Spaces:
Running
Running
Upload 4 files
Browse files- f5-tts/model/backbones/README.md +20 -0
- f5-tts/model/backbones/dit.py +163 -0
- f5-tts/model/backbones/mmdit.py +146 -0
- f5-tts/model/backbones/unett.py +219 -0
f5-tts/model/backbones/README.md
ADDED
|
@@ -0,0 +1,20 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
## Backbones quick introduction
|
| 2 |
+
|
| 3 |
+
|
| 4 |
+
### unett.py
|
| 5 |
+
- flat unet transformer
|
| 6 |
+
- structure same as in e2-tts & voicebox paper except using rotary pos emb
|
| 7 |
+
- update: allow possible abs pos emb & convnextv2 blocks for embedded text before concat
|
| 8 |
+
|
| 9 |
+
### dit.py
|
| 10 |
+
- adaln-zero dit
|
| 11 |
+
- embedded timestep as condition
|
| 12 |
+
- concatted noised_input + masked_cond + embedded_text, linear proj in
|
| 13 |
+
- possible abs pos emb & convnextv2 blocks for embedded text before concat
|
| 14 |
+
- possible long skip connection (first layer to last layer)
|
| 15 |
+
|
| 16 |
+
### mmdit.py
|
| 17 |
+
- sd3 structure
|
| 18 |
+
- timestep as condition
|
| 19 |
+
- left stream: text embedded and applied a abs pos emb
|
| 20 |
+
- right stream: masked_cond & noised_input concatted and with same conv pos emb as unett
|
f5-tts/model/backbones/dit.py
ADDED
|
@@ -0,0 +1,163 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""
|
| 2 |
+
ein notation:
|
| 3 |
+
b - batch
|
| 4 |
+
n - sequence
|
| 5 |
+
nt - text sequence
|
| 6 |
+
nw - raw wave length
|
| 7 |
+
d - dimension
|
| 8 |
+
"""
|
| 9 |
+
|
| 10 |
+
from __future__ import annotations
|
| 11 |
+
|
| 12 |
+
import torch
|
| 13 |
+
from torch import nn
|
| 14 |
+
import torch.nn.functional as F
|
| 15 |
+
|
| 16 |
+
from x_transformers.x_transformers import RotaryEmbedding
|
| 17 |
+
|
| 18 |
+
from f5_tts.model.modules import (
|
| 19 |
+
TimestepEmbedding,
|
| 20 |
+
ConvNeXtV2Block,
|
| 21 |
+
ConvPositionEmbedding,
|
| 22 |
+
DiTBlock,
|
| 23 |
+
AdaLayerNormZero_Final,
|
| 24 |
+
precompute_freqs_cis,
|
| 25 |
+
get_pos_embed_indices,
|
| 26 |
+
)
|
| 27 |
+
|
| 28 |
+
|
| 29 |
+
# Text embedding
|
| 30 |
+
|
| 31 |
+
|
| 32 |
+
class TextEmbedding(nn.Module):
|
| 33 |
+
def __init__(self, text_num_embeds, text_dim, conv_layers=0, conv_mult=2):
|
| 34 |
+
super().__init__()
|
| 35 |
+
self.text_embed = nn.Embedding(text_num_embeds + 1, text_dim) # use 0 as filler token
|
| 36 |
+
|
| 37 |
+
if conv_layers > 0:
|
| 38 |
+
self.extra_modeling = True
|
| 39 |
+
self.precompute_max_pos = 4096 # ~44s of 24khz audio
|
| 40 |
+
self.register_buffer("freqs_cis", precompute_freqs_cis(text_dim, self.precompute_max_pos), persistent=False)
