask_pad.size(2) > 0:  # time > 0
+            x.masked_fill_(~mask_pad, 0.0)
+
+        if self.lorder > 0:
+            if cache.size(2) == 0:  # cache_t == 0
+                x = nn.functional.pad(x, (self.lorder, 0), 'constant', 0.0)
+            else:
+                assert cache.size(0) == x.size(0)  # equal batch
+                assert cache.size(1) == x.size(1)  # equal channel
+                x = torch.cat((cache, x), dim=2)
+            assert (x.size(2) > self.lorder)
+            new_cache = x[:, :, -self.lorder:]
+        else:
+            # It's better we just return None if no cache is required,
+            # However, for JIT export, here we just fake one tensor instead of
+            # None.
+            new_cache = torch.zeros((0, 0, 0), dtype=x.dtype, device=x.device)
+
+        # GLU mechanism
+        x = self.pointwise_conv1(x)  # (batch, 2*channel, dim)
+        x = nn.functional.glu(x, dim=1)  # (batch, channel, dim)
+
+        # 1D Depthwise Conv
+        x = self.depthwise_conv(x)
+        if self.use_layer_norm:
+            x = x.transpose(1, 2)
+        x = self.activation(self.norm(x))
+        if self.use_layer_norm:
+            x = x.transpose(1, 2)
+        x = self.pointwise_conv2(x)
+        # mask batch padding
+        if mask_pad.size(2) > 0:  # time > 0
+            x.masked_fill_(~mask_pad, 0.0)
+
+        return x.transpose(1, 2), new_cache

cosyvoice/transformer/decoder.py

+# Copyright (c) 2021 Mobvoi Inc. (authors: Binbin Zhang, Di Wu)
+#               2024 Alibaba Inc (Xiang Lyu)
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# Modified from ESPnet(https://github.com/espnet/espnet)
+"""Decoder definition."""
+from typing import Tuple, List, Optional
+
+import torch
+import torch.utils.checkpoint as ckpt
+import logging
+
+from cosyvoice.transformer.decoder_layer import DecoderLayer
+from cosyvoice.transformer.positionwise_feed_forward import PositionwiseFeedForward
+from cosyvoice.utils.class_utils import (
+    COSYVOICE_EMB_CLASSES,
+    COSYVOICE_ATTENTION_CLASSES,
+    COSYVOICE_ACTIVATION_CLASSES,
+)
+from cosyvoice.utils.mask import (subsequent_mask, make_pad_mask)
+
+
+class TransformerDecoder(torch.nn.Module):
+    """Base class of Transfomer decoder module.
+    Args:
+        vocab_size: output dim
+        encoder_output_size: dimension of attention
+        attention_heads: the number of heads of multi head attention
+        linear_units: the hidden units number of position-wise feedforward
+        num_blocks: the number of decoder blocks
+        dropout_rate: dropout rate
+        self_attention_dropout_rate: dropout rate for attention
+        input_layer: input layer type
+        use_output_layer: whether to use output layer
+        pos_enc_class: PositionalEncoding or ScaledPositionalEncoding
+        normalize_before:
+            True: use layer_norm before each sub-block of a layer.
+            False: use layer_norm after each sub-block of a layer.
+        src_attention: if false, encoder-decoder cross attention is not
+                       applied, such as CIF model
+        key_bias: whether use bias in attention.linear_k, False for whisper models.
+        gradient_checkpointing: rerunning a forward-pass segment for each
+            checkpointed segment during backward.
+        tie_word_embedding: Tie or clone module weights depending of whether we are
+            using TorchScript or not
+    """
+
+    def __init__(
+        self,
+        vocab_size: int,
+        encoder_output_size: int,
+        attention_heads: int = 4,
+        linear_units: int = 2048,
+        num_blocks: int = 6,
+        dropout_rate: float = 0.1,
+        positional_dropout_rate: float = 0.1,
+        self_attention_dropout_rate: float = 0.0,
+        src_attention_dropout_rate: float = 0.0,
+        input_layer: str = "embed",
+        use_output_layer: bool = True,
+        normalize_before: bool = True,
+        src_attention: bool = True,
+        key_bias: bool = True,
+        activation_type: str = "relu",
+        gradient_checkpointing: bool = False,
+        tie_word_embedding: bool = False,
+    ):
+        super().__init__()
+        attention_dim = encoder_output_size
+        activation = COSYVOICE_ACTIVATION_CLASSES[activation_type]()
+
+        self.embed = torch.nn.Sequential(
+            torch.nn.Identity() if input_layer == "no_pos" else
+            torch.nn.Embedding(vocab_size, attention_dim),
+            COSYVOICE_EMB_CLASSES[input_layer](attention_dim,
+                                               positional_dropout_rate),
+        )
+
+        self.normalize_before = normalize_before
+        self.after_norm = torch.nn.LayerNorm(attention_dim, eps=1e-5)
+        self.use_output_layer = use_output_layer
+        if use_output_layer:
+            self.output_layer = torch.nn.Linear(attention_dim, vocab_size)
+        else:
+            self.output_layer = torch.nn.Identity()
+        self.num_blocks = num_blocks
+        self.decoders = torch.nn.ModuleList([
+            DecoderLayer(
+                attention_dim,
+                COSYVOICE_ATTENTION_CLASSES["selfattn"](
+                    attention_heads, attention_dim,
+                    self_attention_dropout_rate, key_bias),
+                COSYVOICE_ATTENTION_CLASSES["selfattn"](
+                    attention_heads, attention_dim, src_attention_dropout_rate,
+                    key_bias) if src_attention else None,
+                PositionwiseFeedForward(attention_dim, linear_units,
+                                        dropout_rate, activation),
+                dropout_rate,
+                normalize_before,
+            ) for _ in range(self.num_blocks)
+        ])
+
+        self.gradient_checkpointing = gradient_checkpointing
+        self.tie_word_embedding = tie_word_embedding
+
+    def forward(
+        self,
+        memory: torch.Tensor,
+        memory_mask: torch.Tensor,
+        ys_in_pad: torch.Tensor,
+        ys_in_lens: torch.Tensor,
+        r_ys_in_pad: torch.Tensor = torch.empty(0),
+        reverse_weight: float = 0.0,
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        """Forward decoder.
+        Args:
+            memory: encoded memory, float32  (batch, maxlen_in, feat)
+            memory_mask: encoder memory mask, (batch, 1, maxlen_in)
+            ys_in_pad: padded input token ids, int64 (batch, maxlen_out)
+            ys_in_lens: input lengths of this batch (batch)
+            r_ys_in_pad: not used in transformer decoder, in order to unify api
+                with bidirectional decoder
+            reverse_weight: not used in transformer decoder, in order to unify
+                api with bidirectional decode
+        Returns:
+            (tuple): tuple containing:
+                x: decoded token score before softmax (batch, maxlen_out,
+                    vocab_size) if use_output_layer is True,
+                torch.tensor(0.0), in order to unify api with bidirectional decoder
+                olens: (batch, )
+        NOTE(xcsong):
+            We pass the `__call__` method of the modules instead of `forward` to the
+            checkpointing API because `__call__` attaches all the hooks of the module.
+            https://discuss.pytorch.org/t/any-different-between-model-input-and-model-forward-input/3690/2
+        """
+        tgt = ys_in_pad
+        maxlen = tgt.size(1)
+        # tgt_mask: (B, 1, L)
+        tgt_mask = ~make_pad_mask(ys_in_lens, maxlen).unsqueeze(1)
+        tgt_mask = tgt_mask.to(tgt.device)
+        # m: (1, L, L)
+        m = subsequent_mask(tgt_mask.size(-1),
+                            device=tgt_mask.device).unsqueeze(0)
+        # tgt_mask: (B, L, L)
+        tgt_mask = tgt_mask & m
+        x, _ = self.embed(tgt)
+        if self.gradient_checkpointing and self.training:
+            x = self.forward_layers_checkpointed(x, tgt_mask, memory,
+                                                 memory_mask)
+        else:
+            x = self.forward_layers(x, tgt_mask, memory, memory_mask)
+        if self.normalize_before:
+            x = self.after_norm(x)
+        if self.use_output_layer:
+            x = self.output_layer(x)
+        olens = tgt_mask.sum(1)
+        return x, torch.tensor(0.0), olens
+
+    def forward_layers(self, x: torch.Tensor, tgt_mask: torch.Tensor,
+                       memory: torch.Tensor,
+                       memory_mask: torch.Tensor) -> torch.Tensor:
+        for layer in self.decoders:
+            x, tgt_mask, memory, memory_mask = layer(x, tgt_mask, memory,
+                                                     memory_mask)
+        return x
+
+    @torch.jit.ignore(drop=True)
+    def forward_layers_checkpointed(self, x: torch.Tensor,
+                                    tgt_mask: torch.Tensor,
+                                    memory: torch.Tensor,
+                                    memory_mask: torch.Tensor) -> torch.Tensor:
+        for layer in self.decoders:
+            x, tgt_mask, memory, memory_mask = ckpt.checkpoint(
+                layer.__call__, x, tgt_mask, memory, memory_mask)
+        return x
+
+    def forward_one_step(
+        self,
+        memory: torch.Tensor,
+        memory_mask: torch.Tensor,
+        tgt: torch.Tensor,
+        tgt_mask: torch.Tensor,
+        cache: Optional[List[torch.Tensor]] = None,
+    ) -> Tuple[torch.Tensor, List[torch.Tensor]]:
+        """Forward one step.
+            This is only used for decoding.
+        Args:
+            memory: encoded memory, float32  (batch, maxlen_in, feat)
+            memory_mask: encoded memory mask, (batch, 1, maxlen_in)
+            tgt: input token ids, int64 (batch, maxlen_out)
+            tgt_mask: input token mask,  (batch, maxlen_out)
+                      dtype=torch.uint8 in PyTorch 1.2-
+                      dtype=torch.bool in PyTorch 1.2+ (include 1.2)
+            cache: cached output list of (batch, max_time_out-1, size)
+        Returns:
+            y, cache: NN output value and cache per `self.decoders`.
+            y.shape` is (batch, maxlen_out, token)
+        """
+        x, _ = self.embed(tgt)
+        new_cache = []
+        for i, decoder in enumerate(self.decoders):
+            if cache is None:
+                c = None
+            else:
+                c = cache[i]
+            x, tgt_mask, memory, memory_mask = decoder(x,
+                                                       tgt_mask,
+                                                       memory,
+                                                       memory_mask,
+                                                       cache=c)
+            new_cache.append(x)
+        if self.normalize_before:
+            y = self.after_norm(x[:, -1])
+        else:
+            y = x[:, -1]
+        if self.use_output_layer:
+            y = torch.log_softmax(self.output_layer(y), dim=-1)
+        return y, new_cache
+
+    def tie_or_clone_weights(self, jit_mode: bool = True):
+        """Tie or clone module weights (between word_emb and output_layer)
+            depending of whether we are using TorchScript or not"""
+        if not self.use_output_layer:
+            return
+        if jit_mode:
+            logging.info("clone emb.weight to output.weight")
+            self.output_layer.weight = torch.nn.Parameter(
+                self.embed[0].weight.clone())
+        else:
+            logging.info("tie emb.weight with output.weight")
+            self.output_layer.weight = self.embed[0].weight
+
+        if getattr(self.output_layer, "bias", None) is not None:
+            self.output_layer.bias.data = torch.nn.functional.pad(
+                self.output_layer.bias.data,
+                (
+                    0,
+                    self.output_layer.weight.shape[0] -
+                    self.output_layer.bias.shape[0],
+                ),
+                "constant",
+                0,
+            )
+
+
+class BiTransformerDecoder(torch.nn.Module):
+    """Base class of Transfomer decoder module.
+    Args:
+        vocab_size: output dim
+        encoder_output_size: dimension of attention
+        attention_heads: the number of heads of multi head attention
+        linear_units: the hidden units number of position-wise feedforward
+        num_blocks: the number of decoder blocks
+        r_num_blocks: the number of right to left decoder blocks
+        dropout_rate: dropout rate
+        self_attention_dropout_rate: dropout rate for attention
+        input_layer: input layer type
+        use_output_layer: whether to use output layer
+        pos_enc_class: PositionalEncoding or ScaledPositionalEncoding
+        normalize_before:
+            True: use layer_norm before each sub-block of a layer.
+            False: use layer_norm after each sub-block of a layer.
+        key_bias: whether use bias in attention.linear_k, False for whisper models.
+    """
+
+    def __init__(
+        self,
+        vocab_size: int,
+        encoder_output_size: int,
+        attention_heads: int = 4,
+        linear_units: int = 2048,
+        num_blocks: int = 6,
+        r_num_blocks: int = 0,
+        dropout_rate: float = 0.1,
+        positional_dropout_rate: float = 0.1,
+        self_attention_dropout_rate: float = 0.0,
+        src_attention_dropout_rate: float = 0.0,
+        input_layer: str = "embed",
+        use_output_layer: bool = True,
+        normalize_before: bool = True,
+        key_bias: bool = True,
+        gradient_checkpointing: bool = False,
+        tie_word_embedding: bool = False,
+    ):
+
+        super().__init__()
+        self.tie_word_embedding = tie_word_embedding
+        self.left_decoder = TransformerDecoder(
+            vocab_size,
+            encoder_output_size,
+            attention_heads,
+            linear_units,
+            num_blocks,
+            dropout_rate,
+            positional_dropout_rate,
+            self_attention_dropout_rate,
+            src_attention_dropout_rate,
+            input_layer,
+            use_output_layer,
+            normalize_before,
+            key_bias=key_bias,
+            gradient_checkpointing=gradient_checkpointing,
+            tie_word_embedding=tie_word_embedding)
+
+        self.right_decoder = TransformerDecoder(
+            vocab_size,
+            encoder_output_size,
+            attention_heads,
+            linear_units,
+            r_num_blocks,
+            dropout_rate,
+            positional_dropout_rate,
+            self_attention_dropout_rate,
+            src_attention_dropout_rate,
+            input_layer,
+            use_output_layer,
+            normalize_before,
+            key_bias=key_bias,
+            gradient_checkpointing=gradient_checkpointing,
+            tie_word_embedding=tie_word_embedding)
+
+    def forward(
+        self,
+        memory: torch.Tensor,
+        memory_mask: torch.Tensor,
+        ys_in_pad: torch.Tensor,
+        ys_in_lens: torch.Tensor,
+        r_ys_in_pad: torch.Tensor,
+        reverse_weight: float = 0.0,
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        """Forward decoder.
+        Args:
+            memory: encoded memory, float32  (batch, maxlen_in, feat)
+            memory_mask: encoder memory mask, (batch, 1, maxlen_in)
+            ys_in_pad: padded input token ids, int64 (batch, maxlen_out)
+            ys_in_lens: input lengths of this batch (batch)
+            r_ys_in_pad: padded input token ids, int64 (batch, maxlen_out),
+                used for right to left decoder
+            reverse_weight: used for right to left decoder
+        Returns:
+            (tuple): tuple containing:
+                x: decoded token score before softmax (batch, maxlen_out,
+                    vocab_size) if use_output_layer is True,
+                r_x: x: decoded token score (right to left decoder)
+                    before softmax (batch, maxlen_out, vocab_size)
+                    if use_output_layer is True,
+                olens: (batch, )
+        """
+        l_x, _, olens = self.left_decoder(memory, memory_mask, ys_in_pad,
+                                          ys_in_lens)
+        r_x = torch.tensor(0.0)
+        if reverse_weight > 0.0:
+            r_x, _, olens = self.right_decoder(memory, memory_mask,
+                                               r_ys_in_pad, ys_in_lens)
+        return l_x, r_x, olens
+
+    def forward_one_step(
+        self,
+        memory: torch.Tensor,
+        memory_mask: torch.Tensor,
+        tgt: torch.Tensor,
+        tgt_mask: torch.Tensor,
+        cache: Optional[List[torch.Tensor]] = None,
+    ) -> Tuple[torch.Tensor, List[torch.Tensor]]:
+        """Forward one step.
+            This is only used for decoding.
+        Args:
+            memory: encoded memory, float32  (batch, maxlen_in, feat)
+            memory_mask: encoded memory mask, (batch, 1, maxlen_in)
+            tgt: input token ids, int64 (batch, maxlen_out)
+            tgt_mask: input token mask,  (batch, maxlen_out)
+                      dtype=torch.uint8 in PyTorch 1.2-
+                      dtype=torch.bool in PyTorch 1.2+ (include 1.2)
+            cache: cached output list of (batch, max_time_out-1, size)
+        Returns:
+            y, cache: NN output value and cache per `self.decoders`.
+            y.shape` is (batch, maxlen_out, token)
+        """
+        return self.left_decoder.forward_one_step(memory, memory_mask, tgt,
+                                                  tgt_mask, cache)
+
+    def tie_or_clone_weights(self, jit_mode: bool = True):
+        """Tie or clone module weights (between word_emb and output_layer)
+            depending of whether we are using TorchScript or not"""
+        self.left_decoder.tie_or_clone_weights(jit_mode)
+        self.right_decoder.tie_or_clone_weights(jit_mode)

