FRCRN_SE_16K-MLX / 6bit /mlx_model.py
jennyxue's picture
6bit build of FRCRN
a8a5c47 verified
Raw History Blame Contribute Delete
12.3 kB
"""FRCRN SE 16K in MLX: speech enhancement (denoising) at 16 kHz, a port of ClearerVoice-Studio's FRCRN
(clearvoice/models/frcrn_se: DCCRN-style complex UNet x2 with complex FSMN memories and SE layers on a 640-point
STFT). Parameter names are the original's; convolutions are channels-last.
import mlx_model
model = mlx_model.load("fp16") # a folder: config.json + model.safetensors
clean = mlx_model.enhance(model, "noisy.wav") # float32 numpy, 16 kHz (input resampled to 16 kHz)
Tensors are pairs (re, im) of [B, F, T, C] arrays (the original's [B, C, F, T, 2]). Everything computes in float32
(the 16-bit and quantized folders only store the weights smaller; load() upcasts them): in 16-bit the network drifts
far from the original. Like ClearerVoice, enhance() first normalizes the input's loudness (two RMS stages to
-25 dBFS) and scales the output back. Set MLX_ENABLE_TF32=0 for float32 matmuls (MLX uses TF32 otherwise).
"""
from __future__ import annotations
import json
from pathlib import Path
import mlx.core as mx
import mlx.nn as nn
import numpy as np
from mlx.utils import tree_unflatten
SAMPLE_RATE = 16000
ENC_K = [(5, 2)] * 6 + [(2, 2)]
DEC_K = [(2, 2), (5, 2), (5, 2), (5, 2), (6, 2), (5, 2), (5, 2)]
ENC_C = [1, 128, 128, 128, 128, 128, 128, 128]
DEC_C = [64, 128, 128, 128, 128, 128, 128, 1]
DEPTH = 7
def leaky(x):
return mx.where(x >= 0, x, 0.01 * x)
class BN(nn.Module):
def __init__(self, c):
super().__init__()
self.weight, self.bias = mx.ones((c,)), mx.zeros((c,))
self.running_mean, self.running_var = mx.zeros((c,)), mx.ones((c,))
def __call__(self, x):
return (x - self.running_mean) * mx.rsqrt(self.running_var + 1e-5) * self.weight + self.bias
class CBN(nn.Module):
def __init__(self, c):
super().__init__()
self.bn_re, self.bn_im = BN(c), BN(c)
def __call__(self, z):
return self.bn_re(z[0]), self.bn_im(z[1])
class CConv(nn.Module):
def __init__(self, i, o, k, s=(1, 1), p=(0, 0)):
super().__init__()
self.conv_re = nn.Conv2d(i, o, k, stride=s, padding=p)
self.conv_im = nn.Conv2d(i, o, k, stride=s, padding=p)
def __call__(self, z):
re, im = z
return self.conv_re(re) - self.conv_im(im), self.conv_re(im) + self.conv_im(re)
class CConvT(nn.Module):
def __init__(self, i, o, k, s, p):
super().__init__()
self.tconv_re = nn.ConvTranspose2d(i, o, k, stride=s, padding=p)
self.tconv_im = nn.ConvTranspose2d(i, o, k, stride=s, padding=p)
def __call__(self, z):
re, im = z
return self.tconv_re(re) - self.tconv_im(im), self.tconv_re(im) + self.tconv_im(re)
class Encoder(nn.Module):
def __init__(self, i, o, k):
super().__init__()
self.conv, self.bn = CConv(i, o, k, (2, 1), (0, 1)), CBN(o)
def __call__(self, z):
return tuple(leaky(a) for a in self.bn(self.conv(z)))
class Decoder(nn.Module):
def __init__(self, i, o, k):
super().__init__()
self.transconv, self.bn = CConvT(i, o, k, (2, 1), (0, 1)), CBN(o)
def __call__(self, z):
return tuple(leaky(a) for a in self.bn(self.transconv(z)))
class Fsmn(nn.Module):
"""UniDeepFsmn: x + project(relu(linear(x))) + causal depthwise memory (order 20) along the sequence axis."""
def __init__(self, d=128, order=20):
super().__init__()
self.order = order
self.linear = nn.Linear(d, d)
self.project = nn.Linear(d, d, bias=False)
self.conv1 = {"weight": mx.zeros((d, order, 1))} # depthwise, [C, order, 1]
def __call__(self, x): # [N, L, C]
p = self.project(nn.relu(self.linear(x)))
y = mx.pad(p, [(0, 0), (self.order - 1, 0), (0, 0)])
return x + p + mx.conv1d(y, self.conv1["weight"].astype(p.dtype), groups=p.shape[-1])
class CFsmnL1(nn.Module):
"""ComplexUniDeepFsmn_L1: memory along frequency, per frame."""