|
| 41 |
+
self.text_blocks = nn.Sequential(
|
| 42 |
+
*[ConvNeXtV2Block(text_dim, text_dim * conv_mult) for _ in range(conv_layers)]
|
| 43 |
+
)
|
| 44 |
+
else:
|
| 45 |
+
self.extra_modeling = False
|
| 46 |
+
|
| 47 |
+
def forward(self, text: int["b nt"], seq_len, drop_text=False): # noqa: F722
|
| 48 |
+
text = text + 1 # use 0 as filler token. preprocess of batch pad -1, see list_str_to_idx()
|
| 49 |
+
text = text[:, :seq_len] # curtail if character tokens are more than the mel spec tokens
|
| 50 |
+
batch, text_len = text.shape[0], text.shape[1]
|
| 51 |
+
text = F.pad(text, (0, seq_len - text_len), value=0)
|
| 52 |
+
|
| 53 |
+
if drop_text: # cfg for text
|
| 54 |
+
text = torch.zeros_like(text)
|
| 55 |
+
|
| 56 |
+
text = self.text_embed(text) # b n -> b n d
|
| 57 |
+
|
| 58 |
+
# possible extra modeling
|
| 59 |
+
if self.extra_modeling:
|
| 60 |
+
# sinus pos emb
|
| 61 |
+
batch_start = torch.zeros((batch,), dtype=torch.long)
|
| 62 |
+
pos_idx = get_pos_embed_indices(batch_start, seq_len, max_pos=self.precompute_max_pos)
|
| 63 |
+
text_pos_embed = self.freqs_cis[pos_idx]
|
| 64 |
+
text = text + text_pos_embed
|
| 65 |
+
|
| 66 |
+
# convnextv2 blocks
|
| 67 |
+
text = self.text_blocks(text)
|
| 68 |
+
|
| 69 |
+
return text
|
| 70 |
+
|
| 71 |
+
|
| 72 |
+
# noised input audio and context mixing embedding
|
| 73 |
+
|
| 74 |
+
|
| 75 |
+
class InputEmbedding(nn.Module):
|
| 76 |
+
def __init__(self, mel_dim, text_dim, out_dim):
|
| 77 |
+
super().__init__()
|
| 78 |
+
self.proj = nn.Linear(mel_dim * 2 + text_dim, out_dim)
|
| 79 |
+
self.conv_pos_embed = ConvPositionEmbedding(dim=out_dim)
|
| 80 |
+
|
| 81 |
+
def forward(self, x: float["b n d"], cond: float["b n d"], text_embed: float["b n d"], drop_audio_cond=False): # noqa: F722
|
| 82 |
+
if drop_audio_cond: # cfg for cond audio
|
| 83 |
+
cond = torch.zeros_like(cond)
|
| 84 |
+
|
| 85 |
+
x = self.proj(torch.cat((x, cond, text_embed), dim=-1))
|
| 86 |
+
x = self.conv_pos_embed(x) + x
|
| 87 |
+
return x
|
| 88 |
+
|
| 89 |
+
|
| 90 |
+
# Transformer backbone using DiT blocks
|
| 91 |
+
|
| 92 |
+
|
| 93 |
+
class DiT(nn.Module):
|
| 94 |
+
def __init__(
|
| 95 |
+
self,
|
| 96 |
+
*,
|
| 97 |
+
dim,
|
| 98 |
+
depth=8,
|
| 99 |
+
heads=8,
|
| 100 |
+
dim_head=64,
|
| 101 |
+
dropout=0.1,
|
| 102 |
+
ff_mult=4,
|
| 103 |
+
mel_dim=100,
|
| 104 |
+
text_num_embeds=256,
|
| 105 |
+
text_dim=None,
|
| 106 |
+
conv_layers=0,
|
| 107 |
+
long_skip_connection=False,
|
| 108 |
+
):
|
| 109 |
+
super().__init__()