cosyvoice/transformer/decoder_layer.py

+# Copyright (c) 2019 Shigeki Karita
+#               2020 Mobvoi Inc (Binbin Zhang)
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#   http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Decoder self-attention layer definition."""
+from typing import Optional, Tuple
+
+import torch
+from torch import nn
+
+
+class DecoderLayer(nn.Module):
+    """Single decoder layer module.
+
+    Args:
+        size (int): Input dimension.
+        self_attn (torch.nn.Module): Self-attention module instance.
+            `MultiHeadedAttention` instance can be used as the argument.
+        src_attn (torch.nn.Module): Inter-attention module instance.
+            `MultiHeadedAttention` instance can be used as the argument.
+            If `None` is passed, Inter-attention is not used, such as
+            CIF, GPT, and other decoder only model.
+        feed_forward (torch.nn.Module): Feed-forward module instance.
+            `PositionwiseFeedForward` instance can be used as the argument.
+        dropout_rate (float): Dropout rate.
+        normalize_before (bool):
+            True: use layer_norm before each sub-block.
+            False: to use layer_norm after each sub-block.
+    """
+
+    def __init__(
+        self,
+        size: int,
+        self_attn: nn.Module,
+        src_attn: Optional[nn.Module],
+        feed_forward: nn.Module,
+        dropout_rate: float,
+        normalize_before: bool = True,
+    ):
+        """Construct an DecoderLayer object."""
+        super().__init__()
+        self.size = size
+        self.self_attn = self_attn
+        self.src_attn = src_attn
+        self.feed_forward = feed_forward
+        self.norm1 = nn.LayerNorm(size, eps=1e-5)
+        self.norm2 = nn.LayerNorm(size, eps=1e-5)
+        self.norm3 = nn.LayerNorm(size, eps=1e-5)
+        self.dropout = nn.Dropout(dropout_rate)
+        self.normalize_before = normalize_before
+
+    def forward(
+        self,
+        tgt: torch.Tensor,
+        tgt_mask: torch.Tensor,
+        memory: torch.Tensor,
+        memory_mask: torch.Tensor,
+        cache: Optional[torch.Tensor] = None
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
+        """Compute decoded features.
+
+        Args:
+            tgt (torch.Tensor): Input tensor (#batch, maxlen_out, size).
+            tgt_mask (torch.Tensor): Mask for input tensor
+                (#batch, maxlen_out).
+            memory (torch.Tensor): Encoded memory
+                (#batch, maxlen_in, size).
+            memory_mask (torch.Tensor): Encoded memory mask
+                (#batch, maxlen_in).
+            cache (torch.Tensor): cached tensors.
+                (#batch, maxlen_out - 1, size).
+
+        Returns:
+            torch.Tensor: Output tensor (#batch, maxlen_out, size).
+            torch.Tensor: Mask for output tensor (#batch, maxlen_out).
+            torch.Tensor: Encoded memory (#batch, maxlen_in, size).
+            torch.Tensor: Encoded memory mask (#batch, maxlen_in).
+
+        """
+        residual = tgt
+        if self.normalize_before:
+            tgt = self.norm1(tgt)
+
+        if cache is None:
+            tgt_q = tgt
+            tgt_q_mask = tgt_mask
+        else:
+            # compute only the last frame query keeping dim: max_time_out -> 1
+            assert cache.shape == (
+                tgt.shape[0],
+                tgt.shape[1] - 1,
+                self.size,
+            ), "{cache.shape} == {(tgt.shape[0], tgt.shape[1] - 1, self.size)}"
+            tgt_q = tgt[:, -1:, :]
+            residual = residual[:, -1:, :]
+            tgt_q_mask = tgt_mask[:, -1:, :]
+
+        x = residual + self.dropout(
+            self.self_attn(tgt_q, tgt, tgt, tgt_q_mask)[0])
+        if not self.normalize_before:
+            x = self.norm1(x)
+
+        if self.src_attn is not None:
+            residual = x
+            if self.normalize_before:
+                x = self.norm2(x)
+            x = residual + self.dropout(
+                self.src_attn(x, memory, memory, memory_mask)[0])
+            if not self.normalize_before:
+                x = self.norm2(x)
+
+        residual = x
+        if self.normalize_before:
+            x = self.norm3(x)
+        x = residual + self.dropout(self.feed_forward(x))
+        if not self.normalize_before:
+            x = self.norm3(x)
+
+        if cache is not None:
+            x = torch.cat([cache, x], dim=1)
+
+        return x, tgt_mask, memory, memory_mask

