Initial commit
This commit is contained in:
王新升
@@ -0,0 +1,2 @@
|
||||
from .bs_roformer import BSRoformer
|
||||
from .mel_band_roformer import MelBandRoformer
|
||||
@@ -0,0 +1,126 @@
|
||||
from functools import wraps
|
||||
from packaging import version
|
||||
from collections import namedtuple
|
||||
|
||||
import os
|
||||
import torch
|
||||
from torch import nn, einsum
|
||||
import torch.nn.functional as F
|
||||
|
||||
from einops import rearrange, reduce
|
||||
|
||||
# constants
|
||||
|
||||
FlashAttentionConfig = namedtuple('FlashAttentionConfig', ['enable_flash', 'enable_math', 'enable_mem_efficient'])
|
||||
|
||||
# helpers
|
||||
|
||||
def exists(val):
|
||||
return val is not None
|
||||
|
||||
def default(v, d):
|
||||
return v if exists(v) else d
|
||||
|
||||
def once(fn):
|
||||
called = False
|
||||
@wraps(fn)
|
||||
def inner(x):
|
||||
nonlocal called
|
||||
if called:
|
||||
return
|
||||
called = True
|
||||
return fn(x)
|
||||
return inner
|
||||
|
||||
print_once = once(print)
|
||||
|
||||
# main class
|
||||
|
||||
class Attend(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
dropout = 0.,
|
||||
flash = False,
|
||||
scale = None
|
||||
):
|
||||
super().__init__()
|
||||
self.scale = scale
|
||||
self.dropout = dropout
|
||||
self.attn_dropout = nn.Dropout(dropout)
|
||||
|
||||
self.flash = flash
|
||||
assert not (flash and version.parse(torch.__version__) < version.parse('2.0.0')), 'in order to use flash attention, you must be using pytorch 2.0 or above'
|
||||
|
||||
# determine efficient attention configs for cuda and cpu
|
||||
|
||||
self.cpu_config = FlashAttentionConfig(True, True, True)
|
||||
self.cuda_config = None
|
||||
|
||||
if not torch.cuda.is_available() or not flash:
|
||||
return
|
||||
|
||||
device_properties = torch.cuda.get_device_properties(torch.device('cuda'))
|
||||
device_version = version.parse(f'{device_properties.major}.{device_properties.minor}')
|
||||
|
||||
if device_version >= version.parse('8.0'):
|
||||
if os.name == 'nt':
|
||||
print_once('Windows OS detected, using math or mem efficient attention if input tensor is on cuda')
|
||||
self.cuda_config = FlashAttentionConfig(False, True, True)
|
||||
else:
|
||||
print_once('GPU Compute Capability equal or above 8.0, using flash attention if input tensor is on cuda')
|
||||
self.cuda_config = FlashAttentionConfig(True, False, False)
|
||||
else:
|
||||
print_once('GPU Compute Capability below 8.0, using math or mem efficient attention if input tensor is on cuda')
|
||||
self.cuda_config = FlashAttentionConfig(False, True, True)
|
||||
|
||||
def flash_attn(self, q, k, v):
|
||||
_, heads, q_len, _, k_len, is_cuda, device = *q.shape, k.shape[-2], q.is_cuda, q.device
|
||||
|
||||
if exists(self.scale):
|
||||
default_scale = q.shape[-1] ** -0.5
|
||||
q = q * (self.scale / default_scale)
|
||||
|
||||
# Check if there is a compatible device for flash attention
|
||||
|
||||
config = self.cuda_config if is_cuda else self.cpu_config
|
||||
|
||||
# pytorch 2.0 flash attn: q, k, v, mask, dropout, softmax_scale
|
||||
|
||||
with torch.backends.cuda.sdp_kernel(**config._asdict()):
|
||||
out = F.scaled_dot_product_attention(
|
||||
q, k, v,
|
||||
dropout_p = self.dropout if self.training else 0.
|
||||
)
|
||||
|
||||
return out
|
||||
|
||||
def forward(self, q, k, v):
|
||||
"""
|
||||
einstein notation
|
||||
b - batch
|
||||
h - heads
|
||||
n, i, j - sequence length (base sequence length, source, target)
|
||||
d - feature dimension
|
||||
"""
|
||||
|
||||
q_len, k_len, device = q.shape[-2], k.shape[-2], q.device
|
||||
|
||||
scale = default(self.scale, q.shape[-1] ** -0.5)
|
||||
|
||||
if self.flash:
|
||||
return self.flash_attn(q, k, v)
|
||||
|
||||
# similarity
|
||||
|
||||
sim = einsum(f"b h i d, b h j d -> b h i j", q, k) * scale
|
||||
|
||||
# attention
|
||||
|
||||
attn = sim.softmax(dim=-1)
|
||||
attn = self.attn_dropout(attn)
|
||||
|
||||
# aggregate values
|
||||
|
||||
out = einsum(f"b h i j, b h j d -> b h i d", attn, v)
|
||||
|
||||
return out
|
||||
@@ -0,0 +1,145 @@
|
||||
from functools import wraps
|
||||
from packaging import version
|
||||
from collections import namedtuple
|
||||
|
||||
import os
|
||||
import torch
|
||||
from torch import nn, einsum
|
||||
import torch.nn.functional as F
|
||||
|
||||
from einops import rearrange, reduce
|
||||
|
||||
def _print_once(msg):
|
||||
printed = False
|
||||
@wraps(print)
|
||||
def inner():
|
||||
nonlocal printed
|
||||
if not printed:
|
||||
print(msg)
|
||||
printed = True
|
||||
return inner
|
||||
|
||||
try:
|
||||
from sageattention import sageattn
|
||||
_has_sage_attention = True
|
||||
# _print_sage_found = _print_once("SageAttention found. Will be used when flash=True.")
|
||||
# _print_sage_found()
|
||||
except ImportError:
|
||||
_has_sage_attention = False
|
||||
_print_sage_not_found = _print_once("SageAttention not found. Will fall back to PyTorch SDPA (if available) or manual einsum.")
|
||||
_print_sage_not_found()
|
||||
|
||||
# helpers
|
||||
def exists(val):
|
||||
return val is not None
|
||||
|
||||
def default(v, d):
|
||||
return v if exists(v) else d
|
||||
|
||||
# main class
|
||||
class Attend(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
dropout = 0.,
|
||||
flash = False, # If True, attempts to use SageAttention or PyTorch SDPA
|
||||
scale = None
|
||||
):
|
||||
super().__init__()
|
||||
self.scale = scale # Store the scale if needed for einsum path
|
||||
self.dropout = dropout # Store dropout if needed for einsum/SDPA path
|
||||
|
||||
# Determine which attention mechanism to *try* first
|
||||
self.use_sage = flash and _has_sage_attention
|
||||
self.use_pytorch_sdpa = False
|
||||
self._sdpa_checked = False # Flag to check PyTorch version only once
|
||||
|
||||
if flash and not self.use_sage:
|
||||
# Only consider PyTorch SDPA if Sage isn't available/chosen
|
||||
if not self._sdpa_checked:
|
||||
if version.parse(torch.__version__) >= version.parse('2.0.0'):
|
||||
self.use_pytorch_sdpa = True
|
||||
_print_sdpa_used = _print_once("Using PyTorch SDPA backend (FlashAttention-2, Memory-Efficient, or Math).")
|
||||
_print_sdpa_used()
|
||||
else:
|
||||
_print_fallback_einsum = _print_once("Flash attention requested but Pytorch < 2.0 and SageAttention not found. Falling back to einsum.")