|
| 110 |
+
|
| 111 |
+
self.time_embed = TimestepEmbedding(dim)
|
| 112 |
+
if text_dim is None:
|
| 113 |
+
text_dim = mel_dim
|
| 114 |
+
self.text_embed = TextEmbedding(text_num_embeds, text_dim, conv_layers=conv_layers)
|
| 115 |
+
self.input_embed = InputEmbedding(mel_dim, text_dim, dim)
|
| 116 |
+
|
| 117 |
+
self.rotary_embed = RotaryEmbedding(dim_head)
|
| 118 |
+
|
| 119 |
+
self.dim = dim
|
| 120 |
+
self.depth = depth
|
| 121 |
+
|
| 122 |
+
self.transformer_blocks = nn.ModuleList(
|
| 123 |
+
[DiTBlock(dim=dim, heads=heads, dim_head=dim_head, ff_mult=ff_mult, dropout=dropout) for _ in range(depth)]
|
| 124 |
+
)
|
| 125 |
+
self.long_skip_connection = nn.Linear(dim * 2, dim, bias=False) if long_skip_connection else None
|
| 126 |
+
|
| 127 |
+
self.norm_out = AdaLayerNormZero_Final(dim) # final modulation
|
| 128 |
+
self.proj_out = nn.Linear(dim, mel_dim)
|
| 129 |
+
|
| 130 |
+
def forward(
|
| 131 |
+
self,
|
| 132 |
+
x: float["b n d"], # nosied input audio # noqa: F722
|
| 133 |
+
cond: float["b n d"], # masked cond audio # noqa: F722
|
| 134 |
+
text: int["b nt"], # text # noqa: F722
|
| 135 |
+
time: float["b"] | float[""], # time step # noqa: F821 F722
|
| 136 |
+
drop_audio_cond, # cfg for cond audio
|
| 137 |
+
drop_text, # cfg for text
|
| 138 |
+
mask: bool["b n"] | None = None, # noqa: F722
|
| 139 |
+
):
|
| 140 |
+
batch, seq_len = x.shape[0], x.shape[1]
|
| 141 |
+
if time.ndim == 0:
|
| 142 |
+
time = time.repeat(batch)
|
| 143 |
+
|
| 144 |
+
# t: conditioning time, c: context (text + masked cond audio), x: noised input audio
|
| 145 |
+
t = self.time_embed(time)
|
| 146 |
+
text_embed = self.text_embed(text, seq_len, drop_text=drop_text)
|
| 147 |
+
x = self.input_embed(x, cond, text_embed, drop_audio_cond=drop_audio_cond)
|
| 148 |
+
|
| 149 |
+
rope = self.rotary_embed.forward_from_seq_len(seq_len)
|
| 150 |
+
|
| 151 |
+
if self.long_skip_connection is not None:
|
| 152 |
+
residual = x
|
| 153 |
+
|
| 154 |
+
for block in self.transformer_blocks:
|
| 155 |
+
x = block(x, t, mask=mask, rope=rope)
|
| 156 |
+
|
| 157 |
+
if self.long_skip_connection is not None:
|
| 158 |
+
x = self.long_skip_connection(torch.cat((x, residual), dim=-1))
|
| 159 |
+
|
| 160 |
+
x = self.norm_out(x, t)
|
| 161 |
+
output = self.proj_out(x)
|
| 162 |
+
|
| 163 |
+
return output
|
f5-tts/model/backbones/mmdit.py
ADDED
|
@@ -0,0 +1,146 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""
|
| 2 |
+
ein notation:
|
| 3 |
+
b - batch
|
| 4 |
+
n - sequence
|
| 5 |
+
nt - text sequence
|
| 6 |
+
nw - raw wave length