cosyvoice/transformer/embedding.py

+# Copyright (c) 2020 Mobvoi Inc. (authors: Binbin Zhang, Di Wu)
+#               2024 Alibaba Inc (Xiang Lyu)
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# Modified from ESPnet(https://github.com/espnet/espnet)
+"""Positonal Encoding Module."""
+
+import math
+from typing import Tuple, Union
+
+import torch
+import torch.nn.functional as F
+import numpy as np
+
+
+class PositionalEncoding(torch.nn.Module):
+    """Positional encoding.
+
+    :param int d_model: embedding dim
+    :param float dropout_rate: dropout rate
+    :param int max_len: maximum input length
+
+    PE(pos, 2i)   = sin(pos/(10000^(2i/dmodel)))
+    PE(pos, 2i+1) = cos(pos/(10000^(2i/dmodel)))
+    """
+
+    def __init__(self,
+                 d_model: int,
+                 dropout_rate: float,
+                 max_len: int = 5000,
+                 reverse: bool = False):
+        """Construct an PositionalEncoding object."""
+        super().__init__()
+        self.d_model = d_model
+        self.xscale = math.sqrt(self.d_model)
+        self.dropout = torch.nn.Dropout(p=dropout_rate)
+        self.max_len = max_len
+
+        self.pe = torch.zeros(self.max_len, self.d_model)
+        position = torch.arange(0, self.max_len,
+                                dtype=torch.float32).unsqueeze(1)
+        div_term = torch.exp(
+            torch.arange(0, self.d_model, 2, dtype=torch.float32) *
+            -(math.log(10000.0) / self.d_model))
+        self.pe[:, 0::2] = torch.sin(position * div_term)
+        self.pe[:, 1::2] = torch.cos(position * div_term)
+        self.pe = self.pe.unsqueeze(0)
+
+    def forward(self,
+                x: torch.Tensor,
+                offset: Union[int, torch.Tensor] = 0) \
+            -> Tuple[torch.Tensor, torch.Tensor]:
+        """Add positional encoding.
+
+        Args:
+            x (torch.Tensor): Input. Its shape is (batch, time, ...)
+            offset (int, torch.tensor): position offset
+
+        Returns:
+            torch.Tensor: Encoded tensor. Its shape is (batch, time, ...)
+            torch.Tensor: for compatibility to RelPositionalEncoding
+        """
+
+        self.pe = self.pe.to(x.device)
+        pos_emb = self.position_encoding(offset, x.size(1), False)
+        x = x * self.xscale + pos_emb
+        return self.dropout(x), self.dropout(pos_emb)
+
+    def position_encoding(self,
+                          offset: Union[int, torch.Tensor],
+                          size: int,
+                          apply_dropout: bool = True) -> torch.Tensor:
+        """ For getting encoding in a streaming fashion
+
+        Attention!!!!!
+        we apply dropout only once at the whole utterance level in a none
+        streaming way, but will call this function several times with
+        increasing input size in a streaming scenario, so the dropout will
+        be applied several times.
+
+        Args:
+            offset (int or torch.tensor): start offset
+            size (int): required size of position encoding
+
+        Returns:
+            torch.Tensor: Corresponding encoding
+        """
+        # How to subscript a Union type:
+        #   https://github.com/pytorch/pytorch/issues/69434
+        if isinstance(offset, int):
+            assert offset + size <= self.max_len
+            pos_emb = self.pe[:, offset:offset + size]
+        elif isinstance(offset, torch.Tensor) and offset.dim() == 0:  # scalar
+            assert offset + size <= self.max_len
+            pos_emb = self.pe[:, offset:offset + size]
+        else:  # for batched streaming decoding on GPU
+            assert torch.max(offset) + size <= self.max_len
+            index = offset.unsqueeze(1) + \
+                torch.arange(0, size).to(offset.device)  # B X T
+            flag = index > 0
+            # remove negative offset
+            index = index * flag
+            pos_emb = F.embedding(index, self.pe[0])  # B X T X d_model
+
+        if apply_dropout:
+            pos_emb = self.dropout(pos_emb)
+        return pos_emb
+
+
+class RelPositionalEncoding(PositionalEncoding):
+    """Relative positional encoding module.
+    See : Appendix B in https://arxiv.org/abs/1901.02860
+    Args:
+        d_model (int): Embedding dimension.
+        dropout_rate (float): Dropout rate.
+        max_len (int): Maximum input length.
+    """
+
+    def __init__(self, d_model: int, dropout_rate: float, max_len: int = 5000):
+        """Initialize class."""
+        super().__init__(d_model, dropout_rate, max_len, reverse=True)
+
+    def forward(self,
+                x: torch.Tensor,
+                offset: Union[int, torch.Tensor] = 0) \
+            -> Tuple[torch.Tensor, torch.Tensor]:
+        """Compute positional encoding.
+        Args:
+            x (torch.Tensor): Input tensor (batch, time, `*`).
+        Returns:
+            torch.Tensor: Encoded tensor (batch, time, `*`).
+            torch.Tensor: Positional embedding tensor (1, time, `*`).
+        """
+        self.pe = self.pe.to(x.device)
+        x = x * self.xscale
+        pos_emb = self.position_encoding(offset, x.size(1), False)
+        return self.dropout(x), self.dropout(pos_emb)
+
+
+class WhisperPositionalEncoding(PositionalEncoding):
+    """ Sinusoids position encoding used in openai-whisper.encoder
+    """
+
+    def __init__(self, d_model: int, dropout_rate: float, max_len: int = 1500):
+        super().__init__(d_model, dropout_rate, max_len)
+        self.xscale = 1.0
+        log_timescale_increment = np.log(10000) / (d_model // 2 - 1)
+        inv_timescales = torch.exp(-log_timescale_increment *
+                                   torch.arange(d_model // 2))
+        scaled_time = torch.arange(max_len)[:, np.newaxis] * \
+            inv_timescales[np.newaxis, :]
+        pe = torch.cat([torch.sin(scaled_time), torch.cos(scaled_time)], dim=1)
+        delattr(self, "pe")
+        self.register_buffer("pe", pe.unsqueeze(0))
+
+
+class LearnablePositionalEncoding(PositionalEncoding):
+    """ Learnable position encoding used in openai-whisper.decoder
+    """
+
+    def __init__(self, d_model: int, dropout_rate: float, max_len: int = 448):
+        super().__init__(d_model, dropout_rate, max_len)
+        # NOTE(xcsong): overwrite self.pe & self.xscale
+        self.pe = torch.nn.Parameter(torch.empty(1, max_len, d_model))
+        self.xscale = 1.0
+
+
+class NoPositionalEncoding(torch.nn.Module):
+    """ No position encoding
+    """
+
+    def __init__(self, d_model: int, dropout_rate: float):
+        super().__init__()
+        self.d_model = d_model
+        self.dropout = torch.nn.Dropout(p=dropout_rate)
+
+    def forward(self,
+                x: torch.Tensor,
+                offset: Union[int, torch.Tensor] = 0) \
+            -> Tuple[torch.Tensor, torch.Tensor]:
+        """ Just return zero vector for interface compatibility
+        """
+        pos_emb = torch.zeros(1, x.size(1), self.d_model).to(x.device)
+        return self.dropout(x), pos_emb
+
+    def position_encoding(self, offset: Union[int, torch.Tensor],
+                          size: int) -> torch.Tensor:
+        return torch.zeros(1, size, self.d_model)
+
+
+class EspnetRelPositionalEncoding(torch.nn.Module):
+    """Relative positional encoding module (new implementation).
+
+    Details can be found in https://github.com/espnet/espnet/pull/2816.
+
+    See : Appendix B in https://arxiv.org/abs/1901.02860
+
+    Args:
+        d_model (int): Embedding dimension.
+        dropout_rate (float): Dropout rate.
+        max_len (int): Maximum input length.
+
+    """
+
+    def __init__(self, d_model, dropout_rate, max_len=5000):
+        """Construct an PositionalEncoding object."""
+        super(EspnetRelPositionalEncoding, self).__init__()
+        self.d_model = d_model
+        self.xscale = math.sqrt(self.d_model)
+        self.dropout = torch.nn.Dropout(p=dropout_rate)
+        self.pe = None
+        self.extend_pe(torch.tensor(0.0).expand(1, max_len))
+
+    def extend_pe(self, x):
+        """Reset the positional encodings."""
+        if self.pe is not None:
+            # self.pe contains both positive and negative parts
+            # the length of self.pe is 2 * input_len - 1
+            if self.pe.size(1) >= x.size(1) * 2 - 1:
+                if self.pe.dtype != x.dtype or self.pe.device != x.device:
+                    self.pe = self.pe.to(dtype=x.dtype, device=x.device)
+                return
+        # Suppose `i` means to the position of query vecotr and `j` means the
+        # position of key vector. We use position relative positions when keys
+        # are to the left (i>j) and negative relative positions otherwise (i<j).
+        pe_positive = torch.zeros(x.size(1), self.d_model)
+        pe_negative = torch.zeros(x.size(1), self.d_model)
+        position = torch.arange(0, x.size(1), dtype=torch.float32).unsqueeze(1)
+        div_term = torch.exp(
+            torch.arange(0, self.d_model, 2, dtype=torch.float32)
+            * -(math.log(10000.0) / self.d_model)
+        )
+        pe_positive[:, 0::2] = torch.sin(position * div_term)
+        pe_positive[:, 1::2] = torch.cos(position * div_term)
+        pe_negative[:, 0::2] = torch.sin(-1 * position * div_term)
+        pe_negative[:, 1::2] = torch.cos(-1 * position * div_term)
+
+        # Reserve the order of positive indices and concat both positive and
+        # negative indices. This is used to support the shifting trick
+        # as in https://arxiv.org/abs/1901.02860
+        pe_positive = torch.flip(pe_positive, [0]).unsqueeze(0)
+        pe_negative = pe_negative[1:].unsqueeze(0)
+        pe = torch.cat([pe_positive, pe_negative], dim=1)
+        self.pe = pe.to(device=x.device, dtype=x.dtype)
+
+    def forward(self, x: torch.Tensor, offset: Union[int, torch.Tensor] = 0):
+        """Add positional encoding.
+
+        Args:
+            x (torch.Tensor): Input tensor (batch, time, `*`).
+
+        Returns:
+            torch.Tensor: Encoded tensor (batch, time, `*`).
+
+        """
+        self.extend_pe(x)
+        x = x * self.xscale
+        pos_emb = self.position_encoding(size=x.size(1), offset=offset)
+        return self.dropout(x), self.dropout(pos_emb)
+
+    def position_encoding(self,
+                          offset: Union[int, torch.Tensor],
+                          size: int) -> torch.Tensor:
+        """ For getting encoding in a streaming fashion
+
+        Attention!!!!!
+        we apply dropout only once at the whole utterance level in a none
+        streaming way, but will call this function several times with
+        increasing input size in a streaming scenario, so the dropout will
+        be applied several times.
+
+        Args:
+            offset (int or torch.tensor): start offset
+            size (int): required size of position encoding
+
+        Returns:
+            torch.Tensor: Corresponding encoding
+        """
+        pos_emb = self.pe[
+            :,
+            self.pe.size(1) // 2 - size + 1 : self.pe.size(1) // 2 + size,
+        ]
+        return pos_emb