|
||||
_print_fallback_einsum()
|
||||
self._sdpa_checked = True
|
||||
|
||||
# Dropout layer for manual einsum implementation ONLY
|
||||
# SDPA and SageAttention handle dropout differently (or not at all in Sage's base API)
|
||||
self.attn_dropout = nn.Dropout(dropout)
|
||||
|
||||
def forward(self, q, k, v):
|
||||
"""
|
||||
einstein notation
|
||||
b - batch
|
||||
h - heads
|
||||
n, i, j - sequence length (base sequence length, source, target)
|
||||
d - feature dimension
|
||||
|
||||
Input tensors q, k, v expected in shape: (batch, heads, seq_len, dim_head) -> HND layout
|
||||
"""
|
||||
q_len, k_len, device = q.shape[-2], k.shape[-2], q.device
|
||||
|
||||
# --- Priority 1: SageAttention ---
|
||||
if self.use_sage:
|
||||
# Assumes q, k, v are FP16/BF16 (handled by autocast upstream)
|
||||
# Assumes scale is handled internally by sageattn
|
||||
# Assumes dropout is NOT handled by sageattn kernel
|
||||
# is_causal=False based on how Attend is called in mel_band_roformer
|
||||
out = sageattn(q, k, v, tensor_layout='HND', is_causal=False)
|
||||
return out
|
||||
try:
|
||||
return out
|
||||
# print("Attempting SageAttention") # Optional: for debugging
|
||||
out = sageattn(q, k, v, tensor_layout='HND', is_causal=False)
|
||||
return out
|
||||
except Exception as e:
|
||||
print(f"SageAttention failed with error: {e}. Falling back.")
|
||||
self.use_sage = False # Don't try Sage again if it failed once
|
||||
# Decide fallback: Check if PyTorch SDPA is an option
|
||||
if not self._sdpa_checked:
|
||||
if version.parse(torch.__version__) >= version.parse('2.0.0'):
|
||||
self.use_pytorch_sdpa = True
|
||||
_print_sdpa_fallback = _print_once("Falling back to PyTorch SDPA.")
|
||||
_print_sdpa_fallback()
|
||||
else:
|
||||
_print_einsum_fallback = _print_once("Falling back to einsum.")
|
||||
_print_einsum_fallback()
|
||||
self._sdpa_checked = True
|
||||
|
||||
|
||||
# --- Priority 2: PyTorch SDPA ---
|
||||
if self.use_pytorch_sdpa:
|
||||
# Use PyTorch's Scaled Dot Product Attention (SDPA)
|
||||
# It handles scaling and dropout internally.
|
||||
try:
|
||||
# print("Attempting PyTorch SDPA") # Optional: for debugging
|
||||
# Let PyTorch choose the best backend (Flash V2, Mem Efficient, Math)
|
||||
with torch.backends.cuda.sdp_kernel(enable_flash=True, enable_math=True, enable_mem_efficient=True):
|
||||
out = F.scaled_dot_product_attention(
|
||||
q, k, v,
|
||||
attn_mask=None, # Assuming no explicit mask needed here
|
||||
dropout_p = self.dropout if self.training else 0.,
|
||||
is_causal=False # Assuming not needed based on usage context
|
||||
)
|
||||
return out
|
||||
except Exception as e:
|
||||
print(f"PyTorch SDPA failed with error: {e}. Falling back to einsum.")
|
||||
self.use_pytorch_sdpa = False # Fallback to einsum on error
|
||||
|
||||
|
||||
# Calculate scale
|
||||
scale = default(self.scale, q.shape[-1] ** -0.5)
|
||||
|
||||
# similarity
|
||||
sim = einsum(f"b h i d, b h j d -> b h i j", q, k) * scale
|
||||
|
||||
# attention
|
||||
attn = sim.softmax(dim=-1)
|
||||
attn = self.attn_dropout(attn) # Apply dropout ONLY in einsum path
|
||||
|
||||
# aggregate values
|
||||
out = einsum(f"b h i j, b h j d -> b h i d", attn, v)
|
||||
|
||||
return out
|
||||
@@ -0,0 +1,658 @@
|
||||
from functools import partial
|
||||
|
||||
import torch
|
||||
from torch import nn, einsum, Tensor
|
||||
from torch.nn import Module, ModuleList
|
||||
import torch.nn.functional as F
|
||||
|
||||
from .attend import Attend
|
||||
try:
|
||||
from .attend_sage import Attend as AttendSage
|
||||
except:
|
||||
pass
|
||||
from torch.utils.checkpoint import checkpoint
|
||||
|
||||
from beartype.typing import Tuple, Optional, List, Callable
|
||||
from beartype import beartype
|
||||
|
||||
from rotary_embedding_torch import RotaryEmbedding
|
||||
|
||||
from einops import rearrange, pack, unpack
|
||||
from einops.layers.torch import Rearrange
|
||||
|
||||
# helper functions
|
||||
|
||||
def exists(val):
|
||||
return val is not None
|
||||
|
||||
|
||||
def default(v, d):
|
||||
return v if exists(v) else d
|
||||
|
||||
|
||||
def pack_one(t, pattern):
|
||||
return pack([t], pattern)
|
||||
|
||||
|
||||
def unpack_one(t, ps, pattern):
|
||||
return unpack(t, ps, pattern)[0]
|
||||
|
||||
|
||||
# norm
|
||||
|
||||
def l2norm(t):
|
||||
return F.normalize(t, dim = -1, p = 2)
|
||||
|
||||
|
||||
class RMSNorm(Module):
|
||||
def __init__(self, dim):
|
||||
super().__init__()
|
||||
self.scale = dim ** 0.5
|
||||
self.gamma = nn.Parameter(torch.ones(dim))
|
||||
|
||||
def forward(self, x):
|
||||
return F.normalize(x, dim=-1) * self.scale * self.gamma
|
||||
|
||||
|
||||
# attention
|
||||
|
||||
class FeedForward(Module):
|
||||
def __init__(
|
||||
self,
|
||||
dim,
|
||||
mult=4,
|
||||
dropout=0.
|
||||
):
|
||||
super().__init__()
|
||||
dim_inner = int(dim * mult)
|
||||
self.net = nn.Sequential(
|
||||
RMSNorm(dim),
|
||||
nn.Linear(dim, dim_inner),
|
||||
nn.GELU(),
|
||||
nn.Dropout(dropout),
|
||||
nn.Linear(dim_inner, dim),
|
||||
nn.Dropout(dropout)
|
||||
)
|
||||
|
||||
def forward(self, x):
|
||||
return self.net(x)
|
||||
|
||||
|
||||
class Attention(Module):
|
||||
def __init__(
|
||||
self,
|
||||
dim,
|
||||
heads=8,
|
||||
dim_head=64,
|
||||
dropout=0.,
|
||||
rotary_embed=None,
|
||||
flash=True,
|
||||
sage_attention=False,
|
||||
):
|
||||
super().__init__()
|
||||
self.heads = heads
|
||||
self.scale = dim_head ** -0.5
|
||||
dim_inner = heads * dim_head
|
||||
|
||||
self.rotary_embed = rotary_embed
|
||||
|
||||
if sage_attention:
|
||||
self.attend = AttendSage(flash=flash, dropout=dropout)
|
||||
else:
|
||||
self.attend = Attend(flash=flash, dropout=dropout)
|
||||
|
||||
self.norm = RMSNorm(dim)
|
||||
self.to_qkv = nn.Linear(dim, dim_inner * 3, bias=False)
|
||||
|
||||
self.to_gates = nn.Linear(dim, heads)
|
||||
|
||||
self.to_out = nn.Sequential(
|
||||
nn.Linear(dim_inner, dim, bias=False),
|
||||
nn.Dropout(dropout)
|
||||
)
|
||||
|
||||
def forward(self, x):
|
||||
x = self.norm(x)
|
||||
|
||||
q, k, v = rearrange(self.to_qkv(x), 'b n (qkv h d) -> qkv b h n d', qkv=3, h=self.heads)
|
||||
|
||||
if exists(self.rotary_embed):
|
||||
q = self.rotary_embed.rotate_queries_or_keys(q)
|
||||
k = self.rotary_embed.rotate_queries_or_keys(k)
|
||||
|
||||
out = self.attend(q, k, v)
|
||||
|
||||
gates = self.to_gates(x)
|
||||
out = out * rearrange(gates, 'b n h -> b h n 1').sigmoid()
|
||||
|
||||
out = rearrange(out, 'b h n d -> b n (h d)')
|
||||
return self.to_out(out)
|
||||
|
||||
|
||||
class LinearAttention(Module):
|
||||
"""
|
||||
this flavor of linear attention proposed in https://arxiv.org/abs/2106.09681 by El-Nouby et al.