|
| 7 |
+
d - dimension
|
| 8 |
+
"""
|
| 9 |
+
|
| 10 |
+
from __future__ import annotations
|
| 11 |
+
|
| 12 |
+
import torch
|
| 13 |
+
from torch import nn
|
| 14 |
+
|
| 15 |
+
from x_transformers.x_transformers import RotaryEmbedding
|
| 16 |
+
|
| 17 |
+
from f5_tts.model.modules import (
|
| 18 |
+
TimestepEmbedding,
|
| 19 |
+
ConvPositionEmbedding,
|
| 20 |
+
MMDiTBlock,
|
| 21 |
+
AdaLayerNormZero_Final,
|
| 22 |
+
precompute_freqs_cis,
|
| 23 |
+
get_pos_embed_indices,
|
| 24 |
+
)
|
| 25 |
+
|
| 26 |
+
|
| 27 |
+
# text embedding
|
| 28 |
+
|
| 29 |
+
|
| 30 |
+
class TextEmbedding(nn.Module):
|
| 31 |
+
def __init__(self, out_dim, text_num_embeds):
|
| 32 |
+
super().__init__()
|
| 33 |
+
self.text_embed = nn.Embedding(text_num_embeds + 1, out_dim) # will use 0 as filler token
|
| 34 |
+
|
| 35 |
+
self.precompute_max_pos = 1024
|
| 36 |
+
self.register_buffer("freqs_cis", precompute_freqs_cis(out_dim, self.precompute_max_pos), persistent=False)
|
| 37 |
+
|
| 38 |
+
def forward(self, text: int["b nt"], drop_text=False) -> int["b nt d"]: # noqa: F722
|
| 39 |
+
text = text + 1
|
| 40 |
+
if drop_text:
|
| 41 |
+
text = torch.zeros_like(text)
|
| 42 |
+
text = self.text_embed(text)
|
| 43 |
+
|
| 44 |
+
# sinus pos emb
|
| 45 |
+
batch_start = torch.zeros((text.shape[0],), dtype=torch.long)
|
| 46 |
+
batch_text_len = text.shape[1]
|
| 47 |
+
pos_idx = get_pos_embed_indices(batch_start, batch_text_len, max_pos=self.precompute_max_pos)
|
| 48 |
+
text_pos_embed = self.freqs_cis[pos_idx]
|
| 49 |
+
|
| 50 |
+
text = text + text_pos_embed
|
| 51 |
+
|
| 52 |
+
return text
|
| 53 |
+
|
| 54 |
+
|
| 55 |
+
# noised input & masked cond audio embedding
|
| 56 |
+
|
| 57 |
+
|
| 58 |
+
class AudioEmbedding(nn.Module):
|
| 59 |
+
def __init__(self, in_dim, out_dim):
|
| 60 |
+
super().__init__()
|
| 61 |
+
self.linear = nn.Linear(2 * in_dim, out_dim)
|
| 62 |
+
self.conv_pos_embed = ConvPositionEmbedding(out_dim)
|
| 63 |
+
|
| 64 |
+
def forward(self, x: float["b n d"], cond: float["b n d"], drop_audio_cond=False): # noqa: F722
|
| 65 |
+
if drop_audio_cond:
|
| 66 |
+
cond = torch.zeros_like(cond)
|
| 67 |
+
x = torch.cat((x, cond), dim=-1)
|
| 68 |
+
x = self.linear(x)
|
| 69 |
+
x = self.conv_pos_embed(x) + x
|
| 70 |
+
return x
|
| 71 |
+
|
| 72 |
+
|
| 73 |
+
# Transformer backbone using MM-DiT blocks
|
| 74 |
+
|
| 75 |
+
|
| 76 |
+
class MMDiT(nn.Module):
|
| 77 |
+
def __init__(
|
| 78 |
+
self,
|
| 79 |
+
*,
|
| 80 |
+
dim,
|
| 81 |
+
depth=8,
|
| 82 |
+
heads=8,
|
| 83 |
+
dim_head=64,
|
| 84 |
+