cosyvoice/transformer/encoder.py

+# Copyright (c) 2021 Mobvoi Inc (Binbin Zhang, Di Wu)
+#               2022 Xingchen Song (sxc19@mails.tsinghua.edu.cn)
+#               2024 Alibaba Inc (Xiang Lyu)
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#   http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# Modified from ESPnet(https://github.com/espnet/espnet)
+"""Encoder definition."""
+from typing import Tuple
+
+import torch
+import torch.utils.checkpoint as ckpt
+
+from cosyvoice.transformer.convolution import ConvolutionModule
+from cosyvoice.transformer.encoder_layer import TransformerEncoderLayer
+from cosyvoice.transformer.encoder_layer import ConformerEncoderLayer
+from cosyvoice.transformer.positionwise_feed_forward import PositionwiseFeedForward
+from cosyvoice.utils.class_utils import (
+    COSYVOICE_EMB_CLASSES,
+    COSYVOICE_SUBSAMPLE_CLASSES,
+    COSYVOICE_ATTENTION_CLASSES,
+    COSYVOICE_ACTIVATION_CLASSES,
+)
+from cosyvoice.utils.mask import make_pad_mask
+from cosyvoice.utils.mask import add_optional_chunk_mask
+
+
+class BaseEncoder(torch.nn.Module):
+
+    def __init__(
+        self,
+        input_size: int,
+        output_size: int = 256,
+        attention_heads: int = 4,
+        linear_units: int = 2048,
+        num_blocks: int = 6,
+        dropout_rate: float = 0.1,
+        positional_dropout_rate: float = 0.1,
+        attention_dropout_rate: float = 0.0,
+        input_layer: str = "conv2d",
+        pos_enc_layer_type: str = "abs_pos",
+        normalize_before: bool = True,
+        static_chunk_size: int = 0,
+        use_dynamic_chunk: bool = False,
+        global_cmvn: torch.nn.Module = None,
+        use_dynamic_left_chunk: bool = False,
+        gradient_checkpointing: bool = False,
+    ):
+        """
+        Args:
+            input_size (int): input dim
+            output_size (int): dimension of attention
+            attention_heads (int): the number of heads of multi head attention
+            linear_units (int): the hidden units number of position-wise feed
+                forward
+            num_blocks (int): the number of decoder blocks
+            dropout_rate (float): dropout rate
+            attention_dropout_rate (float): dropout rate in attention
+            positional_dropout_rate (float): dropout rate after adding
+                positional encoding
+            input_layer (str): input layer type.
+                optional [linear, conv2d, conv2d6, conv2d8]
+            pos_enc_layer_type (str): Encoder positional encoding layer type.
+                opitonal [abs_pos, scaled_abs_pos, rel_pos, no_pos]
+            normalize_before (bool):
+                True: use layer_norm before each sub-block of a layer.
+                False: use layer_norm after each sub-block of a layer.
+            static_chunk_size (int): chunk size for static chunk training and
+                decoding
+            use_dynamic_chunk (bool): whether use dynamic chunk size for
+                training or not, You can only use fixed chunk(chunk_size > 0)
+                or dyanmic chunk size(use_dynamic_chunk = True)
+            global_cmvn (Optional[torch.nn.Module]): Optional GlobalCMVN module
+            use_dynamic_left_chunk (bool): whether use dynamic left chunk in
+                dynamic chunk training
+            key_bias: whether use bias in attention.linear_k, False for whisper models.
+            gradient_checkpointing: rerunning a forward-pass segment for each
+                checkpointed segment during backward.
+        """
+        super().__init__()
+        self._output_size = output_size
+
+        self.global_cmvn = global_cmvn
+        self.embed = COSYVOICE_SUBSAMPLE_CLASSES[input_layer](
+            input_size,
+            output_size,
+            dropout_rate,
+            COSYVOICE_EMB_CLASSES[pos_enc_layer_type](output_size,
+                                                      positional_dropout_rate),
+        )
+
+        self.normalize_before = normalize_before
+        self.after_norm = torch.nn.LayerNorm(output_size, eps=1e-5)
+        self.static_chunk_size = static_chunk_size
+        self.use_dynamic_chunk = use_dynamic_chunk
+        self.use_dynamic_left_chunk = use_dynamic_left_chunk
+        self.gradient_checkpointing = gradient_checkpointing
+
+    def output_size(self) -> int:
+        return self._output_size
+
+    def forward(
+        self,
+        xs: torch.Tensor,
+        xs_lens: torch.Tensor,
+        decoding_chunk_size: int = 0,
+        num_decoding_left_chunks: int = -1,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Embed positions in tensor.
+
+        Args:
+            xs: padded input tensor (B, T, D)
+            xs_lens: input length (B)
+            decoding_chunk_size: decoding chunk size for dynamic chunk
+                0: default for training, use random dynamic chunk.
+                <0: for decoding, use full chunk.
+                >0: for decoding, use fixed chunk size as set.
+            num_decoding_left_chunks: number of left chunks, this is for decoding,
+            the chunk size is decoding_chunk_size.
+                >=0: use num_decoding_left_chunks
+                <0: use all left chunks
+        Returns:
+            encoder output tensor xs, and subsampled masks
+            xs: padded output tensor (B, T' ~= T/subsample_rate, D)
+            masks: torch.Tensor batch padding mask after subsample
+                (B, 1, T' ~= T/subsample_rate)
+        NOTE(xcsong):
+            We pass the `__call__` method of the modules instead of `forward` to the
+            checkpointing API because `__call__` attaches all the hooks of the module.
+            https://discuss.pytorch.org/t/any-different-between-model-input-and-model-forward-input/3690/2
+        """
+        T = xs.size(1)
+        masks = ~make_pad_mask(xs_lens, T).unsqueeze(1)  # (B, 1, T)
+        if self.global_cmvn is not None:
+            xs = self.global_cmvn(xs)
+        xs, pos_emb, masks = self.embed(xs, masks)
+        mask_pad = masks  # (B, 1, T/subsample_rate)
+        chunk_masks = add_optional_chunk_mask(xs, masks,
+                                              self.use_dynamic_chunk,
+                                              self.use_dynamic_left_chunk,
+                                              decoding_chunk_size,
+                                              self.static_chunk_size,
+                                              num_decoding_left_chunks)
+        if self.gradient_checkpointing and self.training:
+            xs = self.forward_layers_checkpointed(xs, chunk_masks, pos_emb,
+                                                  mask_pad)
+        else:
+            xs = self.forward_layers(xs, chunk_masks, pos_emb, mask_pad)
+        if self.normalize_before:
+            xs = self.after_norm(xs)
+        # Here we assume the mask is not changed in encoder layers, so just
+        # return the masks before encoder layers, and the masks will be used
+        # for cross attention with decoder later
+        return xs, masks
+
+    def forward_layers(self, xs: torch.Tensor, chunk_masks: torch.Tensor,
+                       pos_emb: torch.Tensor,
+                       mask_pad: torch.Tensor) -> torch.Tensor:
+        for layer in self.encoders:
+            xs, chunk_masks, _, _ = layer(xs, chunk_masks, pos_emb, mask_pad)
+        return xs
+
+    @torch.jit.ignore(drop=True)
+    def forward_layers_checkpointed(self, xs: torch.Tensor,
+                                    chunk_masks: torch.Tensor,
+                                    pos_emb: torch.Tensor,
+                                    mask_pad: torch.Tensor) -> torch.Tensor:
+        for layer in self.encoders:
+            xs, chunk_masks, _, _ = ckpt.checkpoint(layer.__call__, xs,
+                                                    chunk_masks, pos_emb,
+                                                    mask_pad)
+        return xs
+
+    def forward_chunk(
+        self,
+        xs: torch.Tensor,
+        offset: int,
+        required_cache_size: int,
+        att_cache: torch.Tensor = torch.zeros(0, 0, 0, 0),
+        cnn_cache: torch.Tensor = torch.zeros(0, 0, 0, 0),
+        att_mask: torch.Tensor = torch.ones((0, 0, 0), dtype=torch.bool),
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        """ Forward just one chunk
+
+        Args:
+            xs (torch.Tensor): chunk input, with shape (b=1, time, mel-dim),
+                where `time == (chunk_size - 1) * subsample_rate + \
+                        subsample.right_context + 1`
+            offset (int): current offset in encoder output time stamp
+            required_cache_size (int): cache size required for next chunk
+                compuation
+                >=0: actual cache size
+                <0: means all history cache is required
+            att_cache (torch.Tensor): cache tensor for KEY & VALUE in
+                transformer/conformer attention, with shape
+                (elayers, head, cache_t1, d_k * 2), where
+                `head * d_k == hidden-dim` and
+                `cache_t1 == chunk_size * num_decoding_left_chunks`.
+            cnn_cache (torch.Tensor): cache tensor for cnn_module in conformer,
+                (elayers, b=1, hidden-dim, cache_t2), where
+                `cache_t2 == cnn.lorder - 1`
+
+        Returns:
+            torch.Tensor: output of current input xs,
+                with shape (b=1, chunk_size, hidden-dim).
+            torch.Tensor: new attention cache required for next chunk, with
+                dynamic shape (elayers, head, ?, d_k * 2)
+                depending on required_cache_size.
+            torch.Tensor: new conformer cnn cache required for next chunk, with
+                same shape as the original cnn_cache.
+
+        """
+        assert xs.size(0) == 1
+        # tmp_masks is just for interface compatibility
+        tmp_masks = torch.ones(1,
+                               xs.size(1),
+                               device=xs.device,
+                               dtype=torch.bool)
+        tmp_masks = tmp_masks.unsqueeze(1)
+        if self.global_cmvn is not None:
+            xs = self.global_cmvn(xs)
+        # NOTE(xcsong): Before embed, shape(xs) is (b=1, time, mel-dim)
+        xs, pos_emb, _ = self.embed(xs, tmp_masks, offset)
+        # NOTE(xcsong): After  embed, shape(xs) is (b=1, chunk_size, hidden-dim)
+        elayers, cache_t1 = att_cache.size(0), att_cache.size(2)
+        chunk_size = xs.size(1)
+        attention_key_size = cache_t1 + chunk_size
+        pos_emb = self.embed.position_encoding(offset=offset - cache_t1,