|
||||
"""
|
||||
|
||||
@beartype
|
||||
def __init__(
|
||||
self,
|
||||
*,
|
||||
dim,
|
||||
dim_head=32,
|
||||
heads=8,
|
||||
scale=8,
|
||||
flash=False,
|
||||
dropout=0.,
|
||||
sage_attention=False,
|
||||
):
|
||||
super().__init__()
|
||||
dim_inner = dim_head * heads
|
||||
self.norm = RMSNorm(dim)
|
||||
|
||||
self.to_qkv = nn.Sequential(
|
||||
nn.Linear(dim, dim_inner * 3, bias=False),
|
||||
Rearrange('b n (qkv h d) -> qkv b h d n', qkv=3, h=heads)
|
||||
)
|
||||
|
||||
self.temperature = nn.Parameter(torch.ones(heads, 1, 1))
|
||||
|
||||
if sage_attention:
|
||||
self.attend = AttendSage(
|
||||
scale=scale,
|
||||
dropout=dropout,
|
||||
flash=flash
|
||||
)
|
||||
else:
|
||||
self.attend = Attend(
|
||||
scale=scale,
|
||||
dropout=dropout,
|
||||
flash=flash
|
||||
)
|
||||
|
||||
self.to_out = nn.Sequential(
|
||||
Rearrange('b h d n -> b n (h d)'),
|
||||
nn.Linear(dim_inner, dim, bias=False)
|
||||
)
|
||||
|
||||
def forward(
|
||||
self,
|
||||
x
|
||||
):
|
||||
x = self.norm(x)
|
||||
|
||||
q, k, v = self.to_qkv(x)
|
||||
|
||||
q, k = map(l2norm, (q, k))
|
||||
q = q * self.temperature.exp()
|
||||
|
||||
out = self.attend(q, k, v)
|
||||
|
||||
return self.to_out(out)
|
||||
|
||||
|
||||
class Transformer(Module):
|
||||
def __init__(
|
||||
self,
|
||||
*,
|
||||
dim,
|
||||
depth,
|
||||
dim_head=64,
|
||||
heads=8,
|
||||
attn_dropout=0.,
|
||||
ff_dropout=0.,
|
||||
ff_mult=4,
|
||||
norm_output=True,
|
||||
rotary_embed=None,
|
||||
flash_attn=True,
|
||||
linear_attn=False,
|
||||
sage_attention=False,
|
||||
):
|
||||
super().__init__()
|
||||
self.layers = ModuleList([])
|
||||
|
||||
for _ in range(depth):
|
||||
if linear_attn:
|
||||
attn = LinearAttention(
|
||||
dim=dim,
|
||||
dim_head=dim_head,
|
||||
heads=heads,
|
||||
dropout=attn_dropout,
|
||||
flash=flash_attn,
|
||||
sage_attention=sage_attention
|
||||
)
|
||||
else:
|
||||
attn = Attention(
|
||||
dim=dim,
|
||||
dim_head=dim_head,
|
||||
heads=heads,
|
||||
dropout=attn_dropout,
|
||||
rotary_embed=rotary_embed,
|
||||
flash=flash_attn,
|
||||
sage_attention=sage_attention
|
||||
)
|
||||
|
||||
self.layers.append(ModuleList([
|
||||
attn,
|
||||
FeedForward(dim=dim, mult=ff_mult, dropout=ff_dropout)
|
||||
]))
|
||||
|
||||
self.norm = RMSNorm(dim) if norm_output else nn.Identity()
|
||||
|
||||
def forward(self, x):
|
||||
|
||||
for attn, ff in self.layers:
|
||||
x = attn(x) + x
|
||||
x = ff(x) + x
|
||||
|
||||
return self.norm(x)
|
||||
|
||||
|
||||
# bandsplit module
|
||||
|
||||
class BandSplit(Module):
|
||||
@beartype
|
||||
def __init__(
|
||||
self,
|
||||
dim,
|
||||
dim_inputs: Tuple[int, ...]
|
||||
):
|
||||
super().__init__()
|
||||
self.dim_inputs = dim_inputs
|
||||
self.to_features = ModuleList([])
|
||||
|
||||
for dim_in in dim_inputs:
|
||||
net = nn.Sequential(
|
||||
RMSNorm(dim_in),
|
||||
nn.Linear(dim_in, dim)
|
||||
)
|
||||
|
||||
self.to_features.append(net)
|
||||
|
||||
def forward(self, x):
|
||||
x = x.split(self.dim_inputs, dim=-1)
|
||||
|
||||
outs = []
|
||||
for split_input, to_feature in zip(x, self.to_features):
|
||||
split_output = to_feature(split_input)
|
||||
outs.append(split_output)
|
||||
|
||||
return torch.stack(outs, dim=-2)
|
||||
|
||||
|
||||
def MLP(
|
||||
dim_in,
|
||||
dim_out,
|
||||
dim_hidden=None,
|
||||
depth=1,
|
||||
activation=nn.Tanh
|
||||
):
|
||||
dim_hidden = default(dim_hidden, dim_in)
|
||||
|
||||
net = []
|
||||
dims = (dim_in, *((dim_hidden,) * (depth - 1)), dim_out)
|
||||
|
||||
for ind, (layer_dim_in, layer_dim_out) in enumerate(zip(dims[:-1], dims[1:])):
|
||||
is_last = ind == (len(dims) - 2)
|
||||
|
||||
net.append(nn.Linear(layer_dim_in, layer_dim_out))
|
||||
|
||||
if is_last:
|
||||
continue
|
||||
|
||||
net.append(activation())
|
||||
|
||||
return nn.Sequential(*net)
|
||||
|
||||
|
||||
class MaskEstimator(Module):
|
||||
@beartype
|
||||
def __init__(
|
||||
self,
|
||||
dim,
|
||||
dim_inputs: Tuple[int, ...],
|
||||
depth,
|
||||
mlp_expansion_factor=4
|
||||
):
|
||||
super().__init__()
|
||||
self.dim_inputs = dim_inputs
|
||||
self.to_freqs = ModuleList([])
|
||||
dim_hidden = dim * mlp_expansion_factor
|
||||
|
||||
for dim_in in dim_inputs:
|
||||
net = []
|
||||
|
||||
mlp = nn.Sequential(
|
||||
MLP(dim, dim_in * 2, dim_hidden=dim_hidden, depth=depth),
|
||||
nn.GLU(dim=-1)
|
||||
)
|
||||
|
||||
self.to_freqs.append(mlp)
|
||||
|
||||
def forward(self, x):
|
||||
x = x.unbind(dim=-2)
|
||||
|
||||
outs = []
|
||||
|
||||
for band_features, mlp in zip(x, self.to_freqs):
|
||||
freq_out = mlp(band_features)
|
||||
outs.append(freq_out)
|
||||
|
||||
return torch.cat(outs, dim=-1)
|
||||
|
||||
|
||||
# main class
|
||||
|
||||
DEFAULT_FREQS_PER_BANDS = (
|
||||
2, 2, 2, 2, 2, 2, 2, 2, 2, 2,
|
||||
2, 2, 2, 2, 2, 2, 2, 2, 2, 2,
|
||||
2, 2, 2, 2,
|
||||
4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4,
|
||||
12, 12, 12, 12, 12, 12, 12, 12,
|
||||
24, 24, 24, 24, 24, 24, 24, 24,
|
||||
48, 48, 48, 48, 48, 48, 48, 48,
|
||||
128, 129,
|
||||
)
|
||||
|
||||
|
||||
class BSRoformer(Module):
|
||||
|
||||
@beartype
|
||||
def __init__(
|
||||
self,
|
||||
dim,
|
||||
*,
|
||||
depth,
|
||||
stereo=False,
|
||||
num_stems=1,
|
||||
time_transformer_depth=2,
|
||||
freq_transformer_depth=2,
|
||||
linear_transformer_depth=0,
|
||||
freqs_per_bands: Tuple[int, ...] = DEFAULT_FREQS_PER_BANDS,
|
||||
# in the paper, they divide into ~60 bands, test with 1 for starters
|
||||
dim_head=64,
|
||||
heads=8,
|
||||
attn_dropout=0.,
|
||||
ff_dropout=0.,
|
||||
flash_attn=True,
|
||||
dim_freqs_in=1025,
|
||||
stft_n_fft=2048,
|
||||
stft_hop_length=512,
|
||||
# 10ms at 44100Hz, from sections 4.1, 4.4 in the paper - @faroit recommends // 2 or // 4 for better reconstruction
|
||||
stft_win_length=2048,
|
||||
stft_normalized=False,
|
||||