dropout=0.1,
|
| 85 |
+
ff_mult=4,
|
| 86 |
+
text_num_embeds=256,
|
| 87 |
+
mel_dim=100,
|
| 88 |
+
):
|
| 89 |
+
super().__init__()
|
| 90 |
+
|
| 91 |
+
self.time_embed = TimestepEmbedding(dim)
|
| 92 |
+
self.text_embed = TextEmbedding(dim, text_num_embeds)
|
| 93 |
+
self.audio_embed = AudioEmbedding(mel_dim, dim)
|
| 94 |
+
|
| 95 |
+
self.rotary_embed = RotaryEmbedding(dim_head)
|
| 96 |
+
|
| 97 |
+
self.dim = dim
|
| 98 |
+
self.depth = depth
|
| 99 |
+
|
| 100 |
+
self.transformer_blocks = nn.ModuleList(
|
| 101 |
+
[
|
| 102 |
+
MMDiTBlock(
|
| 103 |
+
dim=dim,
|
| 104 |
+
heads=heads,
|
| 105 |
+
dim_head=dim_head,
|
| 106 |
+
dropout=dropout,
|
| 107 |
+
ff_mult=ff_mult,
|
| 108 |
+
context_pre_only=i == depth - 1,
|
| 109 |
+
)
|
| 110 |
+
for i in range(depth)
|
| 111 |
+
]
|
| 112 |
+
)
|
| 113 |
+
self.norm_out = AdaLayerNormZero_Final(dim) # final modulation
|
| 114 |
+
self.proj_out = nn.Linear(dim, mel_dim)
|
| 115 |
+
|
| 116 |
+
def forward(
|
| 117 |
+
self,
|
| 118 |
+
x: float["b n d"], # nosied input audio # noqa: F722
|
| 119 |
+
cond: float["b n d"], # masked cond audio # noqa: F722
|
| 120 |
+
text: int["b nt"], # text # noqa: F722
|
| 121 |
+
time: float["b"] | float[""], # time step # noqa: F821 F722
|
| 122 |
+
drop_audio_cond, # cfg for cond audio
|
| 123 |
+
drop_text, # cfg for text
|
| 124 |
+
mask: bool["b n"] | None = None, # noqa: F722
|
| 125 |
+
):
|
| 126 |
+
batch = x.shape[0]
|
| 127 |
+
if time.ndim == 0:
|
| 128 |
+
time = time.repeat(batch)
|
| 129 |
+
|
| 130 |
+
# t: conditioning (time), c: context (text + masked cond audio), x: noised input audio
|
| 131 |
+
t = self.time_embed(time)
|
| 132 |
+
c = self.text_embed(text, drop_text=drop_text)
|
| 133 |
+
x = self.audio_embed(x, cond, drop_audio_cond=drop_audio_cond)
|
| 134 |
+
|
| 135 |
+
seq_len = x.shape[1]
|
| 136 |
+
text_len = text.shape[1]
|
| 137 |
+
rope_audio = self.rotary_embed.forward_from_seq_len(seq_len)
|
| 138 |
+
rope_text = self.rotary_embed.forward_from_seq_len(text_len)
|
| 139 |
+
|
| 140 |
+
for block in self.transformer_blocks:
|
| 141 |
+
c, x = block(x, c, t, mask=mask, rope=rope_audio, c_rope=rope_text)
|
| 142 |
+
|
| 143 |
+
x = self.norm_out(x, t)
|
| 144 |
+
output = self.proj_out(x)
|
| 145 |
+
|
| 146 |
+
return output
|
f5-tts/model/backbones/unett.py
ADDED
|
@@ -0,0 +1,219 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""
|
| 2 |
+
ein notation:
|
| 3 |
+
b - batch
|
| 4 |
+
n - sequence
|
| 5 |
+
nt - text sequence
|
| 6 |
+
nw - raw wave length
|
| 7 |
+
d - dimension
|
| 8 |
+
"""
|
| 9 |
+