+                                               size=attention_key_size)
+        if required_cache_size < 0:
+            next_cache_start = 0
+        elif required_cache_size == 0:
+            next_cache_start = attention_key_size
+        else:
+            next_cache_start = max(attention_key_size - required_cache_size, 0)
+        r_att_cache = []
+        r_cnn_cache = []
+        for i, layer in enumerate(self.encoders):
+            # NOTE(xcsong): Before layer.forward
+            #   shape(att_cache[i:i + 1]) is (1, head, cache_t1, d_k * 2),
+            #   shape(cnn_cache[i])       is (b=1, hidden-dim, cache_t2)
+            xs, _, new_att_cache, new_cnn_cache = layer(
+                xs,
+                att_mask,
+                pos_emb,
+                att_cache=att_cache[i:i + 1] if elayers > 0 else att_cache,
+                cnn_cache=cnn_cache[i] if cnn_cache.size(0) > 0 else cnn_cache)
+            # NOTE(xcsong): After layer.forward
+            #   shape(new_att_cache) is (1, head, attention_key_size, d_k * 2),
+            #   shape(new_cnn_cache) is (b=1, hidden-dim, cache_t2)
+            r_att_cache.append(new_att_cache[:, :, next_cache_start:, :])
+            r_cnn_cache.append(new_cnn_cache.unsqueeze(0))
+        if self.normalize_before:
+            xs = self.after_norm(xs)
+
+        # NOTE(xcsong): shape(r_att_cache) is (elayers, head, ?, d_k * 2),
+        #   ? may be larger than cache_t1, it depends on required_cache_size
+        r_att_cache = torch.cat(r_att_cache, dim=0)
+        # NOTE(xcsong): shape(r_cnn_cache) is (e, b=1, hidden-dim, cache_t2)
+        r_cnn_cache = torch.cat(r_cnn_cache, dim=0)
+
+        return (xs, r_att_cache, r_cnn_cache)
+
+    def forward_chunk_by_chunk(
+        self,
+        xs: torch.Tensor,
+        decoding_chunk_size: int,
+        num_decoding_left_chunks: int = -1,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """ Forward input chunk by chunk with chunk_size like a streaming
+            fashion
+
+        Here we should pay special attention to computation cache in the
+        streaming style forward chunk by chunk. Three things should be taken
+        into account for computation in the current network:
+            1. transformer/conformer encoder layers output cache
+            2. convolution in conformer
+            3. convolution in subsampling
+
+        However, we don't implement subsampling cache for:
+            1. We can control subsampling module to output the right result by
+               overlapping input instead of cache left context, even though it
+               wastes some computation, but subsampling only takes a very
+               small fraction of computation in the whole model.
+            2. Typically, there are several covolution layers with subsampling
+               in subsampling module, it is tricky and complicated to do cache
+               with different convolution layers with different subsampling
+               rate.
+            3. Currently, nn.Sequential is used to stack all the convolution
+               layers in subsampling, we need to rewrite it to make it work
+               with cache, which is not preferred.
+        Args:
+            xs (torch.Tensor): (1, max_len, dim)
+            chunk_size (int): decoding chunk size
+        """
+        assert decoding_chunk_size > 0
+        # The model is trained by static or dynamic chunk
+        assert self.static_chunk_size > 0 or self.use_dynamic_chunk
+        subsampling = self.embed.subsampling_rate
+        context = self.embed.right_context + 1  # Add current frame
+        stride = subsampling * decoding_chunk_size
+        decoding_window = (decoding_chunk_size - 1) * subsampling + context
+        num_frames = xs.size(1)
+        att_cache: torch.Tensor = torch.zeros((0, 0, 0, 0), device=xs.device)
+        cnn_cache: torch.Tensor = torch.zeros((0, 0, 0, 0), device=xs.device)
+        outputs = []
+        offset = 0
+        required_cache_size = decoding_chunk_size * num_decoding_left_chunks
+
+        # Feed forward overlap input step by step
+        for cur in range(0, num_frames - context + 1, stride):
+            end = min(cur + decoding_window, num_frames)
+            chunk_xs = xs[:, cur:end, :]
+            (y, att_cache,
+             cnn_cache) = self.forward_chunk(chunk_xs, offset,
+                                             required_cache_size, att_cache,
+                                             cnn_cache)
+            outputs.append(y)
+            offset += y.size(1)
+        ys = torch.cat(outputs, 1)
+        masks = torch.ones((1, 1, ys.size(1)),
+                           device=ys.device,
+                           dtype=torch.bool)
+        return ys, masks
+
+
+class TransformerEncoder(BaseEncoder):
+    """Transformer encoder module."""
+
+    def __init__(
+        self,
+        input_size: int,
+        output_size: int = 256,
+        attention_heads: int = 4,
+        linear_units: int = 2048,
+        num_blocks: int = 6,
+        dropout_rate: float = 0.1,
+        positional_dropout_rate: float = 0.1,
+        attention_dropout_rate: float = 0.0,
+        input_layer: str = "conv2d",
+        pos_enc_layer_type: str = "abs_pos",
+        normalize_before: bool = True,
+        static_chunk_size: int = 0,
+        use_dynamic_chunk: bool = False,
+        global_cmvn: torch.nn.Module = None,
+        use_dynamic_left_chunk: bool = False,
+        key_bias: bool = True,
+        selfattention_layer_type: str = "selfattn",
+        activation_type: str = "relu",
+        gradient_checkpointing: bool = False,
+    ):
+        """ Construct TransformerEncoder
+
+        See Encoder for the meaning of each parameter.
+        """
+        super().__init__(input_size, output_size, attention_heads,
+                         linear_units, num_blocks, dropout_rate,
+                         positional_dropout_rate, attention_dropout_rate,
+                         input_layer, pos_enc_layer_type, normalize_before,
+                         static_chunk_size, use_dynamic_chunk, global_cmvn,
+                         use_dynamic_left_chunk, gradient_checkpointing)
+        activation = COSYVOICE_ACTIVATION_CLASSES[activation_type]()
+        self.encoders = torch.nn.ModuleList([
+            TransformerEncoderLayer(
+                output_size,
+                COSYVOICE_ATTENTION_CLASSES[selfattention_layer_type](attention_heads,
+                                                                      output_size,
+                                                                      attention_dropout_rate,
+                                                                      key_bias),
+                PositionwiseFeedForward(output_size, linear_units,
+                                        dropout_rate, activation),
+                dropout_rate, normalize_before) for _ in range(num_blocks)
+        ])
+
+
+class ConformerEncoder(BaseEncoder):
+    """Conformer encoder module."""
+
+    def __init__(
+        self,
+        input_size: int,
+        output_size: int = 256,
+        attention_heads: int = 4,
+        linear_units: int = 2048,
+        num_blocks: int = 6,
+        dropout_rate: float = 0.1,
+        positional_dropout_rate: float = 0.1,
+        attention_dropout_rate: float = 0.0,
+        input_layer: str = "conv2d",
+        pos_enc_layer_type: str = "rel_pos",
+        normalize_before: bool = True,
+        static_chunk_size: int = 0,
+        use_dynamic_chunk: bool = False,
+        global_cmvn: torch.nn.Module = None,
+        use_dynamic_left_chunk: bool = False,
+        positionwise_conv_kernel_size: int = 1,
+        macaron_style: bool = True,
+        selfattention_layer_type: str = "rel_selfattn",
+        activation_type: str = "swish",
+        use_cnn_module: bool = True,
+        cnn_module_kernel: int = 15,
+        causal: bool = False,
+        cnn_module_norm: str = "batch_norm",
+        key_bias: bool = True,
+        gradient_checkpointing: bool = False,
+    ):
+        """Construct ConformerEncoder
+
+        Args:
+            input_size to use_dynamic_chunk, see in BaseEncoder
+            positionwise_conv_kernel_size (int): Kernel size of positionwise
+                conv1d layer.
+            macaron_style (bool): Whether to use macaron style for
+                positionwise layer.
+            selfattention_layer_type (str): Encoder attention layer type,
+                the parameter has no effect now, it's just for configure
+                compatibility.
+            activation_type (str): Encoder activation function type.
+            use_cnn_module (bool): Whether to use convolution module.
+            cnn_module_kernel (int): Kernel size of convolution module.
+            causal (bool): whether to use causal convolution or not.
+            key_bias: whether use bias in attention.linear_k, False for whisper models.
+        """
+        super().__init__(input_size, output_size, attention_heads,
+                         linear_units, num_blocks, dropout_rate,
+                         positional_dropout_rate, attention_dropout_rate,
+                         input_layer, pos_enc_layer_type, normalize_before,
+                         static_chunk_size, use_dynamic_chunk, global_cmvn,
+                         use_dynamic_left_chunk, gradient_checkpointing)
+        activation = COSYVOICE_ACTIVATION_CLASSES[activation_type]()
+
+        # self-attention module definition
+        encoder_selfattn_layer_args = (
+            attention_heads,
+            output_size,
+            attention_dropout_rate,
+            key_bias,
+        )
+        # feed-forward module definition
+        positionwise_layer_args = (
+            output_size,
+            linear_units,
+            dropout_rate,
+            activation,
+        )
+        # convolution module definition
+        convolution_layer_args = (output_size, cnn_module_kernel, activation,
+                                  cnn_module_norm, causal)
+
+        self.encoders = torch.nn.ModuleList([
+            ConformerEncoderLayer(
+                output_size,
+                COSYVOICE_ATTENTION_CLASSES[selfattention_layer_type](
+                    *encoder_selfattn_layer_args),
+                PositionwiseFeedForward(*positionwise_layer_args),
+                PositionwiseFeedForward(
+                    *positionwise_layer_args) if macaron_style else None,
+                ConvolutionModule(
+                    *convolution_layer_args) if use_cnn_module else None,
+                dropout_rate,
+                normalize_before,
+            ) for _ in range(num_blocks)
+        ])