stft_window_fn: Optional[Callable] = None,
|
||||
mask_estimator_depth=2,
|
||||
multi_stft_resolution_loss_weight=1.,
|
||||
multi_stft_resolutions_window_sizes: Tuple[int, ...] = (4096, 2048, 1024, 512, 256),
|
||||
multi_stft_hop_size=147,
|
||||
multi_stft_normalized=False,
|
||||
multi_stft_window_fn: Callable = torch.hann_window,
|
||||
mlp_expansion_factor=4,
|
||||
use_torch_checkpoint=False,
|
||||
skip_connection=False,
|
||||
sage_attention=False,
|
||||
):
|
||||
super().__init__()
|
||||
|
||||
self.stereo = stereo
|
||||
self.audio_channels = 2 if stereo else 1
|
||||
self.num_stems = num_stems
|
||||
self.use_torch_checkpoint = use_torch_checkpoint
|
||||
self.skip_connection = skip_connection
|
||||
|
||||
self.layers = ModuleList([])
|
||||
|
||||
if sage_attention:
|
||||
print("Use Sage Attention")
|
||||
|
||||
transformer_kwargs = dict(
|
||||
dim=dim,
|
||||
heads=heads,
|
||||
dim_head=dim_head,
|
||||
attn_dropout=attn_dropout,
|
||||
ff_dropout=ff_dropout,
|
||||
flash_attn=flash_attn,
|
||||
norm_output=False,
|
||||
sage_attention=sage_attention,
|
||||
)
|
||||
|
||||
time_rotary_embed = RotaryEmbedding(dim=dim_head)
|
||||
freq_rotary_embed = RotaryEmbedding(dim=dim_head)
|
||||
|
||||
for _ in range(depth):
|
||||
tran_modules = []
|
||||
if linear_transformer_depth > 0:
|
||||
tran_modules.append(Transformer(depth=linear_transformer_depth, linear_attn=True, **transformer_kwargs))
|
||||
tran_modules.append(
|
||||
Transformer(depth=time_transformer_depth, rotary_embed=time_rotary_embed, **transformer_kwargs)
|
||||
)
|
||||
tran_modules.append(
|
||||
Transformer(depth=freq_transformer_depth, rotary_embed=freq_rotary_embed, **transformer_kwargs)
|
||||
)
|
||||
self.layers.append(nn.ModuleList(tran_modules))
|
||||
|
||||
self.final_norm = RMSNorm(dim)
|
||||
|
||||
self.stft_kwargs = dict(
|
||||
n_fft=stft_n_fft,
|
||||
hop_length=stft_hop_length,
|
||||
win_length=stft_win_length,
|
||||
normalized=stft_normalized
|
||||
)
|
||||
|
||||
self.stft_window_fn = partial(default(stft_window_fn, torch.hann_window), stft_win_length)
|
||||
|
||||
freqs = torch.stft(torch.randn(1, 4096), **self.stft_kwargs, window=torch.ones(stft_win_length), return_complex=True).shape[1]
|
||||
|
||||
assert len(freqs_per_bands) > 1
|
||||
assert sum(
|
||||
freqs_per_bands) == freqs, f'the number of freqs in the bands must equal {freqs} based on the STFT settings, but got {sum(freqs_per_bands)}'
|
||||
|
||||
freqs_per_bands_with_complex = tuple(2 * f * self.audio_channels for f in freqs_per_bands)
|
||||
|
||||
self.band_split = BandSplit(
|
||||
dim=dim,
|
||||
dim_inputs=freqs_per_bands_with_complex
|
||||
)
|
||||
|
||||
self.mask_estimators = nn.ModuleList([])
|
||||
|
||||
for _ in range(num_stems):
|
||||
mask_estimator = MaskEstimator(
|
||||
dim=dim,
|
||||
dim_inputs=freqs_per_bands_with_complex,
|
||||
depth=mask_estimator_depth,
|
||||
mlp_expansion_factor=mlp_expansion_factor,
|
||||
)
|
||||
|
||||
self.mask_estimators.append(mask_estimator)
|
||||
|
||||
# for the multi-resolution stft loss
|
||||
|
||||
self.multi_stft_resolution_loss_weight = multi_stft_resolution_loss_weight
|
||||
self.multi_stft_resolutions_window_sizes = multi_stft_resolutions_window_sizes
|
||||
self.multi_stft_n_fft = stft_n_fft
|
||||
self.multi_stft_window_fn = multi_stft_window_fn
|
||||
|
||||
self.multi_stft_kwargs = dict(
|
||||
hop_length=multi_stft_hop_size,
|
||||
normalized=multi_stft_normalized
|
||||
)
|
||||
|
||||
def forward(
|
||||
self,
|
||||
raw_audio,
|
||||
target=None,
|
||||
return_loss_breakdown=False
|
||||
):
|
||||
"""
|
||||
einops
|
||||
|
||||
b - batch
|
||||
f - freq
|
||||
t - time
|
||||
s - audio channel (1 for mono, 2 for stereo)
|
||||
n - number of 'stems'
|
||||
c - complex (2)
|
||||
d - feature dimension
|
||||
"""
|
||||
|
||||
device = raw_audio.device
|
||||
|
||||
# defining whether model is loaded on MPS (MacOS GPU accelerator)
|
||||
x_is_mps = True if device.type == "mps" else False
|
||||
|
||||
if raw_audio.ndim == 2:
|
||||
raw_audio = rearrange(raw_audio, 'b t -> b 1 t')
|
||||
|
||||
channels = raw_audio.shape[1]
|
||||
assert (not self.stereo and channels == 1) or (self.stereo and channels == 2), 'stereo needs to be set to True if passing in audio signal that is stereo (channel dimension of 2). also need to be False if mono (channel dimension of 1)'
|
||||
|
||||
# to stft
|
||||
|
||||
raw_audio, batch_audio_channel_packed_shape = pack_one(raw_audio, '* t')
|
||||
|
||||
stft_window = self.stft_window_fn(device=device)
|
||||
|
||||
# RuntimeError: FFT operations are only supported on MacOS 14+
|
||||
# Since it's tedious to define whether we're on correct MacOS version - simple try-catch is used
|
||||
try:
|
||||
stft_repr = torch.stft(raw_audio, **self.stft_kwargs, window=stft_window, return_complex=True)
|
||||
except:
|
||||
stft_repr = torch.stft(raw_audio.cpu() if x_is_mps else raw_audio, **self.stft_kwargs,
|
||||
window=stft_window.cpu() if x_is_mps else stft_window, return_complex=True).to(
|
||||
device)
|
||||
stft_repr = torch.view_as_real(stft_repr)
|
||||
|
||||
stft_repr = unpack_one(stft_repr, batch_audio_channel_packed_shape, '* f t c')
|
||||
|
||||
# merge stereo / mono into the frequency, with frequency leading dimension, for band splitting
|
||||
stft_repr = rearrange(stft_repr,'b s f t c -> b (f s) t c')
|
||||
|
||||
x = rearrange(stft_repr, 'b f t c -> b t (f c)')
|
||||
|
||||
if self.use_torch_checkpoint:
|
||||
x = checkpoint(self.band_split, x, use_reentrant=False)
|
||||
else:
|
||||
x = self.band_split(x)
|
||||
|
||||
# axial / hierarchical attention
|
||||
|
||||
store = [None] * len(self.layers)
|
||||
for i, transformer_block in enumerate(self.layers):
|
||||
|
||||
if len(transformer_block) == 3:
|
||||
linear_transformer, time_transformer, freq_transformer = transformer_block
|
||||
