File size: 8,533 Bytes
fe781a6
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
from __future__ import annotations

from typing import Any, Dict, Tuple, Union, Optional

import torch
import yaml
from huggingface_hub import hf_hub_download
from torch import nn
from vocos.feature_extractors import FeatureExtractor, EncodecFeatures
from vocos.heads import FourierHead
from vocos.models import Backbone


def instantiate_class(args: Union[Any, Tuple[Any, ...]], init: Dict[str, Any]) -> Any:
    """Instantiates a class with the given args and init.

    Args:
        args: Positional arguments required for instantiation.
        init: Dict of the form {"class_path":...,"init_args":...}.

    Returns:
        The instantiated class object.
    """
    kwargs = init.get("init_args", {})
    if not isinstance(args, tuple):
        args = (args,)
    class_module, class_name = init["class_path"].rsplit(".", 1)
    module = __import__(class_module, fromlist=[class_name])
    args_class = getattr(module, class_name)
    return args_class(*args, **kwargs)


class Vocos(nn.Module):
    """
    The Vocos class represents a Fourier-based neural vocoder for audio synthesis.
    This class is primarily designed for inference, with support for loading from pretrained
    model checkpoints. It consists of three main components: a feature extractor,
    a backbone, and a head.
    """

    def __init__(
        self, feature_extractor: FeatureExtractor, backbone: Backbone, head: FourierHead,
    ):
        super().__init__()
        self.feature_extractor = feature_extractor
        self.backbone = backbone
        self.head = head

    @classmethod
    def from_hparams(cls, config_path: str) -> Vocos:
        """
        Class method to create a new Vocos model instance from hyperparameters stored in a yaml configuration file.
        """
        with open(config_path, "r") as f:
            config = yaml.safe_load(f)
        feature_extractor = instantiate_class(args=(), init=config["feature_extractor"])
        backbone = instantiate_class(args=(), init=config["backbone"])
        head = instantiate_class(args=(), init=config["head"])
        model = cls(feature_extractor=feature_extractor, backbone=backbone, head=head)
        return model

    @classmethod
    def from_pretrained(cls, repo_id: str, revision: Optional[str] = None) -> Vocos:
        """
        Class method to create a new Vocos model instance from a pre-trained model stored in the Hugging Face model hub.
        """
        config_path = hf_hub_download(repo_id=repo_id, filename="config.yaml", revision=revision)
        model_path = hf_hub_download(repo_id=repo_id, filename="pytorch_model.bin", revision=revision)
        model = cls.from_hparams(config_path)
        state_dict = torch.load(model_path, map_location="cpu")
        if isinstance(model.feature_extractor, EncodecFeatures):
            encodec_parameters = {
                "feature_extractor.encodec." + key: value
                for key, value in model.feature_extractor.encodec.state_dict().items()
            }
            state_dict.update(encodec_parameters)
        model.load_state_dict(state_dict)
        model.eval()
        return model

    @torch.inference_mode()
    def forward(self, audio_input: torch.Tensor, **kwargs: Any) -> torch.Tensor:
        """
        Method to run a copy-synthesis from audio waveform. The feature extractor first processes the audio input,
        which is then passed through the backbone and the head to reconstruct the audio output.

        Args:
            audio_input (Tensor): The input tensor representing the audio waveform of shape (B, T),
                                        where B is the batch size and L is the waveform length.


        Returns:
            Tensor: The output tensor representing the reconstructed audio waveform of shape (B, T).
        """
        features = self.feature_extractor(audio_input, **kwargs)
        audio_output = self.decode(features, **kwargs)
        return audio_output

    @torch.inference_mode()
    def decode(self, features_input: torch.Tensor, **kwargs: Any) -> torch.Tensor:
        """
        Method to decode audio waveform from already calculated features. The features input is passed through
        the backbone and the head to reconstruct the audio output.

        Args:
            features_input (Tensor): The input tensor of features of shape (B, C, L), where B is the batch size,
                                     C denotes the feature dimension, and L is the sequence length.

        Returns:
            Tensor: The output tensor representing the reconstructed audio waveform of shape (B, T).
        """
        x = self.backbone(features_input, **kwargs)
        audio_output = self.head(x)
        return audio_output

    @torch.inference_mode()
    def codes_to_features(self, codes: torch.Tensor) -> torch.Tensor:
        """
        Transforms an input sequence of discrete tokens (codes) into feature embeddings using the feature extractor's
        codebook weights.

        Args:
            codes (Tensor): The input tensor. Expected shape is (K, L) or (K, B, L),
                            where K is the number of codebooks, B is the batch size and L is the sequence length.

        Returns:
            Tensor: Features of shape (B, C, L), where B is the batch size, C denotes the feature dimension,
                    and L is the sequence length.
        """
        assert isinstance(
            self.feature_extractor, EncodecFeatures
        ), "Feature extractor should be an instance of EncodecFeatures"

        if codes.dim() == 2:
            codes = codes.unsqueeze(1)

        n_bins = self.feature_extractor.encodec.quantizer.bins
        offsets = torch.arange(0, n_bins * len(codes), n_bins, device=codes.device)
        embeddings_idxs = codes + offsets.view(-1, 1, 1)
        features = torch.nn.functional.embedding(embeddings_idxs, self.feature_extractor.codebook_weights).sum(dim=0)
        features = features.transpose(1, 2)

        return features

class VocosDecoder(nn.Module):
    """
    The Vocos class represents a Fourier-based neural vocoder for audio synthesis.
    This class is primarily designed for inference, with support for loading from pretrained
    model checkpoints. It consists of three main components: a feature extractor,
    a backbone, and a head.
    """

    def __init__(
        self, backbone: Backbone, head: FourierHead,
    ):
        super().__init__()
        self.backbone = backbone
        self.head = head

    @classmethod
    def from_hparams(cls, config_path: str) -> Vocos:
        """
        Class method to create a new Vocos model instance from hyperparameters stored in a yaml configuration file.
        """
        with open(config_path, "r") as f:
            config = yaml.safe_load(f)
        backbone = instantiate_class(args=(), init=config["backbone"])
        head = instantiate_class(args=(), init=config["head"])
        model = cls(backbone=backbone, head=head)
        return model

    @torch.inference_mode()
    def forward(self, features: torch.Tensor, **kwargs: Any) -> torch.Tensor:
        """
        Method to run a copy-synthesis from audio waveform. The feature extractor first processes the audio input,
        which is then passed through the backbone and the head to reconstruct the audio output.

        Args:
            audio_input (Tensor): The input tensor representing the audio waveform of shape (B, T),
                                        where B is the batch size and L is the waveform length.


        Returns:
            Tensor: The output tensor representing the reconstructed audio waveform of shape (B, T).
        """
        audio_output = self.decode(features, **kwargs)
        return audio_output

    @torch.inference_mode()
    def decode(self, features_input: torch.Tensor, **kwargs: Any) -> torch.Tensor:
        """
        Method to decode audio waveform from already calculated features. The features input is passed through
        the backbone and the head to reconstruct the audio output.

        Args:
            features_input (Tensor): The input tensor of features of shape (B, C, L), where B is the batch size,
                                     C denotes the feature dimension, and L is the sequence length.

        Returns:
            Tensor: The output tensor representing the reconstructed audio waveform of shape (B, T).
        """
        x = self.backbone(features_input, **kwargs)
        audio_output = self.head(x)
        return audio_output