cosyvoice/transformer/encoder_layer.py

+# Copyright (c) 2021 Mobvoi Inc (Binbin Zhang, Di Wu)
+#               2022 Xingchen Song (sxc19@mails.tsinghua.edu.cn)
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#   http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# Modified from ESPnet(https://github.com/espnet/espnet)
+"""Encoder self-attention layer definition."""
+
+from typing import Optional, Tuple
+
+import torch
+from torch import nn
+
+
+class TransformerEncoderLayer(nn.Module):
+    """Encoder layer module.
+
+    Args:
+        size (int): Input dimension.
+        self_attn (torch.nn.Module): Self-attention module instance.
+            `MultiHeadedAttention` or `RelPositionMultiHeadedAttention`
+            instance can be used as the argument.
+        feed_forward (torch.nn.Module): Feed-forward module instance.
+            `PositionwiseFeedForward`, instance can be used as the argument.
+        dropout_rate (float): Dropout rate.
+        normalize_before (bool):
+            True: use layer_norm before each sub-block.
+            False: to use layer_norm after each sub-block.
+    """
+
+    def __init__(
+        self,
+        size: int,
+        self_attn: torch.nn.Module,
+        feed_forward: torch.nn.Module,
+        dropout_rate: float,
+        normalize_before: bool = True,
+    ):
+        """Construct an EncoderLayer object."""
+        super().__init__()
+        self.self_attn = self_attn
+        self.feed_forward = feed_forward
+        self.norm1 = nn.LayerNorm(size, eps=1e-5)
+        self.norm2 = nn.LayerNorm(size, eps=1e-5)
+        self.dropout = nn.Dropout(dropout_rate)
+        self.size = size
+        self.normalize_before = normalize_before
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        mask: torch.Tensor,
+        pos_emb: torch.Tensor,
+        mask_pad: torch.Tensor = torch.ones((0, 0, 0), dtype=torch.bool),
+        att_cache: torch.Tensor = torch.zeros((0, 0, 0, 0)),
+        cnn_cache: torch.Tensor = torch.zeros((0, 0, 0, 0)),
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
+        """Compute encoded features.
+
+        Args:
+            x (torch.Tensor): (#batch, time, size)
+            mask (torch.Tensor): Mask tensor for the input (#batch, time，time),
+                (0, 0, 0) means fake mask.
+            pos_emb (torch.Tensor): just for interface compatibility
+                to ConformerEncoderLayer
+            mask_pad (torch.Tensor): does not used in transformer layer,
+                just for unified api with conformer.
+            att_cache (torch.Tensor): Cache tensor of the KEY & VALUE
+                (#batch=1, head, cache_t1, d_k * 2), head * d_k == size.
+            cnn_cache (torch.Tensor): Convolution cache in conformer layer
+                (#batch=1, size, cache_t2), not used here, it's for interface
+                compatibility to ConformerEncoderLayer.
+        Returns:
+            torch.Tensor: Output tensor (#batch, time, size).
+            torch.Tensor: Mask tensor (#batch, time, time).
+            torch.Tensor: att_cache tensor,
+                (#batch=1, head, cache_t1 + time, d_k * 2).
+            torch.Tensor: cnn_cahce tensor (#batch=1, size, cache_t2).
+
+        """
+        residual = x
+        if self.normalize_before:
+            x = self.norm1(x)
+        x_att, new_att_cache = self.self_attn(x, x, x, mask, pos_emb=pos_emb, cache=att_cache)
+        x = residual + self.dropout(x_att)
+        if not self.normalize_before:
+            x = self.norm1(x)
+
+        residual = x
+        if self.normalize_before:
+            x = self.norm2(x)
+        x = residual + self.dropout(self.feed_forward(x))
+        if not self.normalize_before:
+            x = self.norm2(x)
+
+        fake_cnn_cache = torch.zeros((0, 0, 0), dtype=x.dtype, device=x.device)
+        return x, mask, new_att_cache, fake_cnn_cache
+
+
+class ConformerEncoderLayer(nn.Module):
+    """Encoder layer module.
+    Args:
+        size (int): Input dimension.
+        self_attn (torch.nn.Module): Self-attention module instance.
+            `MultiHeadedAttention` or `RelPositionMultiHeadedAttention`
+            instance can be used as the argument.
+        feed_forward (torch.nn.Module): Feed-forward module instance.
+            `PositionwiseFeedForward` instance can be used as the argument.
+        feed_forward_macaron (torch.nn.Module): Additional feed-forward module
+             instance.
+            `PositionwiseFeedForward` instance can be used as the argument.
+        conv_module (torch.nn.Module): Convolution module instance.
+            `ConvlutionModule` instance can be used as the argument.
+        dropout_rate (float): Dropout rate.
+        normalize_before (bool):
+            True: use layer_norm before each sub-block.
+            False: use layer_norm after each sub-block.
+    """
+
+    def __init__(
+        self,
+        size: int,
+        self_attn: torch.nn.Module,
+        feed_forward: Optional[nn.Module] = None,
+        feed_forward_macaron: Optional[nn.Module] = None,
+        conv_module: Optional[nn.Module] = None,
+        dropout_rate: float = 0.1,
+        normalize_before: bool = True,
+    ):
+        """Construct an EncoderLayer object."""
+        super().__init__()
+        self.self_attn = self_attn
+        self.feed_forward = feed_forward
+        self.feed_forward_macaron = feed_forward_macaron
+        self.conv_module = conv_module
+        self.norm_ff = nn.LayerNorm(size, eps=1e-5)  # for the FNN module
+        self.norm_mha = nn.LayerNorm(size, eps=1e-5)  # for the MHA module
+        if feed_forward_macaron is not None:
+            self.norm_ff_macaron = nn.LayerNorm(size, eps=1e-5)
+            self.ff_scale = 0.5
+        else:
+            self.ff_scale = 1.0
+        if self.conv_module is not None:
+            self.norm_conv = nn.LayerNorm(size, eps=1e-5)  # for the CNN module
+            self.norm_final = nn.LayerNorm(
+                size, eps=1e-5)  # for the final output of the block
+        self.dropout = nn.Dropout(dropout_rate)
+        self.size = size
+        self.normalize_before = normalize_before
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        mask: torch.Tensor,
+        pos_emb: torch.Tensor,
+        mask_pad: torch.Tensor = torch.ones((0, 0, 0), dtype=torch.bool),
+        att_cache: torch.Tensor = torch.zeros((0, 0, 0, 0)),
+        cnn_cache: torch.Tensor = torch.zeros((0, 0, 0, 0)),
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
+        """Compute encoded features.
+
+        Args:
+            x (torch.Tensor): (#batch, time, size)
+            mask (torch.Tensor): Mask tensor for the input (#batch, time，time),
+                (0, 0, 0) means fake mask.
+            pos_emb (torch.Tensor): positional encoding, must not be None
+                for ConformerEncoderLayer.
+            mask_pad (torch.Tensor): batch padding mask used for conv module.
+                (#batch, 1，time), (0, 0, 0) means fake mask.
+            att_cache (torch.Tensor): Cache tensor of the KEY & VALUE
+                (#batch=1, head, cache_t1, d_k * 2), head * d_k == size.
+            cnn_cache (torch.Tensor): Convolution cache in conformer layer
+                (#batch=1, size, cache_t2)
+        Returns:
+            torch.Tensor: Output tensor (#batch, time, size).
+            torch.Tensor: Mask tensor (#batch, time, time).
+            torch.Tensor: att_cache tensor,
+                (#batch=1, head, cache_t1 + time, d_k * 2).
+            torch.Tensor: cnn_cahce tensor (#batch, size, cache_t2).
+        """
+
+        # whether to use macaron style
+        if self.feed_forward_macaron is not None:
+            residual = x
+            if self.normalize_before:
+                x = self.norm_ff_macaron(x)
+            x = residual + self.ff_scale * self.dropout(
+                self.feed_forward_macaron(x))
+            if not self.normalize_before:
+                x = self.norm_ff_macaron(x)
+
+        # multi-headed self-attention module
+        residual = x
+        if self.normalize_before:
+            x = self.norm_mha(x)
+        x_att, new_att_cache = self.self_attn(x, x, x, mask, pos_emb,
+                                              att_cache)
+        x = residual + self.dropout(x_att)
+        if not self.normalize_before:
+            x = self.norm_mha(x)
+
+        # convolution module
+        # Fake new cnn cache here, and then change it in conv_module
+        new_cnn_cache = torch.zeros((0, 0, 0), dtype=x.dtype, device=x.device)
+        if self.conv_module is not None:
+            residual = x
+            if self.normalize_before:
+                x = self.norm_conv(x)
+            x, new_cnn_cache = self.conv_module(x, mask_pad, cnn_cache)
+            x = residual + self.dropout(x)
+
+            if not self.normalize_before:
+                x = self.norm_conv(x)
+
+        # feed forward module
+        residual = x
+        if self.normalize_before:
+            x = self.norm_ff(x)
+
+        x = residual + self.ff_scale * self.dropout(self.feed_forward(x))
+        if not self.normalize_before:
+            x = self.norm_ff(x)
+
+        if self.conv_module is not None:
+            x = self.norm_final(x)
+
+        return x, mask, new_att_cache, new_cnn_cache

