File size: 14,960 Bytes
f09d4e1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
import base64
import gzip
from dataclasses import dataclass
from typing import Dict, Iterable, Optional, List

import numpy as np
import torch
import torch.nn.functional as F
from torch import Tensor, nn
from subprocess import CalledProcessError, run, Popen, PIPE

import os
from functools import lru_cache
from typing import Optional, Union

def exact_div(x, y):
    assert x % y == 0
    return x // y

# hard-coded audio hyperparameters
SAMPLE_RATE = 16000
N_FFT = 400
N_MELS = 80
HOP_LENGTH = 160
CHUNK_LENGTH = 30
N_SAMPLES = CHUNK_LENGTH * SAMPLE_RATE  # 480000 samples in a 30-second chunk
N_FRAMES = exact_div(N_SAMPLES, HOP_LENGTH)  # 3000 frames in a mel spectrogram input

N_SAMPLES_PER_TOKEN = HOP_LENGTH * 2  # the initial convolutions has stride 2
FRAMES_PER_SECOND = exact_div(SAMPLE_RATE, HOP_LENGTH)  # 10ms per audio frame
TOKENS_PER_SECOND = exact_div(SAMPLE_RATE, N_SAMPLES_PER_TOKEN)  # 20ms per audio token



def get_T_after_cnn(L_in, dilation=1):
    for (padding, kernel_size, stride) in eval("[(1,3,1)] + [(1,3,2)] "):
        L_out = L_in + 2 * padding - dilation * (kernel_size - 1) - 1
        L_out = 1 + L_out // stride
        L_in = L_out
    return L_out

def load_bytesio_audio(content, sr: int = SAMPLE_RATE):
    cmd = [
        "ffmpeg",
        "-nostdin",
        "-threads", "0",
        "-i", "pipe:",
        "-f", "s16le",
        "-ac", "1",
        "-acodec", "pcm_s16le",
        "-ar", str(sr),
        "pipe:"
    ]
    p = Popen(cmd, stdin=PIPE, stdout=PIPE, stderr=PIPE, bufsize=-1)
    out, _ = p.communicate(input=content)
    return np.frombuffer(out, np.int16).flatten().astype(np.float32) / 32768.0

def load_audio(file: str, sr: int = SAMPLE_RATE):
    """
    Open an audio file and read as mono waveform, resampling as necessary

    Parameters
    ----------
    file: str
        The audio file to open

    sr: int
        The sample rate to resample the audio if necessary

    Returns
    -------
    A NumPy array containing the audio waveform, in float32 dtype.
    """

    # This launches a subprocess to decode audio while down-mixing
    # and resampling as necessary.  Requires the ffmpeg CLI in PATH.
    # fmt: off
    cmd = [
        "ffmpeg",
        "-nostdin",
        "-threads", "0",
        "-i", file,
        "-f", "s16le",
        "-ac", "1",
        "-acodec", "pcm_s16le",
        "-ar", str(sr),
        "-"
    ]
    # fmt: on
    try:
        out = run(cmd, capture_output=True, check=True).stdout
    except CalledProcessError as e:
        raise RuntimeError(f"Failed to load audio: {e.stderr.decode()}") from e

    return np.frombuffer(out, np.int16).flatten().astype(np.float32) / 32768.0


def pad_or_trim(array, length: int = N_SAMPLES, *, axis: int = -1):
    """
    Pad or trim the audio array to N_SAMPLES, as expected by the encoder.
    """
    if torch.is_tensor(array):
        if array.shape[axis] > length:
            array = array.index_select(
                dim=axis, index=torch.arange(length, device=array.device)
            )

        if array.shape[axis] < length:
            pad_widths = [(0, 0)] * array.ndim
            pad_widths[axis] = (0, length - array.shape[axis])
            array = F.pad(array, [pad for sizes in pad_widths[::-1] for pad in sizes])
    else:
        if array.shape[axis] > length:
            array = array.take(indices=range(length), axis=axis)

        if array.shape[axis] < length:
            pad_widths = [(0, 0)] * array.ndim
            pad_widths[axis] = (0, length - array.shape[axis])
            array = np.pad(array, pad_widths)

    return array

def trim(array, length: int = N_SAMPLES, *, axis: int = -1):
    """
    Pad or trim the audio array to N_SAMPLES, as expected by the encoder.
    """
    if torch.is_tensor(array):
        if array.shape[axis] > length:
            array = array.index_select(
                dim=axis, index=torch.arange(length, device=array.device)
            )
    else:
        if array.shape[axis] > length:
            array = array.take(indices=range(length), axis=axis)
    return array


