Spaces:
Runtime error
Runtime error
add mdxnet model
Browse files- mdxnet_model.py +313 -0
mdxnet_model.py
ADDED
|
@@ -0,0 +1,313 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# reference: https://huggingface.co/spaces/r3gm/Audio_separator
|
| 2 |
+
import torch
|
| 3 |
+
import numpy as np
|
| 4 |
+
import onnxruntime as ort
|
| 5 |
+
import hashlib
|
| 6 |
+
import queue
|
| 7 |
+
import threading
|
| 8 |
+
from tqdm import tqdm
|
| 9 |
+
|
| 10 |
+
|
| 11 |
+
class MDXModel:
|
| 12 |
+
def __init__(
|
| 13 |
+
self,
|
| 14 |
+
device,
|
| 15 |
+
dim_f,
|
| 16 |
+
dim_t,
|
| 17 |
+
n_fft,
|
| 18 |
+
hop=1024,
|
| 19 |
+
stem_name=None,
|
| 20 |
+
compensation=1.000,
|
| 21 |
+
):
|
| 22 |
+
self.dim_f = dim_f # frequency bins
|
| 23 |
+
self.dim_t = dim_t
|
| 24 |
+
self.dim_c = 4
|
| 25 |
+
self.n_fft = n_fft
|
| 26 |
+
self.hop = hop
|
| 27 |
+
self.stem_name = stem_name
|
| 28 |
+
self.compensation = compensation
|
| 29 |
+
|
| 30 |
+
self.n_bins = self.n_fft // 2 + 1
|
| 31 |
+
self.chunk_size = hop * (self.dim_t - 1)
|
| 32 |
+
self.window = torch.hann_window(
|
| 33 |
+
window_length=self.n_fft, periodic=True
|
| 34 |
+
).to(device)
|
| 35 |
+
|
| 36 |
+
out_c = self.dim_c
|
| 37 |
+
|
| 38 |
+
self.freq_pad = torch.zeros(
|
| 39 |
+
[1, out_c, self.n_bins - self.dim_f, self.dim_t]
|
| 40 |
+
).to(device)
|
| 41 |
+
|
| 42 |
+
def stft(self, x):
|
| 43 |
+
"""
|
| 44 |
+
computes the Fourier transform of short overlapping windows of the input
|
| 45 |
+
"""
|
| 46 |
+
x = x.reshape([-1, self.chunk_size])
|
| 47 |
+
x = torch.stft(
|
| 48 |
+
x,
|
| 49 |
+
n_fft=self.n_fft,
|
| 50 |
+
hop_length=self.hop,
|
| 51 |
+
window=self.window,
|
| 52 |
+
center=True,
|
| 53 |
+
return_complex=True,
|
| 54 |
+
)
|
| 55 |
+
x = torch.view_as_real(x)
|
| 56 |
+
x = x.permute([0, 3, 1, 2])
|
| 57 |
+
x = x.reshape([-1, 2, 2, self.n_bins, self.dim_t]).reshape(
|
| 58 |
+
[-1, 4, self.n_bins, self.dim_t]
|
| 59 |
+
)
|
| 60 |
+
return x[:, :, : self.dim_f]
|
| 61 |
+
|
| 62 |
+
def istft(self, x, freq_pad=None):
|
| 63 |
+
"""
|
| 64 |
+
computes the inverse Fourier transform of short overlapping windows of the input
|
| 65 |
+
"""
|
| 66 |
+
freq_pad = (
|
| 67 |
+
self.freq_pad.repeat([x.shape[0], 1, 1, 1])
|
| 68 |
+
if freq_pad is None
|
| 69 |
+
else freq_pad
|
| 70 |
+
)
|
| 71 |
+
x = torch.cat([x, freq_pad], -2)
|
| 72 |
+
# c = 4*2 if self.target_name=='*' else 2
|
| 73 |
+
x = x.reshape([-1, 2, 2, self.n_bins, self.dim_t]).reshape(
|
| 74 |
+
[-1, 2, self.n_bins, self.dim_t]
|
| 75 |
+
)
|
| 76 |
+
x = x.permute([0, 2, 3, 1])
|
| 77 |
+
x = x.contiguous()
|
| 78 |
+
x = torch.view_as_complex(x)
|
| 79 |
+
x = torch.istft(
|
| 80 |
+
x,
|
| 81 |
+
n_fft=self.n_fft,
|
| 82 |
+
hop_length=self.hop,
|
| 83 |
+
window=self.window,
|
| 84 |
+
center=True,
|
| 85 |
+
)
|
| 86 |
+
return x.reshape([-1, 2, self.chunk_size])
|
| 87 |
+
|
| 88 |
+
|
| 89 |
+
class MDX:
|
| 90 |
+
DEFAULT_SR = 44100 # unit: Hz
|
| 91 |
+
# Unit: seconds
|
| 92 |
+
DEFAULT_CHUNK_SIZE = 0 * DEFAULT_SR
|
| 93 |
+
DEFAULT_MARGIN_SIZE = 1 * DEFAULT_SR
|
| 94 |
+
|
| 95 |
+
def __init__(self, model_path: str, params: MDXModel, processor=0):
|
| 96 |
+
# Set the device and the provider (CPU or CUDA)
|
| 97 |
+
self.device = (
|
| 98 |
+
torch.device(f"cuda:{processor}")
|
| 99 |
+
if processor >= 0
|
| 100 |
+
else torch.device("cpu")
|
| 101 |
+
)
|
| 102 |
+
self.provider = (
