Spaces:
Running
Running
File size: 10,861 Bytes
0883aa1 |
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 |
# Copyright (c) 2023 Amphion.
#
# This source code is licensed under the MIT license found in the
# LICENSE file in the root directory of this source tree.
# This module is modified from [Whisper](https://github.com/openai/whisper.git).
# ## Citations
# ```bibtex
# @inproceedings{openai-whisper,
# author = {Alec Radford and
# Jong Wook Kim and
# Tao Xu and
# Greg Brockman and
# Christine McLeavey and
# Ilya Sutskever},
# title = {Robust Speech Recognition via Large-Scale Weak Supervision},
# booktitle = {{ICML}},
# series = {Proceedings of Machine Learning Research},
# volume = {202},
# pages = {28492--28518},
# publisher = {{PMLR}},
# year = {2023}
# }
# ```
#
from dataclasses import dataclass
from typing import Dict
from typing import Iterable, Optional
import numpy as np
import torch
import torch.nn.functional as F
from torch import Tensor
from torch import nn
from .transcribe import transcribe as transcribe_function
from .decoding import (
detect_language as detect_language_function,
decode as decode_function,
)
@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)
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 = qk + mask[:n_ctx, :n_ctx]
qk = qk.float()
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
):
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)
def forward(self, x: Tensor):
"""
x : torch.Tensor, shape = (batch_size, n_mels, n_ctx)
the mel spectrogram of the audio
"""
x = F.gelu(self.conv1(x))
x = F.gelu(self.conv2(x))
x = x.permute(0, 2, 1)
assert x.shape[1:] == self.positional_embedding.shape, "incorrect audio shape"
x = (x + self.positional_embedding).to(x.dtype)
for block in self.blocks:
x = block(x)
x = self.ln_post(x)
return x
class TextDecoder(nn.Module):
def __init__(
self, n_vocab: int, n_ctx: int, n_state: int, n_head: int, n_layer: int
):
super().__init__()
self.token_embedding = nn.Embedding(n_vocab, n_state)
self.positional_embedding = nn.Parameter(torch.empty(n_ctx, n_state))
self.blocks: Iterable[ResidualAttentionBlock] = nn.ModuleList(
[
ResidualAttentionBlock(n_state, n_head, cross_attention=True)
for _ in range(n_layer)
]
)
self.ln = LayerNorm(n_state)
mask = torch.empty(n_ctx, n_ctx).fill_(-np.inf).triu_(1)
self.register_buffer("mask", mask, persistent=False)
def forward(self, x: Tensor, xa: Tensor, kv_cache: Optional[dict] = None):
"""
x : torch.LongTensor, shape = (batch_size, <= n_ctx)
the text tokens
xa : torch.Tensor, shape = (batch_size, n_mels, n_audio_ctx)
the encoded audio features to be attended on
"""
offset = next(iter(kv_cache.values())).shape[1] if kv_cache else 0
x = (
self.token_embedding(x)
+ self.positional_embedding[offset : offset + x.shape[-1]]
)
x = x.to(xa.dtype)
for block in self.blocks:
x = block(x, xa, mask=self.mask, kv_cache=kv_cache)
x = self.ln(x)
logits = (
x @ torch.transpose(self.token_embedding.weight.to(x.dtype), 0, 1)
).float()
return logits
class Whisper(nn.Module):
def __init__(self, dims: ModelDimensions):
super().__init__()
self.dims = dims
self.encoder = AudioEncoder(
self.dims.n_mels,
self.dims.n_audio_ctx,
self.dims.n_audio_state,
self.dims.n_audio_head,
self.dims.n_audio_layer,
)
self.decoder = TextDecoder(
self.dims.n_vocab,
self.dims.n_text_ctx,
self.dims.n_text_state,
self.dims.n_text_head,
self.dims.n_text_layer,
)
def embed_audio(self, mel: torch.Tensor):
return self.encoder(mel)
def logits(self, tokens: torch.Tensor, audio_features: torch.Tensor):
return self.decoder(tokens, audio_features)
def forward(
self, mel: torch.Tensor, tokens: torch.Tensor
) -> Dict[str, torch.Tensor]:
return self.decoder(tokens, self.encoder(mel))
@property
def device(self):
return next(self.parameters()).device
@property
def is_multilingual(self):
return self.dims.n_vocab == 51865
def install_kv_cache_hooks(self, cache: Optional[dict] = None):
"""
The `MultiHeadAttention` module optionally accepts `kv_cache` which stores the key and value
tensors calculated for the previous positions. This method returns a dictionary that stores
all caches, and the necessary hooks for the key and value projection modules that save the
intermediate tensors to be reused during later calculations.
Returns
-------
cache : Dict[nn.Module, torch.Tensor]
A dictionary object mapping the key/value projection modules to its cache
hooks : List[RemovableHandle]
List of PyTorch RemovableHandle objects to stop the hooks to be called
"""
cache = {**cache} if cache is not None else {}
hooks = []
def save_to_cache(module, _, output):
if (
module not in cache
or output.shape[1] > self.decoder.positional_embedding.shape[0]
):
cache[
module
] = output # save as-is, for the first token or cross attention
else:
cache[module] = torch.cat([cache[module], output], dim=1).detach()
return cache[module]
def install_hooks(layer: nn.Module):
if isinstance(layer, MultiHeadAttention):
hooks.append(layer.key.register_forward_hook(save_to_cache))
hooks.append(layer.value.register_forward_hook(save_to_cache))
self.decoder.apply(install_hooks)
return cache, hooks
detect_language = detect_language_function
transcribe = transcribe_function
decode = decode_function
|