@lru_cache(maxsize=None)
def mel_filters(device, n_mels: int = N_MELS) -> torch.Tensor:
    """
    load the mel filterbank matrix for projecting STFT into a Mel spectrogram.
    Allows decoupling librosa dependency; saved using:

        np.savez_compressed(
            "mel_filters.npz",
            mel_80=librosa.filters.mel(sr=16000, n_fft=400, n_mels=80),
        )
    """
    assert n_mels == 80, f"Unsupported n_mels: {n_mels}"
    with np.load(
        os.path.join(os.path.dirname(__file__), "mel_filters.npz") # todo
        # os.path.join("assets", "mel_filters.npz")
    ) as f:
        return torch.from_numpy(f[f"mel_{n_mels}"]).to(device)


def log_mel_spectrogram(
    audio: Union[str, np.ndarray, torch.Tensor],
    n_mels: int = N_MELS,
    padding: int = 0,
    device: Optional[Union[str, torch.device]] = None,
):
    """
    Compute the log-Mel spectrogram of

    Parameters
    ----------
    audio: Union[str, np.ndarray, torch.Tensor], shape = (*)
        The path to audio or either a NumPy array or Tensor containing the audio waveform in 16 kHz

    n_mels: int
        The number of Mel-frequency filters, only 80 is supported

    padding: int
        Number of zero samples to pad to the right

    device: Optional[Union[str, torch.device]]
        If given, the audio tensor is moved to this device before STFT

    Returns
    -------
    torch.Tensor, shape = (80, n_frames)
        A Tensor that contains the Mel spectrogram
    """
    if not torch.is_tensor(audio):
        if isinstance(audio, str):
            audio = load_audio(audio)
        audio = torch.from_numpy(audio)

    if device is not None:
        audio = audio.to(device)
    if padding > 0:
        audio = F.pad(audio, (0, padding))
    window = torch.hann_window(N_FFT).to(audio.device)
    stft = torch.stft(audio, N_FFT, HOP_LENGTH, window=window, return_complex=True)
    magnitudes = stft[..., :-1].abs() ** 2

    filters = mel_filters(audio.device, n_mels)
    mel_spec = filters @ magnitudes

    log_spec = torch.clamp(mel_spec, min=1e-10).log10()
    log_spec = torch.maximum(log_spec, log_spec.max() - 8.0)
    log_spec = (log_spec + 4.0) / 4.0
    return log_spec


@dataclass
class ModelDimensions:
    n_mels: int
    n_audio_ctx: int
    n_audio_state: int
    n_audio_head: int
    n_audio_layer: int
    n_vocab: int
    n_text_ctx: int
    n_text_state: int
    n_text_head: int
    n_text_layer: int


class LayerNorm(nn.LayerNorm):
    def forward(self, x: Tensor) -> Tensor:
        # return super().forward(x.float()).type(x.dtype)
        return super().forward(x).type(x.dtype)




class Linear(nn.Linear):
    def forward(self, x: Tensor) -> Tensor:
        return F.linear(
            x,
            self.weight.to(x.dtype),
            None if self.bias is None else self.bias.to(x.dtype),
        )


class Conv1d(nn.Conv1d):
    def _conv_forward(
        self, x: Tensor, weight: Tensor, bias: Optional[Tensor]
    ) -> Tensor:
        return super()._conv_forward(
            x, weight.to(x.dtype), None if bias is None else bias.to(x.dtype)
        )


