Spaces:
Running
on
Zero
Running
on
Zero
File size: 5,350 Bytes
c38f823 |
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 |
"""
ein notation:
b - batch
n - sequence
nt - text sequence
nw - raw wave length
d - dimension
"""
from __future__ import annotations
import torch
from torch import nn
import torch.nn.functional as F
from einops import repeat
from x_transformers.x_transformers import RotaryEmbedding
from model.modules import (
TimestepEmbedding,
ConvNeXtV2Block,
ConvPositionEmbedding,
DiTBlock,
AdaLayerNormZero_Final,
precompute_freqs_cis, get_pos_embed_indices,
)
# Text embedding
class TextEmbedding(nn.Module):
def __init__(self, text_num_embeds, text_dim, conv_layers = 0, conv_mult = 2):
super().__init__()
self.text_embed = nn.Embedding(text_num_embeds + 1, text_dim) # use 0 as filler token
if conv_layers > 0:
self.extra_modeling = True
self.precompute_max_pos = 4096 # ~44s of 24khz audio
self.register_buffer("freqs_cis", precompute_freqs_cis(text_dim, self.precompute_max_pos), persistent=False)
self.text_blocks = nn.Sequential(*[ConvNeXtV2Block(text_dim, text_dim * conv_mult) for _ in range(conv_layers)])
else:
self.extra_modeling = False
def forward(self, text: int['b nt'], seq_len, drop_text = False):
batch, text_len = text.shape[0], text.shape[1]
text = text + 1 # use 0 as filler token. preprocess of batch pad -1, see list_str_to_idx()
text = text[:, :seq_len] # curtail if character tokens are more than the mel spec tokens
text = F.pad(text, (0, seq_len - text_len), value = 0)
if drop_text: # cfg for text
text = torch.zeros_like(text)
text = self.text_embed(text) # b n -> b n d
# possible extra modeling
if self.extra_modeling:
# sinus pos emb
batch_start = torch.zeros((batch,), dtype=torch.long)
pos_idx = get_pos_embed_indices(batch_start, seq_len, max_pos=self.precompute_max_pos)
text_pos_embed = self.freqs_cis[pos_idx]
text = text + text_pos_embed
# convnextv2 blocks
text = self.text_blocks(text)
return text
# noised input audio and context mixing embedding
class InputEmbedding(nn.Module):
def __init__(self, mel_dim, text_dim, out_dim):
super().__init__()
self.proj = nn.Linear(mel_dim * 2 + text_dim, out_dim)
self.conv_pos_embed = ConvPositionEmbedding(dim = out_dim)
def forward(self, x: float['b n d'], cond: float['b n d'], text_embed: float['b n d'], drop_audio_cond = False):
if drop_audio_cond: # cfg for cond audio
cond = torch.zeros_like(cond)
x = self.proj(torch.cat((x, cond, text_embed), dim = -1))
x = self.conv_pos_embed(x) + x
return x
# Transformer backbone using DiT blocks
class DiT(nn.Module):
def __init__(self, *,
dim, depth = 8, heads = 8, dim_head = 64, dropout = 0.1, ff_mult = 4,
mel_dim = 100, text_num_embeds = 256, text_dim = None, conv_layers = 0,
long_skip_connection = False,
):
super().__init__()
self.time_embed = TimestepEmbedding(dim)
if text_dim is None:
text_dim = mel_dim
self.text_embed = TextEmbedding(text_num_embeds, text_dim, conv_layers = conv_layers)
self.input_embed = InputEmbedding(mel_dim, text_dim, dim)
self.rotary_embed = RotaryEmbedding(dim_head)
self.dim = dim
self.depth = depth
self.transformer_blocks = nn.ModuleList(
[
DiTBlock(
dim = dim,
heads = heads,
dim_head = dim_head,
ff_mult = ff_mult,
dropout = dropout
)
for _ in range(depth)
]
)
self.long_skip_connection = nn.Linear(dim * 2, dim, bias = False) if long_skip_connection else None
self.norm_out = AdaLayerNormZero_Final(dim) # final modulation
self.proj_out = nn.Linear(dim, mel_dim)
def forward(
self,
x: float['b n d'], # nosied input audio
cond: float['b n d'], # masked cond audio
text: int['b nt'], # text
time: float['b'] | float[''], # time step
drop_audio_cond, # cfg for cond audio
drop_text, # cfg for text
mask: bool['b n'] | None = None,
):
batch, seq_len = x.shape[0], x.shape[1]
if time.ndim == 0:
time = repeat(time, ' -> b', b = batch)
# t: conditioning time, c: context (text + masked cond audio), x: noised input audio
t = self.time_embed(time)
text_embed = self.text_embed(text, seq_len, drop_text = drop_text)
x = self.input_embed(x, cond, text_embed, drop_audio_cond = drop_audio_cond)
rope = self.rotary_embed.forward_from_seq_len(seq_len)
if self.long_skip_connection is not None:
residual = x
for block in self.transformer_blocks:
x = block(x, t, mask = mask, rope = rope)
if self.long_skip_connection is not None:
x = self.long_skip_connection(torch.cat((x, residual), dim = -1))
x = self.norm_out(x, t)
output = self.proj_out(x)
return output
|