cosyvoice/transformer/label_smoothing_loss.py

+# Copyright (c) 2019 Shigeki Karita
+#               2020 Mobvoi Inc (Binbin Zhang)
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#   http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Label smoothing module."""
+
+import torch
+from torch import nn
+
+
+class LabelSmoothingLoss(nn.Module):
+    """Label-smoothing loss.
+
+    In a standard CE loss, the label's data distribution is:
+    [0,1,2] ->
+    [
+        [1.0, 0.0, 0.0],
+        [0.0, 1.0, 0.0],
+        [0.0, 0.0, 1.0],
+    ]
+
+    In the smoothing version CE Loss,some probabilities
+    are taken from the true label prob (1.0) and are divided
+    among other labels.
+
+    e.g.
+    smoothing=0.1
+    [0,1,2] ->
+    [
+        [0.9, 0.05, 0.05],
+        [0.05, 0.9, 0.05],
+        [0.05, 0.05, 0.9],
+    ]
+
+    Args:
+        size (int): the number of class
+        padding_idx (int): padding class id which will be ignored for loss
+        smoothing (float): smoothing rate (0.0 means the conventional CE)
+        normalize_length (bool):
+            normalize loss by sequence length if True
+            normalize loss by batch size if False
+    """
+
+    def __init__(self,
+                 size: int,
+                 padding_idx: int,
+                 smoothing: float,
+                 normalize_length: bool = False):
+        """Construct an LabelSmoothingLoss object."""
+        super(LabelSmoothingLoss, self).__init__()
+        self.criterion = nn.KLDivLoss(reduction="none")
+        self.padding_idx = padding_idx
+        self.confidence = 1.0 - smoothing
+        self.smoothing = smoothing
+        self.size = size
+        self.normalize_length = normalize_length
+
+    def forward(self, x: torch.Tensor, target: torch.Tensor) -> torch.Tensor:
+        """Compute loss between x and target.
+
+        The model outputs and data labels tensors are flatten to
+        (batch*seqlen, class) shape and a mask is applied to the
+        padding part which should not be calculated for loss.
+
+        Args:
+            x (torch.Tensor): prediction (batch, seqlen, class)
+            target (torch.Tensor):
+                target signal masked with self.padding_id (batch, seqlen)
+        Returns:
+            loss (torch.Tensor) : The KL loss, scalar float value
+        """
+        assert x.size(2) == self.size
+        batch_size = x.size(0)
+        x = x.view(-1, self.size)
+        target = target.view(-1)
+        # use zeros_like instead of torch.no_grad() for true_dist,
+        # since no_grad() can not be exported by JIT
+        true_dist = torch.zeros_like(x)
+        true_dist.fill_(self.smoothing / (self.size - 1))
+        ignore = target == self.padding_idx  # (B,)
+        total = len(target) - ignore.sum().item()
+        target = target.masked_fill(ignore, 0)  # avoid -1 index
+        true_dist.scatter_(1, target.unsqueeze(1), self.confidence)
+        kl = self.criterion(torch.log_softmax(x, dim=1), true_dist)
+        denom = total if self.normalize_length else batch_size
+        return kl.masked_fill(ignore.unsqueeze(1), 0).sum() / denom

cosyvoice/transformer/positionwise_feed_forward.py

+# Copyright (c) 2019 Shigeki Karita
+#               2020 Mobvoi Inc (Binbin Zhang)
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#   http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Positionwise feed forward layer definition."""
+
+import torch
+
+
+class PositionwiseFeedForward(torch.nn.Module):
+    """Positionwise feed forward layer.
+
+    FeedForward are appied on each position of the sequence.
+    The output dim is same with the input dim.
+
+    Args:
+        idim (int): Input dimenstion.
+        hidden_units (int): The number of hidden units.
+        dropout_rate (float): Dropout rate.
+        activation (torch.nn.Module): Activation function
+    """
+
+    def __init__(
+            self,
+            idim: int,
+            hidden_units: int,
+            dropout_rate: float,
+            activation: torch.nn.Module = torch.nn.ReLU(),
+    ):
+        """Construct a PositionwiseFeedForward object."""
+        super(PositionwiseFeedForward, self).__init__()
+        self.w_1 = torch.nn.Linear(idim, hidden_units)
+        self.activation = activation
+        self.dropout = torch.nn.Dropout(dropout_rate)
+        self.w_2 = torch.nn.Linear(hidden_units, idim)
+
+    def forward(self, xs: torch.Tensor) -> torch.Tensor:
+        """Forward function.
+
+        Args:
+            xs: input tensor (B, L, D)
+        Returns:
+            output tensor, (B, L, D)
+        """
+        return self.w_2(self.dropout(self.activation(self.w_1(xs))))
+
+
+class MoEFFNLayer(torch.nn.Module):
+    """
+    Mixture of expert with Positionwise feed forward layer
+    See also figure 1 in https://arxiv.org/pdf/2305.15663.pdf
+    The output dim is same with the input dim.
+
+    Modified from https://github.com/Lightning-AI/lit-gpt/pull/823
+                  https://github.com/mistralai/mistral-src/blob/b46d6/moe_one_file_ref.py#L203-L219
+    Args:
+        n_expert: number of expert.
+        n_expert_per_token: The actual number of experts used for each frame
+        idim (int): Input dimenstion.
+        hidden_units (int): The number of hidden units.
+        dropout_rate (float): Dropout rate.
+        activation (torch.nn.Module): Activation function
+    """
+
+    def __init__(
+            self,
+            n_expert: int,
+            n_expert_per_token: int,
+            idim: int,
+            hidden_units: int,
+            dropout_rate: float,
+            activation: torch.nn.Module = torch.nn.ReLU(),
+    ):
+        super(MoEFFNLayer, self).__init__()
+        self.gate = torch.nn.Linear(idim, n_expert, bias=False)
+        self.experts = torch.nn.ModuleList(
+            PositionwiseFeedForward(idim, hidden_units, dropout_rate,
+                                    activation) for _ in range(n_expert))
+        self.n_expert_per_token = n_expert_per_token
+
+    def forward(self, xs: torch.Tensor) -> torch.Tensor:
+        """Foward function.
+        Args:
+            xs: input tensor (B, L, D)
+        Returns:
+            output tensor, (B, L, D)
+
+        """
+        B, L, D = xs.size(
+        )  # batch size, sequence length, embedding dimension (idim)
+        xs = xs.view(-1, D)  # (B*L, D)
+        router = self.gate(xs)  # (B*L, n_expert)
+        logits, indices = torch.topk(
+            router, self.n_expert_per_token
+        )  # probs:(B*L, n_expert), indices: (B*L, n_expert)
+        weights = torch.nn.functional.softmax(
+            logits, dim=1,
+            dtype=torch.float).to(dtype=xs.dtype)  # (B*L, n_expert_per_token)
+        output = torch.zeros_like(xs)  # (B*L, D)
+        for i, expert in enumerate(self.experts):
+            mask = indices == i
+            batch_idx, ith_expert = torch.where(mask)
+            output[batch_idx] += weights[batch_idx, ith_expert, None] * expert(
+                xs[batch_idx])
+        return output.view(B, L, D)