def sinusoids(length, channels, max_timescale=10000):
    """Returns sinusoids for positional embedding"""
    assert channels % 2 == 0
    log_timescale_increment = np.log(max_timescale) / (channels // 2 - 1)
    inv_timescales = torch.exp(-log_timescale_increment * torch.arange(channels // 2))
    scaled_time = torch.arange(length)[:, np.newaxis] * inv_timescales[np.newaxis, :]
    return torch.cat([torch.sin(scaled_time), torch.cos(scaled_time)], dim=1)


class MultiHeadAttention(nn.Module):
    def __init__(self, n_state: int, n_head: int):
        super().__init__()
        self.n_head = n_head
        self.query = Linear(n_state, n_state)
        self.key = Linear(n_state, n_state, bias=False)
        self.value = Linear(n_state, n_state)
        self.out = Linear(n_state, n_state)

    def forward(
        self,
        x: Tensor,
        xa: Optional[Tensor] = None,
        mask: Optional[Tensor] = None,
        kv_cache: Optional[dict] = None,
    ):
        q = self.query(x)

        if kv_cache is None or xa is None or self.key not in kv_cache:
            # hooks, if installed (i.e. kv_cache is not None), will prepend the cached kv tensors;
            # otherwise, perform key/value projections for self- or cross-attention as usual.
            k = self.key(x if xa is None else xa)
            v = self.value(x if xa is None else xa)
        else:
            # for cross-attention, calculate keys and values once and reuse in subsequent calls.
            k = kv_cache[self.key]
            v = kv_cache[self.value]

        wv, qk = self.qkv_attention(q, k, v, mask)
        return self.out(wv), qk

    def qkv_attention(
        self, q: Tensor, k: Tensor, v: Tensor, mask: Optional[Tensor] = None
    ):
        n_batch, n_ctx, n_state = q.shape
        scale = (n_state // self.n_head) ** -0.25
        q = q.view(*q.shape[:2], self.n_head, -1).permute(0, 2, 1, 3) * scale
        k = k.view(*k.shape[:2], self.n_head, -1).permute(0, 2, 3, 1) * scale
        v = v.view(*v.shape[:2], self.n_head, -1).permute(0, 2, 1, 3)

        qk = q @ k
        if mask is not None:
            qk += mask

        w = F.softmax(qk, dim=-1).to(q.dtype)
        return (w @ v).permute(0, 2, 1, 3).flatten(start_dim=2), qk.detach()


class ResidualAttentionBlock(nn.Module):
    def __init__(self, n_state: int, n_head: int, cross_attention: bool = False):
        super().__init__()

        self.attn = MultiHeadAttention(n_state, n_head)
        self.attn_ln = LayerNorm(n_state)

        self.cross_attn = (
            MultiHeadAttention(n_state, n_head) if cross_attention else None
        )
        self.cross_attn_ln = LayerNorm(n_state) if cross_attention else None

        n_mlp = n_state * 4
        self.mlp = nn.Sequential(
            Linear(n_state, n_mlp), nn.GELU(), Linear(n_mlp, n_state)
        )
        self.mlp_ln = LayerNorm(n_state)

    def forward(
        self,
        x: Tensor,
        xa: Optional[Tensor] = None,
        mask: Optional[Tensor] = None,
        kv_cache: Optional[dict] = None,
    ):
        x = x + self.attn(self.attn_ln(x), mask=mask, kv_cache=kv_cache)[0]
        if self.cross_attn:
            x = x + self.cross_attn(self.cross_attn_ln(x), xa, kv_cache=kv_cache)[0]
        x = x + self.mlp(self.mlp_ln(x))
        return x


class AudioEncoder(nn.Module):
    def __init__(
            self,
            n_mels: int,
            n_ctx: int,
            n_state: int,
            n_head: int,
            n_layer: int,
            output_dim: int = 512,
            avg_pool: bool = True,
            add_audio_bos_eos_token: bool = True,
            **kwargs
    ):
        super().__init__()
        self.conv1 = Conv1d(n_mels, n_state, kernel_size=3, padding=1)
        self.conv2 = Conv1d(n_state, n_state, kernel_size=3, stride=2, padding=1)
        self.register_buffer("positional_embedding", sinusoids(n_ctx, n_state))