|
| 10 |
+
from __future__ import annotations
|
| 11 |
+
from typing import Literal
|
| 12 |
+
|
| 13 |
+
import torch
|
| 14 |
+
from torch import nn
|
| 15 |
+
import torch.nn.functional as F
|
| 16 |
+
|
| 17 |
+
from x_transformers import RMSNorm
|
| 18 |
+
from x_transformers.x_transformers import RotaryEmbedding
|
| 19 |
+
|
| 20 |
+
from f5_tts.model.modules import (
|
| 21 |
+
TimestepEmbedding,
|
| 22 |
+
ConvNeXtV2Block,
|
| 23 |
+
ConvPositionEmbedding,
|
| 24 |
+
Attention,
|
| 25 |
+
AttnProcessor,
|
| 26 |
+
FeedForward,
|
| 27 |
+
precompute_freqs_cis,
|
| 28 |
+
get_pos_embed_indices,
|
| 29 |
+
)
|
| 30 |
+
|
| 31 |
+
|
| 32 |
+
# Text embedding
|
| 33 |
+
|
| 34 |
+
|
| 35 |
+
class TextEmbedding(nn.Module):
|
| 36 |
+
def __init__(self, text_num_embeds, text_dim, conv_layers=0, conv_mult=2):
|
| 37 |
+
super().__init__()
|
| 38 |
+
self.text_embed = nn.Embedding(text_num_embeds + 1, text_dim) # use 0 as filler token
|
| 39 |
+
|
| 40 |
+
if conv_layers > 0:
|
| 41 |
+
self.extra_modeling = True
|
| 42 |
+
self.precompute_max_pos = 4096 # ~44s of 24khz audio
|
| 43 |
+
self.register_buffer("freqs_cis", precompute_freqs_cis(text_dim, self.precompute_max_pos), persistent=False)
|
| 44 |
+
self.text_blocks = nn.Sequential(
|
| 45 |
+
*[ConvNeXtV2Block(text_dim, text_dim * conv_mult) for _ in range(conv_layers)]
|
| 46 |
+
)
|
| 47 |
+
else:
|
| 48 |
+
self.extra_modeling = False
|
| 49 |
+
|
| 50 |
+
def forward(self, text: int["b nt"], seq_len, drop_text=False): # noqa: F722
|
| 51 |
+
text = text + 1 # use 0 as filler token. preprocess of batch pad -1, see list_str_to_idx()
|
| 52 |
+
text = text[:, :seq_len] # curtail if character tokens are more than the mel spec tokens
|
| 53 |
+
batch, text_len = text.shape[0], text.shape[1]
|
| 54 |
+
text = F.pad(text, (0, seq_len - text_len), value=0)
|
| 55 |
+
|
| 56 |
+
if drop_text: # cfg for text
|
| 57 |
+
text = torch.zeros_like(text)
|
| 58 |
+
|
| 59 |
+
text = self.text_embed(text) # b n -> b n d
|
| 60 |
+
|
| 61 |
+
# possible extra modeling
|
| 62 |
+
if self.extra_modeling:
|
| 63 |
+
# sinus pos emb
|
| 64 |
+
batch_start = torch.zeros((batch,), dtype=torch.long)
|
| 65 |
+
pos_idx = get_pos_embed_indices(batch_start, seq_len, max_pos=self.precompute_max_pos)
|
| 66 |
+
text_pos_embed = self.freqs_cis[pos_idx]
|
| 67 |
+
text = text + text_pos_embed
|
| 68 |
+
|
| 69 |
+
# convnextv2 blocks
|
| 70 |
+
text = self.text_blocks(text)
|
| 71 |
+
|
| 72 |
+
return text
|
| 73 |
+
|
| 74 |
+
|
| 75 |
+
# noised input audio and context mixing embedding
|
| 76 |