cosyvoice/transformer/subsampling.py

+# Copyright (c) 2021 Mobvoi Inc (Binbin Zhang, Di Wu)
+#               2024 Alibaba Inc (Xiang Lyu)
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#   http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# Modified from ESPnet(https://github.com/espnet/espnet)
+"""Subsampling layer definition."""
+
+from typing import Tuple, Union
+
+import torch
+
+
+class BaseSubsampling(torch.nn.Module):
+
+    def __init__(self):
+        super().__init__()
+        self.right_context = 0
+        self.subsampling_rate = 1
+
+    def position_encoding(self, offset: Union[int, torch.Tensor],
+                          size: int) -> torch.Tensor:
+        return self.pos_enc.position_encoding(offset, size)
+
+
+class EmbedinigNoSubsampling(BaseSubsampling):
+    """Embedding input without subsampling
+    """
+
+    def __init__(self, idim: int, odim: int, dropout_rate: float,
+                 pos_enc_class: torch.nn.Module):
+        super().__init__()
+        self.embed = torch.nn.Embedding(idim, odim)
+        self.pos_enc = pos_enc_class
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_mask: torch.Tensor,
+        offset: Union[int, torch.Tensor] = 0
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        """Input x.
+
+        Args:
+            x (torch.Tensor): Input tensor (#batch, time, idim).
+            x_mask (torch.Tensor): Input mask (#batch, 1, time).
+
+        Returns:
+            torch.Tensor: linear input tensor (#batch, time', odim),
+                where time' = time .
+            torch.Tensor: linear input mask (#batch, 1, time'),
+                where time' = time .
+
+        """
+        x = self.embed(x)
+        x, pos_emb = self.pos_enc(x, offset)
+        return x, pos_emb, x_mask
+
+
+class LinearNoSubsampling(BaseSubsampling):
+    """Linear transform the input without subsampling
+
+    Args:
+        idim (int): Input dimension.
+        odim (int): Output dimension.
+        dropout_rate (float): Dropout rate.
+
+    """
+
+    def __init__(self, idim: int, odim: int, dropout_rate: float,
+                 pos_enc_class: torch.nn.Module):
+        """Construct an linear object."""
+        super().__init__()
+        self.out = torch.nn.Sequential(
+            torch.nn.Linear(idim, odim),
+            torch.nn.LayerNorm(odim, eps=1e-5),
+            torch.nn.Dropout(dropout_rate),
+        )
+        self.pos_enc = pos_enc_class
+        self.right_context = 0
+        self.subsampling_rate = 1
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_mask: torch.Tensor,
+        offset: Union[int, torch.Tensor] = 0
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        """Input x.
+
+        Args:
+            x (torch.Tensor): Input tensor (#batch, time, idim).
+            x_mask (torch.Tensor): Input mask (#batch, 1, time).
+
+        Returns:
+            torch.Tensor: linear input tensor (#batch, time', odim),
+                where time' = time .
+            torch.Tensor: linear input mask (#batch, 1, time'),
+                where time' = time .
+
+        """
+        x = self.out(x)
+        x, pos_emb = self.pos_enc(x, offset)
+        return x, pos_emb, x_mask
+
+
+class Conv1dSubsampling2(BaseSubsampling):
+    """Convolutional 1D subsampling (to 1/2 length).
+       It is designed for Whisper, ref:
+       https://github.com/openai/whisper/blob/main/whisper/model.py
+
+    Args:
+        idim (int): Input dimension.
+        odim (int): Output dimension.
+        dropout_rate (float): Dropout rate.
+
+    """
+
+    def __init__(self, idim: int, odim: int, dropout_rate: float,
+                 pos_enc_class: torch.nn.Module):
+        """Construct an Conv1dSubsampling2 object."""
+        super().__init__()
+        self.conv = torch.nn.Sequential(
+            torch.nn.Conv1d(idim, odim, kernel_size=3, padding=1),
+            torch.nn.GELU(),
+            torch.nn.Conv1d(odim, odim, kernel_size=3, stride=2, padding=1),
+            torch.nn.GELU(),
+        )
+        self.pos_enc = pos_enc_class
+        # The right context for every conv layer is computed by:
+        # (kernel_size - 1) * frame_rate_of_this_layer
+        self.subsampling_rate = 2
+        # 4 = (3 - 1) * 1 + (3 - 1) * 1
+        self.right_context = 4
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_mask: torch.Tensor,
+        offset: Union[int, torch.Tensor] = 0
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        """Subsample x.
+
+        Args:
+            x (torch.Tensor): Input tensor (#batch, time, idim).
+            x_mask (torch.Tensor): Input mask (#batch, 1, time).
+
+        Returns:
+            torch.Tensor: Subsampled tensor (#batch, time', odim),
+                where time' = time // 2.
+            torch.Tensor: Subsampled mask (#batch, 1, time'),
+                where time' = time // 2.
+            torch.Tensor: positional encoding
+
+        """
+        time = x.size(1)
+        x = x.transpose(1, 2)  # (b, f, t)
+        x = self.conv(x)
+        x = x.transpose(1, 2)  # (b, t, f)
+        x, pos_emb = self.pos_enc(x, offset)
+        return x, pos_emb, x_mask[:, :, (time + 1) % 2::2]
+
+
+class Conv2dSubsampling4(BaseSubsampling):
+    """Convolutional 2D subsampling (to 1/4 length).
+
+    Args:
+        idim (int): Input dimension.
+        odim (int): Output dimension.
+        dropout_rate (float): Dropout rate.
+
+    """
+
+    def __init__(self, idim: int, odim: int, dropout_rate: float,
+                 pos_enc_class: torch.nn.Module):
+        """Construct an Conv2dSubsampling4 object."""
+        super().__init__()
+        self.conv = torch.nn.Sequential(
+            torch.nn.Conv2d(1, odim, 3, 2),
+            torch.nn.ReLU(),
+            torch.nn.Conv2d(odim, odim, 3, 2),
+            torch.nn.ReLU(),
+        )
+        self.out = torch.nn.Sequential(
+            torch.nn.Linear(odim * (((idim - 1) // 2 - 1) // 2), odim))
+        self.pos_enc = pos_enc_class
+        # The right context for every conv layer is computed by:
+        # (kernel_size - 1) * frame_rate_of_this_layer
+        self.subsampling_rate = 4
+        # 6 = (3 - 1) * 1 + (3 - 1) * 2
+        self.right_context = 6
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_mask: torch.Tensor,
+        offset: Union[int, torch.Tensor] = 0
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        """Subsample x.
+
+        Args:
+            x (torch.Tensor): Input tensor (#batch, time, idim).
+            x_mask (torch.Tensor): Input mask (#batch, 1, time).
+
+        Returns:
+            torch.Tensor: Subsampled tensor (#batch, time', odim),
+                where time' = time // 4.
+            torch.Tensor: Subsampled mask (#batch, 1, time'),
+                where time' = time // 4.
+            torch.Tensor: positional encoding
+
+        """
+        x = x.unsqueeze(1)  # (b, c=1, t, f)
+        x = self.conv(x)
+        b, c, t, f = x.size()
+        x = self.out(x.transpose(1, 2).contiguous().view(b, t, c * f))
+        x, pos_emb = self.pos_enc(x, offset)
+        return x, pos_emb, x_mask[:, :, 2::2][:, :, 2::2]
+
+
+class Conv2dSubsampling6(BaseSubsampling):
+    """Convolutional 2D subsampling (to 1/6 length).
+    Args:
+        idim (int): Input dimension.
+        odim (int): Output dimension.
+        dropout_rate (float): Dropout rate.
+        pos_enc (torch.nn.Module): Custom position encoding layer.
+    """
+
+    def __init__(self, idim: int, odim: int, dropout_rate: float,
+                 pos_enc_class: torch.nn.Module):
+        """Construct an Conv2dSubsampling6 object."""
+        super().__init__()
+        self.conv = torch.nn.Sequential(
+            torch.nn.Conv2d(1, odim, 3, 2),
+            torch.nn.ReLU(),
+            torch.nn.Conv2d(odim, odim, 5, 3),
+            torch.nn.ReLU(),
+        )
+        self.linear = torch.nn.Linear(odim * (((idim - 1) // 2 - 2) // 3),
+                                      odim)
+        self.pos_enc = pos_enc_class
+        # 10 = (3 - 1) * 1 + (5 - 1) * 2
+        self.subsampling_rate = 6
+        self.right_context = 10
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_mask: torch.Tensor,
+        offset: Union[int, torch.Tensor] = 0
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        """Subsample x.
+        Args:
+            x (torch.Tensor): Input tensor (#batch, time, idim).
+            x_mask (torch.Tensor): Input mask (#batch, 1, time).
+
+        Returns:
+            torch.Tensor: Subsampled tensor (#batch, time', odim),
+                where time' = time // 6.
+            torch.Tensor: Subsampled mask (#batch, 1, time'),
+                where time' = time // 6.
+            torch.Tensor: positional encoding
+        """
+        x = x.unsqueeze(1)  # (b, c, t, f)
+        x = self.conv(x)
+        b, c, t, f = x.size()
+        x = self.linear(x.transpose(1, 2).contiguous().view(b, t, c * f))
+        x, pos_emb = self.pos_enc(x, offset)
+        return x, pos_emb, x_mask[:, :, 2::2][:, :, 4::3]
+
+
+class Conv2dSubsampling8(BaseSubsampling):
+    """Convolutional 2D subsampling (to 1/8 length).
+
+    Args:
+        idim (int): Input dimension.
+        odim (int): Output dimension.
+        dropout_rate (float): Dropout rate.
+
+    """
+
+    def __init__(self, idim: int, odim: int, dropout_rate: float,
+                 pos_enc_class: torch.nn.Module):
+        """Construct an Conv2dSubsampling8 object."""
+        super().__init__()
+        self.conv = torch.nn.Sequential(
+            torch.nn.Conv2d(1, odim, 3, 2),
+            torch.nn.ReLU(),
+            torch.nn.Conv2d(odim, odim, 3, 2),
+            torch.nn.ReLU(),
+            torch.nn.Conv2d(odim, odim, 3, 2),
+            torch.nn.ReLU(),
+        )
+        self.linear = torch.nn.Linear(
+            odim * ((((idim - 1) // 2 - 1) // 2 - 1) // 2), odim)
+        self.pos_enc = pos_enc_class
+        self.subsampling_rate = 8
+        # 14 = (3 - 1) * 1 + (3 - 1) * 2 + (3 - 1) * 4
+        self.right_context = 14
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_mask: torch.Tensor,
+        offset: Union[int, torch.Tensor] = 0
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        """Subsample x.
+
+        Args:
+            x (torch.Tensor): Input tensor (#batch, time, idim).
+            x_mask (torch.Tensor): Input mask (#batch, 1, time).
+
+        Returns:
+            torch.Tensor: Subsampled tensor (#batch, time', odim),
+                where time' = time // 8.
+            torch.Tensor: Subsampled mask (#batch, 1, time'),
+                where time' = time // 8.
+            torch.Tensor: positional encoding
+        """
+        x = x.unsqueeze(1)  # (b, c, t, f)
+        x = self.conv(x)
+        b, c, t, f = x.size()
+        x = self.linear(x.transpose(1, 2).contiguous().view(b, t, c * f))
+        x, pos_emb = self.pos_enc(x, offset)
+        return x, pos_emb, x_mask[:, :, 2::2][:, :, 2::2][:, :, 2::2]
+
+
+class LegacyLinearNoSubsampling(BaseSubsampling):
+    """Linear transform the input without subsampling
+
+    Args:
+        idim (int): Input dimension.
+        odim (int): Output dimension.
+        dropout_rate (float): Dropout rate.
+
+    """
+
+    def __init__(self, idim: int, odim: int, dropout_rate: float,
+                 pos_enc_class: torch.nn.Module):
+        """Construct an linear object."""
+        super().__init__()
+        self.out = torch.nn.Sequential(
+            torch.nn.Linear(idim, odim),
+            torch.nn.LayerNorm(odim, eps=1e-5),
+            torch.nn.Dropout(dropout_rate),
+            torch.nn.ReLU(),
+        )
+        self.pos_enc = pos_enc_class
+        self.right_context = 0
+        self.subsampling_rate = 1
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_mask: torch.Tensor,
+        offset: Union[int, torch.Tensor] = 0
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        """Input x.
+
+        Args:
+            x (torch.Tensor): Input tensor (#batch, time, idim).
+            x_mask (torch.Tensor): Input mask (#batch, 1, time).
+
+        Returns:
+            torch.Tensor: linear input tensor (#batch, time', odim),
+                where time' = time .
+            torch.Tensor: linear input mask (#batch, 1, time'),
+                where time' = time .
+
+        """
+        x = self.out(x)
+        x, pos_emb = self.pos_enc(x, offset)
+        return x, pos_emb, x_mask

cosyvoice/utils/class_utils.py

+# Copyright [2023-11-28] <sxc19@mails.tsinghua.edu.cn, Xingchen Song>
+#            2024 Alibaba Inc (authors: Xiang Lyu)
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import torch
+
+from cosyvoice.transformer.activation import Swish
+from cosyvoice.transformer.subsampling import (
+    LinearNoSubsampling,
+    EmbedinigNoSubsampling,
+    Conv1dSubsampling2,
+    Conv2dSubsampling4,
+    Conv2dSubsampling6,
+    Conv2dSubsampling8,
+)
+from cosyvoice.transformer.embedding import (PositionalEncoding,
+                                             RelPositionalEncoding,
+                                             WhisperPositionalEncoding,
+                                             LearnablePositionalEncoding,
+                                             NoPositionalEncoding)
+from cosyvoice.transformer.attention import (MultiHeadedAttention,
+                                             RelPositionMultiHeadedAttention)
+from cosyvoice.transformer.embedding import EspnetRelPositionalEncoding
+from cosyvoice.transformer.subsampling import LegacyLinearNoSubsampling
+
+
+COSYVOICE_ACTIVATION_CLASSES = {
+    "hardtanh": torch.nn.Hardtanh,
+    "tanh": torch.nn.Tanh,
+    "relu": torch.nn.ReLU,
+    "selu": torch.nn.SELU,
+    "swish": getattr(torch.nn, "SiLU", Swish),
+    "gelu": torch.nn.GELU,
+}
+
+COSYVOICE_SUBSAMPLE_CLASSES = {
+    "linear": LinearNoSubsampling,
+    "linear_legacy": LegacyLinearNoSubsampling,
+    "embed": EmbedinigNoSubsampling,
+    "conv1d2": Conv1dSubsampling2,
+    "conv2d": Conv2dSubsampling4,
+    "conv2d6": Conv2dSubsampling6,
+    "conv2d8": Conv2dSubsampling8,
+    'paraformer_dummy': torch.nn.Identity
+}
+
+COSYVOICE_EMB_CLASSES = {
+    "embed": PositionalEncoding,
+    "abs_pos": PositionalEncoding,
+    "rel_pos": RelPositionalEncoding,
+    "rel_pos_espnet": EspnetRelPositionalEncoding,
+    "no_pos": NoPositionalEncoding,
+    "abs_pos_whisper": WhisperPositionalEncoding,
+    "embed_learnable_pe": LearnablePositionalEncoding,
+}
+
+COSYVOICE_ATTENTION_CLASSES = {
+    "selfattn": MultiHeadedAttention,
+    "rel_selfattn": RelPositionMultiHeadedAttention,
+}