File size: 5,953 Bytes
b68ef84 |
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 |
import torch
import torch.nn as nn
import torch.nn.functional as F
import torchaudio
from typing import Optional
from comfy.ldm.modules.attention import optimized_attention_masked
import comfy.ops
class WhisperFeatureExtractor(nn.Module):
def __init__(self, n_mels=128, device=None):
super().__init__()
self.sample_rate = 16000
self.n_fft = 400
self.hop_length = 160
self.n_mels = n_mels
self.chunk_length = 30
self.n_samples = 480000
self.mel_spectrogram = torchaudio.transforms.MelSpectrogram(
sample_rate=self.sample_rate,
n_fft=self.n_fft,
hop_length=self.hop_length,
n_mels=self.n_mels,
f_min=0,
f_max=8000,
norm="slaney",
mel_scale="slaney",
).to(device)
def __call__(self, audio):
audio = torch.mean(audio, dim=1)
batch_size = audio.shape[0]
processed_audio = []
for i in range(batch_size):
aud = audio[i]
if aud.shape[0] > self.n_samples:
aud = aud[:self.n_samples]
elif aud.shape[0] < self.n_samples:
aud = F.pad(aud, (0, self.n_samples - aud.shape[0]))
processed_audio.append(aud)
audio = torch.stack(processed_audio)
mel_spec = self.mel_spectrogram(audio.to(self.mel_spectrogram.spectrogram.window.device))[:, :, :-1].to(audio.device)
log_mel_spec = torch.clamp(mel_spec, min=1e-10).log10()
log_mel_spec = torch.maximum(log_mel_spec, log_mel_spec.max() - 8.0)
log_mel_spec = (log_mel_spec + 4.0) / 4.0
return log_mel_spec
class MultiHeadAttention(nn.Module):
def __init__(self, d_model: int, n_heads: int, dtype=None, device=None, operations=None):
super().__init__()
assert d_model % n_heads == 0
self.d_model = d_model
self.n_heads = n_heads
self.d_k = d_model // n_heads
self.q_proj = operations.Linear(d_model, d_model, dtype=dtype, device=device)
self.k_proj = operations.Linear(d_model, d_model, bias=False, dtype=dtype, device=device)
self.v_proj = operations.Linear(d_model, d_model, dtype=dtype, device=device)
self.out_proj = operations.Linear(d_model, d_model, dtype=dtype, device=device)
def forward(
self,
query: torch.Tensor,
key: torch.Tensor,
value: torch.Tensor,
mask: Optional[torch.Tensor] = None,
) -> torch.Tensor:
batch_size, seq_len, _ = query.shape
q = self.q_proj(query)
k = self.k_proj(key)
v = self.v_proj(value)
attn_output = optimized_attention_masked(q, k, v, self.n_heads, mask)
attn_output = self.out_proj(attn_output)
return attn_output
class EncoderLayer(nn.Module):
def __init__(self, d_model: int, n_heads: int, d_ff: int, dtype=None, device=None, operations=None):
super().__init__()
self.self_attn = MultiHeadAttention(d_model, n_heads, dtype=dtype, device=device, operations=operations)
self.self_attn_layer_norm = operations.LayerNorm(d_model, dtype=dtype, device=device)
self.fc1 = operations.Linear(d_model, d_ff, dtype=dtype, device=device)
self.fc2 = operations.Linear(d_ff, d_model, dtype=dtype, device=device)
self.final_layer_norm = operations.LayerNorm(d_model, dtype=dtype, device=device)
def forward(
self,
x: torch.Tensor,
attention_mask: Optional[torch.Tensor] = None
) -> torch.Tensor:
residual = x
x = self.self_attn_layer_norm(x)
x = self.self_attn(x, x, x, attention_mask)
x = residual + x
residual = x
x = self.final_layer_norm(x)
x = self.fc1(x)
x = F.gelu(x)
x = self.fc2(x)
x = residual + x
return x
class AudioEncoder(nn.Module):
def __init__(
self,
n_mels: int = 128,
n_ctx: int = 1500,
n_state: int = 1280,
n_head: int = 20,
n_layer: int = 32,
dtype=None,
device=None,
operations=None
):
super().__init__()
self.conv1 = operations.Conv1d(n_mels, n_state, kernel_size=3, padding=1, dtype=dtype, device=device)
self.conv2 = operations.Conv1d(n_state, n_state, kernel_size=3, stride=2, padding=1, dtype=dtype, device=device)
self.embed_positions = operations.Embedding(n_ctx, n_state, dtype=dtype, device=device)
self.layers = nn.ModuleList([
EncoderLayer(n_state, n_head, n_state * 4, dtype=dtype, device=device, operations=operations)
for _ in range(n_layer)
])
self.layer_norm = operations.LayerNorm(n_state, dtype=dtype, device=device)
def forward(self, x: torch.Tensor) -> torch.Tensor:
x = F.gelu(self.conv1(x))
x = F.gelu(self.conv2(x))
x = x.transpose(1, 2)
x = x + comfy.ops.cast_to_input(self.embed_positions.weight[:, :x.shape[1]], x)
all_x = ()
for layer in self.layers:
all_x += (x,)
x = layer(x)
x = self.layer_norm(x)
all_x += (x,)
return x, all_x
class WhisperLargeV3(nn.Module):
def __init__(
self,
n_mels: int = 128,
n_audio_ctx: int = 1500,
n_audio_state: int = 1280,
n_audio_head: int = 20,
n_audio_layer: int = 32,
dtype=None,
device=None,
operations=None
):
super().__init__()
self.feature_extractor = WhisperFeatureExtractor(n_mels=n_mels, device=device)
self.encoder = AudioEncoder(
n_mels, n_audio_ctx, n_audio_state, n_audio_head, n_audio_layer,
dtype=dtype, device=device, operations=operations
)
def forward(self, audio):
mel = self.feature_extractor(audio)
x, all_x = self.encoder(mel)
return x, all_x
|