+
|
| 77 |
+
|
| 78 |
+
class InputEmbedding(nn.Module):
|
| 79 |
+
def __init__(self, mel_dim, text_dim, out_dim):
|
| 80 |
+
super().__init__()
|
| 81 |
+
self.proj = nn.Linear(mel_dim * 2 + text_dim, out_dim)
|
| 82 |
+
self.conv_pos_embed = ConvPositionEmbedding(dim=out_dim)
|
| 83 |
+
|
| 84 |
+
def forward(self, x: float["b n d"], cond: float["b n d"], text_embed: float["b n d"], drop_audio_cond=False): # noqa: F722
|
| 85 |
+
if drop_audio_cond: # cfg for cond audio
|
| 86 |
+
cond = torch.zeros_like(cond)
|
| 87 |
+
|
| 88 |
+
x = self.proj(torch.cat((x, cond, text_embed), dim=-1))
|
| 89 |
+
x = self.conv_pos_embed(x) + x
|
| 90 |
+
return x
|
| 91 |
+
|
| 92 |
+
|
| 93 |
+
# Flat UNet Transformer backbone
|
| 94 |
+
|
| 95 |
+
|
| 96 |
+
class UNetT(nn.Module):
|
| 97 |
+
def __init__(
|
| 98 |
+
self,
|
| 99 |
+
*,
|
| 100 |
+
dim,
|
| 101 |
+
depth=8,
|
| 102 |
+
heads=8,
|
| 103 |
+
dim_head=64,
|
| 104 |
+
dropout=0.1,
|
| 105 |
+
ff_mult=4,
|
| 106 |
+
mel_dim=100,
|
| 107 |
+
text_num_embeds=256,
|
| 108 |
+
text_dim=None,
|
| 109 |
+
conv_layers=0,
|
| 110 |
+
skip_connect_type: Literal["add", "concat", "none"] = "concat",
|
| 111 |
+
):
|
| 112 |
+
super().__init__()
|
| 113 |
+
assert depth % 2 == 0, "UNet-Transformer's depth should be even."
|
| 114 |
+
|
| 115 |
+
self.time_embed = TimestepEmbedding(dim)
|
| 116 |
+
if text_dim is None:
|
| 117 |
+
text_dim = mel_dim
|
| 118 |
+
self.text_embed = TextEmbedding(text_num_embeds, text_dim, conv_layers=conv_layers)
|
| 119 |
+
self.input_embed = InputEmbedding(mel_dim, text_dim, dim)
|
| 120 |
+
|
| 121 |
+
self.rotary_embed = RotaryEmbedding(dim_head)
|
| 122 |
+
|
| 123 |
+
# transformer layers & skip connections
|
| 124 |
+
|
| 125 |
+
self.dim = dim
|
| 126 |
+
self.skip_connect_type = skip_connect_type
|
| 127 |
+
needs_skip_proj = skip_connect_type == "concat"
|
| 128 |
+
|
| 129 |
+
self.depth = depth
|
| 130 |
+
self.layers = nn.ModuleList([])
|
| 131 |
+
|
| 132 |
+
for idx in range(depth):
|
| 133 |
+
is_later_half = idx >= (depth // 2)
|
| 134 |
+
|
| 135 |
+
attn_norm = RMSNorm(dim)
|
| 136 |
+
attn = Attention(
|
| 137 |
+
processor=AttnProcessor(),
|
| 138 |
+
dim=dim,
|
| 139 |
+
heads=heads,
|
| 140 |
+
dim_head=dim_head,
|
| 141 |
+
dropout=dropout,
|
| 142 |
+
)
|
| 143 |
+
|
| 144 |
+
ff_norm = RMSNorm(dim)
|
| 145 |
+
ff = FeedForward(dim=dim, mult=ff_mult, dropout=dropout, approximate="tanh")
|
| 146 |
+
|
| 147 |
+
skip_proj = nn.Linear(dim * 2, dim, bias=False) if needs_skip_proj and is_later_half else None
|
| 148 |
+
|
| 149 |
+