def __init__(self):
super().__init__()
self.fsmn_re_L1, self.fsmn_im_L1 = Fsmn(), Fsmn()
def __call__(self, z):
B, F, T, C = z[0].shape
re, im = (mx.transpose(a, (0, 2, 1, 3)).reshape(B * T, F, C) for a in z)
r = self.fsmn_re_L1(re) - self.fsmn_im_L1(im)
i = self.fsmn_re_L1(im) + self.fsmn_im_L1(re)
return tuple(mx.transpose(a.reshape(B, T, F, C), (0, 2, 1, 3)) for a in (r, i))
class CFsmn(nn.Module):
"""ComplexUniDeepFsmn (the bottleneck): two complex layers, memory along time over C*F features."""
def __init__(self):
super().__init__()
self.fsmn_re_L1, self.fsmn_im_L1 = Fsmn(), Fsmn()
self.fsmn_re_L2, self.fsmn_im_L2 = Fsmn(), Fsmn()
def __call__(self, z):
B, F, T, C = z[0].shape
re, im = (mx.transpose(a, (0, 2, 3, 1)).reshape(B, T, C * F) for a in z)
r1 = self.fsmn_re_L1(re) - self.fsmn_im_L1(im)
i1 = self.fsmn_re_L1(im) + self.fsmn_im_L1(re)
r = self.fsmn_re_L2(r1) - self.fsmn_im_L2(i1)
i = self.fsmn_re_L2(i1) + self.fsmn_im_L2(r1)
return tuple(mx.transpose(a.reshape(B, T, C, F), (0, 3, 1, 2)) for a in (r, i))
class SE(nn.Module):
def __init__(self, c=128, r=8):
super().__init__()
self.fc_r = [nn.Linear(c, c // r), nn.ReLU(), nn.Linear(c // r, c), nn.Sigmoid()]
self.fc_i = [nn.Linear(c, c // r), nn.ReLU(), nn.Linear(c // r, c), nn.Sigmoid()]
@staticmethod
def _seq(layers, x):
for l in layers:
x = l(x)
return x
def __call__(self, z):
re, im = z
xr = mx.mean(re.astype(mx.float32), axis=(1, 2)).astype(re.dtype)
xi = mx.mean(im.astype(mx.float32), axis=(1, 2)).astype(im.dtype)
yr = self._seq(self.fc_r, xr) - self._seq(self.fc_i, xi)
yi = self._seq(self.fc_r, xi) + self._seq(self.fc_i, xr)
return re * yr[:, None, None, :], im * yi[:, None, None, :]
class UNet(nn.Module):
def __init__(self):
super().__init__()
self.fsmn = CFsmn()
for i in range(DEPTH):
setattr(self, f"fsmn_enc{i}", CFsmnL1())
setattr(self, f"encoder{i}", Encoder(ENC_C[i], ENC_C[i + 1], ENC_K[i]))
setattr(self, f"se_layer_enc{i}", SE(ENC_C[i + 1]))
setattr(self, f"fsmn_dec{i}", CFsmnL1())
setattr(self, f"decoder{i}", Decoder(DEC_C[i] * 2, DEC_C[i + 1], DEC_K[i]))
if i < DEPTH - 1:
setattr(self, f"se_layer_dec{i}", SE(DEC_C[i + 1]))
self.linear = CConv(DEC_C[-1], 1, (1, 1))
def __call__(self, z):
xs_se = [z]
x = z
for i in range(DEPTH):
if i > 0:
x = getattr(self, f"fsmn_enc{i}")(x)
x = getattr(self, f"encoder{i}")(x)
xs_se.append(getattr(self, f"se_layer_enc{i}")(x))
p = self.fsmn(x)
for i in range(DEPTH):
p = getattr(self, f"decoder{i}")(p)
if i < DEPTH - 1:
p = getattr(self, f"fsmn_dec{i}")(p)
if i == DEPTH - 1:
break
if i < DEPTH - 2:
p = getattr(self, f"se_layer_dec{i}")(p)
s = xs_se[DEPTH - 1 - i]
p = tuple(mx.concatenate([a, b], axis=-1) for a, b in zip(p, s))
return self.linear(p)
class FRCRN(nn.Module):
def __init__(self, win_len=640, win_inc=320, fft_len=640):
super().__init__()
self.win_len, self.win_inc, self.dim = win_len, win_inc, fft_len // 2 + 1
self.unet, self.unet2 = UNet(), UNet()
self.stft = {"weight": mx.zeros((2 * self.dim, 1, win_len))}
self.istft = {"weight": mx.zeros((2 * self.dim, 1, win_len)), "window": mx.zeros((1, win_len, 1))}
def _frames(self, x): # [L] -> [T, win]
n = 1 + (x.shape[0] - self.win_len) // self.win_inc
idx = mx.arange(n)[:, None] * self.win_inc + mx.arange(self.win_len)[None, :]
return x[idx]
def __call__(self, x: mx.array) -> mx.array: # [L] float32 -> [L'] float32
w = self.stft["weight"][:, 0, :].astype(mx.float32)
spec = self._frames(x.astype(mx.float32)) @ w.T # [T, 2D]
re, im = spec[:, : self.dim].T, spec[:, self.dim:].T # [D, T]