|
||||
x, ft_ps = pack([x], 'b * d')
|
||||
if self.use_torch_checkpoint:
|
||||
x = checkpoint(linear_transformer, x, use_reentrant=False)
|
||||
else:
|
||||
x = linear_transformer(x)
|
||||
x, = unpack(x, ft_ps, 'b * d')
|
||||
else:
|
||||
time_transformer, freq_transformer = transformer_block
|
||||
|
||||
if self.skip_connection:
|
||||
# Sum all previous
|
||||
for j in range(i):
|
||||
x = x + store[j]
|
||||
|
||||
x = rearrange(x, 'b t f d -> b f t d')
|
||||
x, ps = pack([x], '* t d')
|
||||
|
||||
if self.use_torch_checkpoint:
|
||||
x = checkpoint(time_transformer, x, use_reentrant=False)
|
||||
else:
|
||||
x = time_transformer(x)
|
||||
|
||||
x, = unpack(x, ps, '* t d')
|
||||
x = rearrange(x, 'b f t d -> b t f d')
|
||||
x, ps = pack([x], '* f d')
|
||||
|
||||
if self.use_torch_checkpoint:
|
||||
x = checkpoint(freq_transformer, x, use_reentrant=False)
|
||||
else:
|
||||
x = freq_transformer(x)
|
||||
|
||||
x, = unpack(x, ps, '* f d')
|
||||
|
||||
if self.skip_connection:
|
||||
store[i] = x
|
||||
|
||||
x = self.final_norm(x)
|
||||
|
||||
num_stems = len(self.mask_estimators)
|
||||
|
||||
if self.use_torch_checkpoint:
|
||||
mask = torch.stack([checkpoint(fn, x, use_reentrant=False) for fn in self.mask_estimators], dim=1)
|
||||
else:
|
||||
mask = torch.stack([fn(x) for fn in self.mask_estimators], dim=1)
|
||||
mask = rearrange(mask, 'b n t (f c) -> b n f t c', c=2)
|
||||
|
||||
# modulate frequency representation
|
||||
|
||||
stft_repr = rearrange(stft_repr, 'b f t c -> b 1 f t c')
|
||||
|
||||
# complex number multiplication
|
||||
|
||||
stft_repr = torch.view_as_complex(stft_repr)
|
||||
mask = torch.view_as_complex(mask)
|
||||
|
||||
stft_repr = stft_repr * mask
|
||||
|
||||
# istft
|
||||
|
||||
stft_repr = rearrange(stft_repr, 'b n (f s) t -> (b n s) f t', s=self.audio_channels)
|
||||
|
||||
# same as torch.stft() fix for MacOS MPS above
|
||||
try:
|
||||
recon_audio = torch.istft(stft_repr, **self.stft_kwargs, window=stft_window, return_complex=False, length=raw_audio.shape[-1])
|
||||
except:
|
||||
recon_audio = torch.istft(stft_repr.cpu() if x_is_mps else stft_repr, **self.stft_kwargs, window=stft_window.cpu() if x_is_mps else stft_window, return_complex=False, length=raw_audio.shape[-1]).to(device)
|
||||
|
||||
recon_audio = rearrange(recon_audio, '(b n s) t -> b n s t', s=self.audio_channels, n=num_stems)
|
||||
|
||||
if num_stems == 1:
|
||||
recon_audio = rearrange(recon_audio, 'b 1 s t -> b s t')
|
||||
|
||||
# if a target is passed in, calculate loss for learning
|
||||
|
||||
if not exists(target):
|
||||
return recon_audio
|
||||
|
||||
if self.num_stems > 1:
|
||||
assert target.ndim == 4 and target.shape[1] == self.num_stems
|
||||
|
||||
if target.ndim == 2:
|
||||
target = rearrange(target, '... t -> ... 1 t')
|
||||
|
||||
target = target[..., :recon_audio.shape[-1]] # protect against lost length on istft
|
||||
|
||||
loss = F.l1_loss(recon_audio, target)
|
||||
|
||||
multi_stft_resolution_loss = 0.
|
||||
|
||||
for window_size in self.multi_stft_resolutions_window_sizes:
|
||||
res_stft_kwargs = dict(
|
||||
n_fft=max(window_size, self.multi_stft_n_fft), # not sure what n_fft is across multi resolution stft
|
||||
win_length=window_size,
|
||||
return_complex=True,
|
||||
window=self.multi_stft_window_fn(window_size, device=device),
|
||||
**self.multi_stft_kwargs,
|
||||
)
|
||||
|
||||
recon_Y = torch.stft(rearrange(recon_audio, '... s t -> (... s) t'), **res_stft_kwargs)
|
||||
target_Y = torch.stft(rearrange(target, '... s t -> (... s) t'), **res_stft_kwargs)
|
||||
|
||||
multi_stft_resolution_loss = multi_stft_resolution_loss + F.l1_loss(recon_Y, target_Y)
|
||||
|
||||
weighted_multi_resolution_loss = multi_stft_resolution_loss * self.multi_stft_resolution_loss_weight
|
||||
|
||||
total_loss = loss + weighted_multi_resolution_loss
|
||||
|
||||
if not return_loss_breakdown:
|
||||
return total_loss
|
||||
|
||||
return total_loss, (loss, multi_stft_resolution_loss)
|
||||
@@ -0,0 +1,703 @@
|
||||
from functools import partial
|
||||
|
||||
import torch
|
||||
from torch import nn, einsum, Tensor
|
||||
from torch.nn import Module, ModuleList
|
||||
import torch.nn.functional as F
|
||||
|
||||
from .attend import Attend
|
||||
try:
|
||||
from .attend_sage import Attend as AttendSage
|
||||
except:
|
||||
pass
|
||||
from torch.utils.checkpoint import checkpoint
|
||||
|
||||
from beartype.typing import Tuple, Optional, List, Callable
|
||||
from beartype import beartype
|
||||
|
||||
from rotary_embedding_torch import RotaryEmbedding
|
||||
|
||||
from einops import rearrange, pack, unpack, reduce, repeat
|
||||
from einops.layers.torch import Rearrange
|
||||
|
||||
from librosa import filters
|
||||
|
||||
|
||||
# helper functions
|
||||
|
||||
def exists(val):
|
||||
return val is not None
|
||||
|
||||
|
||||
def default(v, d):
|
||||
return v if exists(v) else d
|
||||
|
||||
|
||||
def pack_one(t, pattern):
|
||||
return pack([t], pattern)
|
||||
|
||||
|
||||
def unpack_one(t, ps, pattern):
|
||||
return unpack(t, ps, pattern)[0]
|
||||
|
||||
|
||||
def pad_at_dim(t, pad, dim=-1, value=0.):
|
||||
dims_from_right = (- dim - 1) if dim < 0 else (t.ndim - dim - 1)
|
||||
zeros = ((0, 0) * dims_from_right)
|
||||
return F.pad(t, (*zeros, *pad), value=value)
|
||||
|
||||
|
||||
def l2norm(t):
|
||||
return F.normalize(t, dim=-1, p=2)
|
||||
|
||||
|
||||
# norm
|
||||
|
||||
class RMSNorm(Module):
|
||||
def __init__(self, dim):
|
||||
super().__init__()
|
||||
self.scale = dim ** 0.5
|
||||
self.gamma = nn.Parameter(torch.ones(dim))
|
||||
|
||||
def forward(self, x):
|
||||
return F.normalize(x, dim=-1) * self.scale * self.gamma
|
||||
|
||||
|
||||
# attention
|
||||
|
||||
class FeedForward(Module):
|
||||
def __init__(
|
||||
self,
|
||||
dim,
|
||||
mult=4,
|
||||
dropout=0.