self.layers.append(
|
| 150 |
+
nn.ModuleList(
|
| 151 |
+
[
|
| 152 |
+
skip_proj,
|
| 153 |
+
attn_norm,
|
| 154 |
+
attn,
|
| 155 |
+
ff_norm,
|
| 156 |
+
ff,
|
| 157 |
+
]
|
| 158 |
+
)
|
| 159 |
+
)
|
| 160 |
+
|
| 161 |
+
self.norm_out = RMSNorm(dim)
|
| 162 |
+
self.proj_out = nn.Linear(dim, mel_dim)
|
| 163 |
+
|
| 164 |
+
def forward(
|
| 165 |
+
self,
|
| 166 |
+
x: float["b n d"], # nosied input audio # noqa: F722
|
| 167 |
+
cond: float["b n d"], # masked cond audio # noqa: F722
|
| 168 |
+
text: int["b nt"], # text # noqa: F722
|
| 169 |
+
time: float["b"] | float[""], # time step # noqa: F821 F722
|
| 170 |
+
drop_audio_cond, # cfg for cond audio
|
| 171 |
+
drop_text, # cfg for text
|
| 172 |
+
mask: bool["b n"] | None = None, # noqa: F722
|
| 173 |
+
):
|
| 174 |
+
batch, seq_len = x.shape[0], x.shape[1]
|
| 175 |
+
if time.ndim == 0:
|
| 176 |
+
time = time.repeat(batch)
|
| 177 |
+
|
| 178 |
+
# t: conditioning time, c: context (text + masked cond audio), x: noised input audio
|
| 179 |
+
t = self.time_embed(time)
|
| 180 |
+
text_embed = self.text_embed(text, seq_len, drop_text=drop_text)
|
| 181 |
+
x = self.input_embed(x, cond, text_embed, drop_audio_cond=drop_audio_cond)
|
| 182 |
+
|
| 183 |
+
# postfix time t to input x, [b n d] -> [b n+1 d]
|
| 184 |
+
x = torch.cat([t.unsqueeze(1), x], dim=1) # pack t to x
|
| 185 |
+
if mask is not None:
|
| 186 |
+
mask = F.pad(mask, (1, 0), value=1)
|
| 187 |
+
|
| 188 |
+
rope = self.rotary_embed.forward_from_seq_len(seq_len + 1)
|
| 189 |
+
|
| 190 |
+
# flat unet transformer
|
| 191 |
+
skip_connect_type = self.skip_connect_type
|
| 192 |
+
skips = []
|
| 193 |
+
for idx, (maybe_skip_proj, attn_norm, attn, ff_norm, ff) in enumerate(self.layers):
|
| 194 |
+
layer = idx + 1
|
| 195 |
+
|
| 196 |
+
# skip connection logic
|
| 197 |
+
is_first_half = layer <= (self.depth // 2)
|
| 198 |
+
is_later_half = not is_first_half
|
| 199 |
+
|
| 200 |
+
if is_first_half:
|
| 201 |
+
skips.append(x)
|
| 202 |
+
|
| 203 |
+
if is_later_half:
|
| 204 |
+
skip = skips.pop()
|
| 205 |
+
if skip_connect_type == "concat":
|
| 206 |
+
x = torch.cat((x, skip), dim=-1)
|
| 207 |
+
x = maybe_skip_proj(x)
|
| 208 |
+
elif skip_connect_type == "add":
|
| 209 |
+
x = x + skip
|
| 210 |
+
|
| 211 |
+
# attention and feedforward blocks
|
| 212 |
+
x = attn(attn_norm(x), rope=rope, mask=mask) + x
|
| 213 |
+
x = ff(ff_norm(x)) + x
|
| 214 |
+
|
| 215 |
+
assert len(skips) == 0
|
| 216 |
+
|
| 217 |
+
x = self.norm_out(x)[:, 1:, :] # unpack t from x
|
| 218 |
+
|
| 219 |
+
return self.proj_out(x)
|