z = (re[None, :, :, None], im[None, :, :, None])
u1 = self.unet(z)
m1 = tuple(mx.tanh(a) for a in u1)
u2 = self.unet2(u1)
mr, mi = (mx.tanh(a).astype(mx.float32) + b.astype(mx.float32) for a, b in zip(u2, m1))
mr, mi = mr[0, :, :, 0], mi[0, :, :, 0]
er, ei = re * mr - im * mi, re * mi + im * mr
est = mx.concatenate([er, ei], axis=0).T # [T, 2D]
iw = self.istft["weight"][:, 0, :].astype(mx.float32)
frames = est @ iw # [T, win]
win2 = self.istft["window"][0, :, 0].astype(mx.float32) ** 2
T = frames.shape[0]
L = (T - 1) * self.win_inc + self.win_len
out = mx.zeros((L,)); coff = mx.zeros((L,))
idx = mx.arange(T)[:, None] * self.win_inc + mx.arange(self.win_len)[None, :]
out = out.at[idx.reshape(-1)].add(frames.reshape(-1))
coff = coff.at[idx.reshape(-1)].add(mx.broadcast_to(win2, (T, self.win_len)).reshape(-1))
return out / (coff + 1e-8)
def quantizable(path: str, module) -> bool:
return isinstance(module, nn.Linear) and module.weight.shape[1] % 64 == 0
def load(folder) -> FRCRN:
folder = Path(folder)
conf = json.loads((folder / "config.json").read_text())
model = FRCRN(conf.get("win_len", 640), conf.get("win_inc", 320), conf.get("fft_len", 640))
q = conf.get("quantization")
if q:
nn.quantize(model, group_size=q["group_size"], bits=q["bits"], class_predicate=quantizable)
w = mx.load(str(folder / "model.safetensors"))
model.update(tree_unflatten([(k, v.astype(mx.float32) if mx.issubdtype(v.dtype, mx.floating) else v)
for k, v in w.items()]))
mx.eval(model.parameters())
return model
def audio_norm(x: np.ndarray, eps: float = 1e-6):
"""ClearerVoice's dataloader.audio_norm: scale to -25 dBFS RMS, then again on the louder-than-average samples."""
x = x.astype(np.float64)
s1 = 10 ** (-25 / 20) / (np.sqrt((x ** 2).mean()) + eps)
x = x * s1
p = x ** 2
s2 = 10 ** (-25 / 20) / (np.sqrt(p[p > p.mean()].mean()) + eps)
return (x * s2).astype(np.float32), 1 / (s1 * s2 + eps)
def enhance(model: FRCRN, audio, sr: int | None = None, one_time_decode_length: float = 120.0,
decode_window: float = 1.0, normalize: bool = True) -> np.ndarray:
"""Denoise `audio` (a path or a float waveform at `sr`); float32 at 16 kHz, the input's length. Pads and segments
exactly as ClearerVoice's decode_one_audio_frcrn_se_16k (inputs over 120 s go through 1 s windows, 0.75 s
stride)."""
if isinstance(audio, (str, Path)):
import soundfile as sf
audio, sr = sf.read(str(audio), dtype="float32")
audio = np.asarray(audio, dtype=np.float32)
audio = audio.mean(1) if audio.ndim > 1 else audio
if sr and sr != SAMPLE_RATE:
import librosa
audio = librosa.resample(audio, orig_sr=sr, target_sr=SAMPLE_RATE)
scale = 1.0
if normalize:
audio, scale = audio_norm(audio)
n = len(audio)
window = int(SAMPLE_RATE * decode_window)
stride = int(window * 0.75)
x = audio
if n < window:
x = np.concatenate([x, np.zeros(window - n, np.float32)])
elif n < window + stride:
x = np.concatenate([x, np.zeros(window + stride - n, np.float32)])
elif (n - window) % stride:
x = np.concatenate([x, np.zeros(n - (n - window) // stride * stride, np.float32)])
if n <= SAMPLE_RATE * one_time_decode_length:
y = np.array(model(mx.array(x)))
else:
y = np.zeros(len(x), np.float32)
give = (window - stride) // 2
i = 0
while i + window <= len(x):
seg = np.array(model(mx.array(x[i:i + window])))
if i == 0:
y[:window - give] = seg[:window - give]
else:
y[i + give:i + window - give] = seg[give:window - give]
i += stride
return (np.pad(y, (0, max(0, n - len(y))))[:n] * scale).astype(np.float32)