        self.blocks: Iterable[ResidualAttentionBlock] = nn.ModuleList(
            [ResidualAttentionBlock(n_state, n_head) for _ in range(n_layer)]
        )
        self.ln_post = LayerNorm(n_state)

        if avg_pool:
            self.avg_pooler = nn.AvgPool1d(2, stride=2)
        else:
            self.avg_pooler = None
        self.proj = nn.Linear(n_state, output_dim)
        if add_audio_bos_eos_token:
            self.audio_bos_eos_token = nn.Embedding(2, output_dim)
        else:
            self.audio_bos_eos_token = None
        self.output_dim = output_dim
        self.n_head = n_head

    def forward(self, x: Tensor, padding_mask: Tensor=None, audio_lengths: Tensor=None):
        """
        x : torch.Tensor, shape = (batch_size, n_mels, n_ctx)
            the mel spectrogram of the audio
        """
        x = x.to(dtype=self.conv1.weight.dtype,
                 device=self.conv1.weight.device)
        if audio_lengths is not None:
            input_mel_len = audio_lengths[:,0] * 2
            max_mel_len_in_batch = input_mel_len.max()
            x = x[:, :, :max_mel_len_in_batch]
        x = F.gelu(self.conv1(x))
        x = F.gelu(self.conv2(x))
        x = x.permute(0, 2, 1)  # B, L, D
        bsz = x.size(0)
        src_len = x.size(1)


        self.input_positional_embedding = self.positional_embedding[:src_len]
        assert x.shape[1:] == self.input_positional_embedding.shape, f"incorrect audio shape: {x.shape[1:], self.input_positional_embedding.shape}"
        x = (x + self.input_positional_embedding).to(x.dtype)
        if padding_mask is not None:
            padding_mask = padding_mask.to(dtype=self.conv1.weight.dtype,
                     device=self.conv1.weight.device)
            batch_src_len = padding_mask.size(1)
            x = x[:, :batch_src_len, :]
            padding_mask = padding_mask.view(
                bsz, -1, batch_src_len
            )
            padding_mask_ = padding_mask.all(1)
            x[padding_mask_] = 0
            key_padding_mask = padding_mask_.view(bsz, 1, 1, batch_src_len). \
                expand(-1, self.n_head, -1, -1).reshape(bsz, self.n_head, 1, batch_src_len)
            new_padding_mask = torch.zeros_like(key_padding_mask, dtype=x.dtype)
            padding_mask = new_padding_mask.masked_fill(key_padding_mask, float("-inf"))

        for block in self.blocks:
            x = block(x, mask=padding_mask)


        if self.avg_pooler:
            x = x.permute(0, 2, 1)
            x = self.avg_pooler(x)
            x = x.permute(0, 2, 1)


        x = self.ln_post(x)
        x = self.proj(x)

        if self.audio_bos_eos_token is not None:
            bos = self.audio_bos_eos_token.weight[0][None, :]
            eos = self.audio_bos_eos_token.weight[1][None, :]
        else:
            bos, eos = None, None
        return x, bos, eos

    def encode(self, input_audios: Tensor, input_audio_lengths: Tensor, audio_span_tokens: List):
        real_input_audio_lens = input_audio_lengths[:, 0].tolist()
        max_len_in_batch = max(real_input_audio_lens)
        padding_mask = torch.ones([input_audios.size(0), max_len_in_batch]).to(dtype=self.conv1.weight.dtype,
                                                                               device=self.conv1.weight.device)
        for index in range(len(input_audios)):
            padding_mask[index, :input_audio_lengths[index][0].item()] = 0
        x, bos, eos = self(input_audios, padding_mask,input_audio_lengths)
        output_audios = []
        for i in range(len(audio_span_tokens)):
            audio_span = audio_span_tokens[i]
            audio = x[i][:audio_span-2]
            if bos is not None:
                audio = torch.concat([bos, audio, eos])
            assert len(audio) == audio_span
            output_audios.append(audio)
        return output_audios