|
||||
):
|
||||
super().__init__()
|
||||
dim_inner = int(dim * mult)
|
||||
self.net = nn.Sequential(
|
||||
RMSNorm(dim),
|
||||
nn.Linear(dim, dim_inner),
|
||||
nn.GELU(),
|
||||
nn.Dropout(dropout),
|
||||
nn.Linear(dim_inner, dim),
|
||||
nn.Dropout(dropout)
|
||||
)
|
||||
|
||||
def forward(self, x):
|
||||
return self.net(x)
|
||||
|
||||
|
||||
class Attention(Module):
|
||||
def __init__(
|
||||
self,
|
||||
dim,
|
||||
heads=8,
|
||||
dim_head=64,
|
||||
dropout=0.,
|
||||
rotary_embed=None,
|
||||
flash=True,
|
||||
sage_attention=False,
|
||||
):
|
||||
super().__init__()
|
||||
self.heads = heads
|
||||
self.scale = dim_head ** -0.5
|
||||
dim_inner = heads * dim_head
|
||||
|
||||
self.rotary_embed = rotary_embed
|
||||
|
||||
if sage_attention:
|
||||
self.attend = AttendSage(flash=flash, dropout=dropout)
|
||||
else:
|
||||
self.attend = Attend(flash=flash, dropout=dropout)
|
||||
self.norm = RMSNorm(dim)
|
||||
self.to_qkv = nn.Linear(dim, dim_inner * 3, bias=False)
|
||||
|
||||
self.to_gates = nn.Linear(dim, heads)
|
||||
|
||||
self.to_out = nn.Sequential(
|
||||
nn.Linear(dim_inner, dim, bias=False),
|
||||
nn.Dropout(dropout)
|
||||
)
|
||||
|
||||
def forward(self, x):
|
||||
x = self.norm(x)
|
||||
|
||||
q, k, v = rearrange(self.to_qkv(x), 'b n (qkv h d) -> qkv b h n d', qkv=3, h=self.heads)
|
||||
|
||||
if exists(self.rotary_embed):
|
||||
q = self.rotary_embed.rotate_queries_or_keys(q)
|
||||
k = self.rotary_embed.rotate_queries_or_keys(k)
|
||||
|
||||
out = self.attend(q, k, v)
|
||||
|
||||
gates = self.to_gates(x)
|
||||
out = out * rearrange(gates, 'b n h -> b h n 1').sigmoid()
|
||||
|
||||
out = rearrange(out, 'b h n d -> b n (h d)')
|
||||
return self.to_out(out)
|
||||
|
||||
|
||||
class LinearAttention(Module):
|
||||
"""
|
||||
this flavor of linear attention proposed in https://arxiv.org/abs/2106.09681 by El-Nouby et al.
|
||||
"""
|
||||
|
||||
@beartype
|
||||
def __init__(
|
||||
self,
|
||||
*,
|
||||
dim,
|
||||
dim_head=32,
|
||||
heads=8,
|
||||
scale=8,
|
||||
flash=False,
|
||||
dropout=0.,
|
||||
sage_attention=False
|
||||
):
|
||||
super().__init__()
|
||||
dim_inner = dim_head * heads
|
||||
self.norm = RMSNorm(dim)
|
||||
|
||||
self.to_qkv = nn.Sequential(
|
||||
nn.Linear(dim, dim_inner * 3, bias=False),
|
||||
Rearrange('b n (qkv h d) -> qkv b h d n', qkv=3, h=heads)
|
||||
)
|
||||
|
||||
self.temperature = nn.Parameter(torch.ones(heads, 1, 1))
|
||||
|
||||
if sage_attention:
|
||||
self.attend = AttendSage(
|
||||
scale=scale,
|
||||
dropout=dropout,
|
||||
flash=flash
|
||||
)
|
||||
else:
|
||||
self.attend = Attend(
|
||||
scale=scale,
|
||||
dropout=dropout,
|
||||
flash=flash
|
||||
)
|
||||
|
||||
self.to_out = nn.Sequential(
|
||||
Rearrange('b h d n -> b n (h d)'),
|
||||
nn.Linear(dim_inner, dim, bias=False)
|
||||
)
|
||||
|
||||
def forward(
|
||||
self,
|
||||
x
|
||||
):
|
||||
x = self.norm(x)
|
||||
|
||||
q, k, v = self.to_qkv(x)
|
||||
|
||||
q, k = map(l2norm, (q, k))
|
||||
q = q * self.temperature.exp()
|
||||
|
||||
out = self.attend(q, k, v)
|
||||
|
||||
return self.to_out(out)
|
||||
|
||||
|
||||
class Transformer(Module):
|
||||
def __init__(
|
||||
self,
|
||||
*,
|
||||
dim,
|
||||
depth,
|
||||
dim_head=64,
|
||||
heads=8,
|
||||
attn_dropout=0.,
|
||||
ff_dropout=0.,
|
||||
ff_mult=4,
|
||||
norm_output=True,
|
||||
rotary_embed=None,
|
||||
flash_attn=True,
|
||||
linear_attn=False,
|
||||
sage_attention=False,
|
||||
):
|
||||
super().__init__()
|
||||
self.layers = ModuleList([])
|
||||
|
||||
for _ in range(depth):
|
||||
if linear_attn:
|
||||
attn = LinearAttention(
|
||||
dim=dim,
|
||||
dim_head=dim_head,
|
||||
heads=heads,
|
||||
dropout=attn_dropout,
|
||||
flash=flash_attn,
|
||||
sage_attention=sage_attention
|
||||
)
|
||||
else:
|
||||
attn = Attention(
|
||||
dim=dim,
|
||||
dim_head=dim_head,
|
||||
heads=heads,
|
||||
dropout=attn_dropout,
|
||||
rotary_embed=rotary_embed,
|
||||
flash=flash_attn,
|
||||
sage_attention=sage_attention
|
||||
)
|
||||
|
||||
self.layers.append(ModuleList([
|
||||
attn,
|
||||
FeedForward(dim=dim, mult=ff_mult, dropout=ff_dropout)
|
||||
]))
|
||||
|
||||
self.norm = RMSNorm(dim) if norm_output else nn.Identity()
|
||||
|
||||
def forward(self, x):
|
||||
|
||||
for attn, ff in self.layers:
|
||||
x = attn(x) + x
|
||||
x = ff(x) + x
|
||||
|
||||
return self.norm(x)
|
||||
|
||||
|
||||
# bandsplit module
|
||||
|
||||
class BandSplit(Module):
|
||||
@beartype
|
||||
def __init__(
|
||||
self,
|
||||
dim,
|
||||
dim_inputs: Tuple[int, ...]
|
||||
):
|
||||
super().__init__()
|
||||
self.dim_inputs = dim_inputs
|
||||
self.to_features = ModuleList([])
|
||||
|
||||
for dim_in in dim_inputs:
|
||||
net = nn.Sequential(
|
||||
RMSNorm(dim_in),
|
||||
nn.Linear(dim_in, dim)
|
||||
)
|
||||
|
||||
self.to_features.append(net)
|
||||
|
||||
def forward(self, x):
|
||||
x = x.split(self.dim_inputs, dim=-1)
|
||||
|
||||
outs = []
|
||||
for split_input, to_feature in zip(x, self.to_features):
|
||||
split_output = to_feature(split_input)
|
||||
outs.append(split_output)
|
||||
|
||||
return torch.stack(outs, dim=-2)
|
||||
|
||||
|
||||
def MLP(
|
||||
dim_in,
|
||||
dim_out,
|
||||
dim_hidden=None,
|
||||
depth=1,
|
||||
activation=nn.Tanh
|
||||
):
|
||||
dim_hidden = default(dim_hidden, dim_in)
|
||||
|
||||
net = []
|
||||
dims = (dim_in, *((dim_hidden,) * depth), dim_out)
|
||||
|
||||
for ind, (layer_dim_in, layer_dim_out) in enumerate(zip(dims[:-1], dims[1:])):
|
||||
is_last = ind == (len(dims) - 2)
|
||||
|
||||
net.append(nn.Linear(layer_dim_in, layer_dim_out))
|
||||
|
||||
if is_last:
|
||||
continue
|
||||
|
||||
net.append(activation())
|
||||
|
||||
return nn.Sequential(*net)
|
||||
|
||||
|
||||
class MaskEstimator(Module):
|
||||
@beartype
|
||||
def __init__(
|
||||
self,
|
||||
dim,
|
||||
dim_inputs: Tuple[int, ...],
|
||||
depth,
|
||||
mlp_expansion_factor=4
|
||||
):
|
||||
super().__init__()
|
||||
self.dim_inputs = dim_inputs
|
||||
self.to_freqs = ModuleList([])
|
||||
dim_hidden = dim * mlp_expansion_factor
|
||||
|
||||
for dim_in in dim_inputs:
|
||||
net = []
|
||||
|
||||
mlp = nn.Sequential(
|
||||
MLP(dim, dim_in * 2, dim_hidden=dim_hidden, depth=depth),
|
||||
nn.GLU(dim=-1)
|
||||
)
|
||||
|
||||
self.to_freqs.append(mlp)
|
||||
|
||||
def forward(self, x):
|
||||
x = x.unbind(dim=-2)
|
||||
|
||||
outs = []
|
||||
|
||||
for band_features, mlp in zip(x, self.to_freqs):
|
||||
freq_out = mlp(band_features)
|
||||
outs.append(freq_out)
|
||||
|
||||
return torch.cat(outs, dim=-1)
|
||||
|
||||
|
||||
# main class
|
||||
|
||||
class MelBandRoformer(Module):
|
||||
|
||||
@beartype
|
||||
def __init__(
|
||||
self,
|
||||
dim,
|
||||
*,
|
||||
depth,
|
||||
stereo=False,
|
||||
num_stems=1,
|
||||
time_transformer_depth=2,
|
||||
freq_transformer_depth=2,
|
||||
linear_transformer_depth=0,
|
||||
num_bands=60,
|
||||
dim_head=64,
|
||||
heads=8,
|
||||
attn_dropout=0.1,
|
||||
ff_dropout=0.1,
|
||||
flash_attn=True,
|
||||
dim_freqs_in=1025,
|
||||
sample_rate=44100, # needed for mel filter bank from librosa
|
||||
stft_n_fft=2048,
|
||||
stft_hop_length=512,
|
||||
# 10ms at 44100Hz, from sections 4.1, 4.4 in the paper - @faroit recommends // 2 or // 4 for better reconstruction
|
||||
stft_win_length=2048,
|
||||
stft_normalized=False,
|
||||
stft_window_fn: Optional[Callable] = None,
|
||||
mask_estimator_depth=1,
|
||||
multi_stft_resolution_loss_weight=1.,
|
||||
multi_stft_resolutions_window_sizes: Tuple[int, ...] = (4096, 2048, 1024, 512, 256),
|
||||
multi_stft_hop_size=147,
|
||||
multi_stft_normalized=False,
|
||||
multi_stft_window_fn: Callable = torch.hann_window,
|
||||
match_input_audio_length=False, # if True, pad output tensor to match length of input tensor
|
||||
mlp_expansion_factor=4,
|
||||
use_torch_checkpoint=False,
|
||||
skip_connection=False,
|
||||
sage_attention=False,
|
||||
):
|
||||
super().__init__()
|
||||
|
||||
self.stereo = stereo
|
||||
self.audio_channels = 2 if stereo else 1
|
||||
self.num_stems = num_stems
|
||||
self.use_torch_checkpoint = use_torch_checkpoint
|
||||
self.skip_connection = skip_connection
|
||||
|
||||
self.layers = ModuleList([])
|
||||
|
||||
if sage_attention:
|
||||
print("Use Sage Attention")
|
||||
|
||||
transformer_kwargs = dict(
|
||||
dim=dim,
|
||||
heads=heads,
|
||||
dim_head=dim_head,
|
||||
attn_dropout=attn_dropout,
|
||||
ff_dropout=ff_dropout,
|
||||
flash_attn=flash_attn,
|
||||
sage_attention=sage_attention,
|
||||
)
|
||||
|
||||
time_rotary_embed = RotaryEmbedding(dim=dim_head)
|
||||
freq_rotary_embed = RotaryEmbedding(dim=dim_head)
|
||||
|
||||
for _ in range(depth):
|
||||
tran_modules = []
|
||||
if linear_transformer_depth > 0:
|
||||
tran_modules.append(Transformer(depth=linear_transformer_depth, linear_attn=True, **transformer_kwargs))
|
||||
tran_modules.append(
|
||||
Transformer(depth=time_transformer_depth, rotary_embed=time_rotary_embed, **transformer_kwargs)
|
||||
)
|
||||
tran_modules.append(
|
||||
Transformer(depth=freq_transformer_depth, rotary_embed=freq_rotary_embed, **transformer_kwargs)
|
||||
)
|
||||
self.layers.append(nn.ModuleList(tran_modules))
|
||||
|
||||
self.stft_window_fn = partial(default(stft_window_fn, torch.hann_window), stft_win_length)
|
||||
|
||||
self.stft_kwargs = dict(
|
||||
n_fft=stft_n_fft,
|
||||
hop_length=stft_hop_length,
|
||||
win_length=stft_win_length,
|
||||
normalized=stft_normalized
|
||||
)
|
||||
|
||||
freqs = torch.stft(torch.randn(1, 4096), **self.stft_kwargs, window=torch.ones(stft_n_fft), return_complex=True).shape[1]
|
||||
|
||||
# create mel filter bank
|
||||
# with librosa.filters.mel as in section 2 of paper
|
||||
|
||||
mel_filter_bank_numpy = filters.mel(sr=sample_rate, n_fft=stft_n_fft, n_mels=num_bands)
|
||||
|
||||
mel_filter_bank = torch.from_numpy(mel_filter_bank_numpy)
|
||||
|
||||
# for some reason, it doesn't include the first freq? just force a value for now
|
||||
|
||||
mel_filter_bank[0][0] = 1.
|
||||
|
||||
# In some systems/envs we get 0.0 instead of ~1.9e-18 in the last position,
|
||||
# so let's force a positive value
|
||||
|
||||
mel_filter_bank[-1, -1] = 1.
|
||||
|
||||
# binary as in paper (then estimated masks are averaged for overlapping regions)
|
||||
|
||||
freqs_per_band = mel_filter_bank > 0
|
||||
assert freqs_per_band.any(dim=0).all(), 'all frequencies need to be covered by all bands for now'
|
||||
|
||||
repeated_freq_indices = repeat(torch.arange(freqs), 'f -> b f', b=num_bands)
|
||||
freq_indices = repeated_freq_indices[freqs_per_band]
|
||||
|
||||
if stereo:
|
||||
freq_indices = repeat(freq_indices, 'f -> f s', s=2)
|
||||
freq_indices = freq_indices * 2 + torch.arange(2)
|
||||
freq_indices = rearrange(freq_indices, 'f s -> (f s)')
|
||||
|
||||
self.register_buffer('freq_indices', freq_indices, persistent=False)
|
||||
self.register_buffer('freqs_per_band', freqs_per_band, persistent=False)
|
||||
|
||||
num_freqs_per_band = reduce(freqs_per_band, 'b f -> b', 'sum')
|
||||
num_bands_per_freq = reduce(freqs_per_band, 'b f -> f', 'sum')
|
||||
|
||||
self.register_buffer('num_freqs_per_band', num_freqs_per_band, persistent=False)
|
||||
self.register_buffer('num_bands_per_freq', num_bands_per_freq, persistent=False)
|
||||
|
||||
# band split and mask estimator
|
||||
|
||||
freqs_per_bands_with_complex = tuple(2 * f * self.audio_channels for f in num_freqs_per_band.tolist())
|
||||
|
||||
self.band_split = BandSplit(
|
||||
dim=dim,
|
||||
dim_inputs=freqs_per_bands_with_complex
|
||||
)
|
||||
|
||||
self.mask_estimators = nn.ModuleList([])
|
||||
|
||||
for _ in range(num_stems):
|
||||
mask_estimator = MaskEstimator(
|
||||
dim=dim,
|
||||
dim_inputs=freqs_per_bands_with_complex,
|
||||
depth=mask_estimator_depth,
|
||||
mlp_expansion_factor=mlp_expansion_factor,
|
||||
)
|
||||
|
||||
self.mask_estimators.append(mask_estimator)
|
||||
|
||||
# for the multi-resolution stft loss
|
||||
|
||||
self.multi_stft_resolution_loss_weight = multi_stft_resolution_loss_weight
|
||||
self.multi_stft_resolutions_window_sizes = multi_stft_resolutions_window_sizes
|
||||
self.multi_stft_n_fft = stft_n_fft
|
||||
self.multi_stft_window_fn = multi_stft_window_fn
|
||||
|
||||
self.multi_stft_kwargs = dict(
|
||||
hop_length=multi_stft_hop_size,
|
||||
normalized=multi_stft_normalized
|
||||
)
|
||||
|
||||
self.match_input_audio_length = match_input_audio_length
|
||||
|
||||
def forward(
|
||||
self,
|
||||
raw_audio,
|
||||
target=None,
|
||||
return_loss_breakdown=False
|
||||
):
|
||||
"""
|
||||
einops
|
||||
|
||||
b - batch
|
||||
f - freq
|
||||
t - time
|
||||
s - audio channel (1 for mono, 2 for stereo)
|
||||
n - number of 'stems'
|
||||
c - complex (2)
|
||||
d - feature dimension
|
||||
"""
|
||||
|
||||
device = raw_audio.device
|
||||
|
||||
if raw_audio.ndim == 2:
|
||||
raw_audio = rearrange(raw_audio, 'b t -> b 1 t')
|
||||
|
||||
batch, channels, raw_audio_length = raw_audio.shape
|
||||
|
||||
istft_length = raw_audio_length if self.match_input_audio_length else None
|
||||
|
||||
assert (not self.stereo and channels == 1) or (
|
||||
self.stereo and channels == 2), 'stereo needs to be set to True if passing in audio signal that is stereo (channel dimension of 2). also need to be False if mono (channel dimension of 1)'
|
||||
|
||||
# to stft
|
||||
|
||||
raw_audio, batch_audio_channel_packed_shape = pack_one(raw_audio, '* t')
|
||||
|
||||
stft_window = self.stft_window_fn(device=device)
|
||||
|
||||
stft_repr = torch.stft(raw_audio, **self.stft_kwargs, window=stft_window, return_complex=True)
|
||||
stft_repr = torch.view_as_real(stft_repr)
|
||||
|
||||
stft_repr = unpack_one(stft_repr, batch_audio_channel_packed_shape, '* f t c')
|
||||
|
||||
# merge stereo / mono into the frequency, with frequency leading dimension, for band splitting
|
||||
stft_repr = rearrange(stft_repr,'b s f t c -> b (f s) t c')
|
||||
|
||||
# index out all frequencies for all frequency ranges across bands ascending in one go
|
||||
|
||||
batch_arange = torch.arange(batch, device=device)[..., None]
|
||||
|
||||
# account for stereo
|
||||
|
||||
x = stft_repr[batch_arange, self.freq_indices]
|
||||
|
||||
# fold the complex (real and imag) into the frequencies dimension
|
||||
|
||||
x = rearrange(x, 'b f t c -> b t (f c)')
|
||||
|
||||
if self.use_torch_checkpoint:
|
||||
x = checkpoint(self.band_split, x, use_reentrant=False)
|
||||
else:
|
||||
x = self.band_split(x)
|
||||
|
||||
# axial / hierarchical attention
|
||||
|
||||
store = [None] * len(self.layers)
|
||||
for i, transformer_block in enumerate(self.layers):
|
||||
|
||||
if len(transformer_block) == 3:
|
||||
linear_transformer, time_transformer, freq_transformer = transformer_block
|
||||
|
||||
x, ft_ps = pack([x], 'b * d')
|
||||
if self.use_torch_checkpoint:
|
||||
x = checkpoint(linear_transformer, x, use_reentrant=False)
|
||||
else:
|
||||
x = linear_transformer(x)
|
||||
x, = unpack(x, ft_ps, 'b * d')
|
||||
else:
|
||||
time_transformer, freq_transformer = transformer_block
|
||||
|
||||
if self.skip_connection:
|
||||
# Sum all previous
|
||||
for j in range(i):
|
||||
x = x + store[j]
|
||||
|
||||
x = rearrange(x, 'b t f d -> b f t d')
|
||||
x, ps = pack([x], '* t d')
|
||||
|
||||
if self.use_torch_checkpoint:
|
||||
x = checkpoint(time_transformer, x, use_reentrant=False)
|
||||
else:
|
||||
x = time_transformer(x)
|
||||
|
||||
x, = unpack(x, ps, '* t d')
|
||||
x = rearrange(x, 'b f t d -> b t f d')
|
||||
x, ps = pack([x], '* f d')
|
||||
|
||||
if self.use_torch_checkpoint:
|
||||
x = checkpoint(freq_transformer, x, use_reentrant=False)
|
||||
else:
|
||||
x = freq_transformer(x)
|
||||
|
||||
x, = unpack(x, ps, '* f d')
|
||||
|
||||
if self.skip_connection:
|
||||
store[i] = x
|
||||
|
||||
num_stems = len(self.mask_estimators)
|
||||
if self.use_torch_checkpoint:
|
||||
masks = torch.stack([checkpoint(fn, x, use_reentrant=False) for fn in self.mask_estimators], dim=1)
|
||||
else:
|
||||
masks = torch.stack([fn(x) for fn in self.mask_estimators], dim=1)
|
||||
masks = rearrange(masks, 'b n t (f c) -> b n f t c', c=2)
|
||||
|
||||
# modulate frequency representation
|
||||
|
||||
stft_repr = rearrange(stft_repr, 'b f t c -> b 1 f t c')
|
||||
|
||||
# complex number multiplication
|
||||
|
||||
stft_repr = torch.view_as_complex(stft_repr)
|
||||
masks = torch.view_as_complex(masks)
|
||||
|
||||
masks = masks.type(stft_repr.dtype)
|
||||
|
||||
# need to average the estimated mask for the overlapped frequencies
|
||||
|
||||
scatter_indices = repeat(self.freq_indices, 'f -> b n f t', b=batch, n=num_stems, t=stft_repr.shape[-1])
|
||||
|
||||
stft_repr_expanded_stems = repeat(stft_repr, 'b 1 ... -> b n ...', n=num_stems)
|
||||
masks_summed = torch.zeros_like(stft_repr_expanded_stems).scatter_add_(2, scatter_indices, masks)
|
||||
|
||||
denom = repeat(self.num_bands_per_freq, 'f -> (f r) 1', r=channels)
|
||||
|
||||
masks_averaged = masks_summed / denom.clamp(min=1e-8)
|
||||
|
||||
# modulate stft repr with estimated mask
|
||||
|
||||
stft_repr = stft_repr * masks_averaged
|
||||
|
||||
# istft
|
||||
|
||||
stft_repr = rearrange(stft_repr, 'b n (f s) t -> (b n s) f t', s=self.audio_channels)
|
||||
|
||||
recon_audio = torch.istft(stft_repr, **self.stft_kwargs, window=stft_window, return_complex=False,
|
||||
length=istft_length)
|
||||
|
||||
recon_audio = rearrange(recon_audio, '(b n s) t -> b n s t', b=batch, s=self.audio_channels, n=num_stems)
|
||||
|
||||
if num_stems == 1:
|
||||
recon_audio = rearrange(recon_audio, 'b 1 s t -> b s t')
|
||||
|
||||
# if a target is passed in, calculate loss for learning
|
||||
|
||||
if not exists(target):
|
||||
return recon_audio
|
||||
|
||||
if self.num_stems > 1:
|
||||
assert target.ndim == 4 and target.shape[1] == self.num_stems
|
||||
|
||||
if target.ndim == 2:
|
||||
target = rearrange(target, '... t -> ... 1 t')
|
||||
|
||||
target = target[..., :recon_audio.shape[-1]] # protect against lost length on istft
|
||||
|
||||
loss = F.l1_loss(recon_audio, target)
|
||||
|
||||
multi_stft_resolution_loss = 0.
|
||||
|
||||
for window_size in self.multi_stft_resolutions_window_sizes:
|
||||
res_stft_kwargs = dict(
|
||||
n_fft=max(window_size, self.multi_stft_n_fft), # not sure what n_fft is across multi resolution stft
|
||||
win_length=window_size,
|
||||
return_complex=True,
|
||||
window=self.multi_stft_window_fn(window_size, device=device),
|
||||
**self.multi_stft_kwargs,
|
||||
)
|
||||
|
||||
recon_Y = torch.stft(rearrange(recon_audio, '... s t -> (... s) t'), **res_stft_kwargs)
|
||||
target_Y = torch.stft(rearrange(target, '... s t -> (... s) t'), **res_stft_kwargs)
|
||||
|
||||
multi_stft_resolution_loss = multi_stft_resolution_loss + F.l1_loss(recon_Y, target_Y)
|
||||
|
||||
weighted_multi_resolution_loss = multi_stft_resolution_loss * self.multi_stft_resolution_loss_weight
|
||||
|
||||
total_loss = loss + weighted_multi_resolution_loss
|
||||
|
||||
if not return_loss_breakdown:
|
||||
return total_loss
|
||||
|
||||
return total_loss, (loss, multi_stft_resolution_loss)
|
||||
Reference in New Issue
Block a user