|
| 103 |
+
["CUDAExecutionProvider"]
|
| 104 |
+
if processor >= 0
|
| 105 |
+
else ["CPUExecutionProvider"]
|
| 106 |
+
)
|
| 107 |
+
|
| 108 |
+
self.model = params
|
| 109 |
+
|
| 110 |
+
# Load the ONNX model using ONNX Runtime
|
| 111 |
+
self.ort = ort.InferenceSession(model_path, providers=self.provider)
|
| 112 |
+
# Preload the model for faster performance
|
| 113 |
+
self.ort.run(
|
| 114 |
+
None,
|
| 115 |
+
{"input": torch.rand(1, 4, params.dim_f, params.dim_t).numpy()},
|
| 116 |
+
)
|
| 117 |
+
self.process = lambda spec: self.ort.run(
|
| 118 |
+
None, {"input": spec.cpu().numpy()}
|
| 119 |
+
)[0]
|
| 120 |
+
|
| 121 |
+
self.prog = None
|
| 122 |
+
|
| 123 |
+
@staticmethod
|
| 124 |
+
def get_hash(model_path: str) -> str:
|
| 125 |
+
try:
|
| 126 |
+
with open(model_path, "rb") as f:
|
| 127 |
+
f.seek(-10000 * 1024, 2)
|
| 128 |
+
model_hash = hashlib.md5(f.read()).hexdigest()
|
| 129 |
+
except: # noqa
|
| 130 |
+
model_hash = hashlib.md5(open(model_path, "rb").read()).hexdigest()
|
| 131 |
+
|
| 132 |
+
return model_hash
|
| 133 |
+
|
| 134 |
+
@staticmethod
|
| 135 |
+
def segment(
|
| 136 |
+
wave,
|
| 137 |
+
combine=True,
|
| 138 |
+
chunk_size=DEFAULT_CHUNK_SIZE,
|
| 139 |
+
margin_size=DEFAULT_MARGIN_SIZE,
|
| 140 |
+
):
|
| 141 |
+
"""
|
| 142 |
+
Segment or join segmented wave array
|
| 143 |
+
|
| 144 |
+
Args:
|
| 145 |
+
wave: (np.array) Wave array to be segmented or joined
|
| 146 |
+
combine: (bool) If True, combines segmented wave array.
|
| 147 |
+
If False, segments wave array.
|
| 148 |
+
chunk_size: (int) Size of each segment (in samples)
|
| 149 |
+
margin_size: (int) Size of margin between segments (in samples)
|
| 150 |
+
|
| 151 |
+
Returns:
|
| 152 |
+
numpy array: Segmented or joined wave array
|
| 153 |
+
"""
|
| 154 |
+
|
| 155 |
+
if combine:
|
| 156 |
+
# Initializing as None instead of [] for later numpy array concatenation
|
| 157 |
+
processed_wave = None
|
| 158 |
+
for segment_count, segment in enumerate(wave):
|
| 159 |
+
start = 0 if segment_count == 0 else margin_size
|
| 160 |
+
end = None if segment_count == len(wave) - 1 else -margin_size
|
| 161 |
+
if margin_size == 0:
|
| 162 |
+
end = None
|
| 163 |
+
if processed_wave is None: # Create array for first segment
|
| 164 |
+
processed_wave = segment[:, start:end]
|
| 165 |
+
else: # Concatenate to existing array for subsequent segments
|
| 166 |
+
processed_wave = np.concatenate(
|
| 167 |
+
(processed_wave, segment[:, start:end]), axis=-1
|
| 168 |
+
)
|
| 169 |
+
|
| 170 |
+
else:
|
| 171 |
+
processed_wave = []
|
| 172 |
+
sample_count = wave.shape[-1]
|
| 173 |
+
|
| 174 |
+
if chunk_size <= 0 or chunk_size > sample_count:
|
| 175 |
+
chunk_size = sample_count
|
| 176 |
+
|
| 177 |
+
if margin_size > chunk_size:
|
| 178 |
+
margin_size = chunk_size
|
| 179 |
+
|
| 180 |
+
for segment_count, skip in enumerate(
|
| 181 |
+
range(0, sample_count, chunk_size)
|
| 182 |
+
):
|
| 183 |
+
margin = 0 if segment_count == 0 else margin_size
|
| 184 |
+
end = min(skip + chunk_size + margin_size, sample_count)
|
| 185 |
+
start = skip - margin
|
| 186 |
+
|
| 187 |
+
cut = wave[:, start:end].copy()
|
| 188 |
+
processed_wave.append(cut)
|
| 189 |
+
|
| 190 |
+
if end == sample_count:
|
| 191 |
+
break
|
| 192 |
+
|
| 193 |
+
return processed_wave
|
| 194 |
+
|
| 195 |
+
def pad_wave(self, wave):
|
| 196 |
+
"""
|
| 197 |
+
Pad the wave array to match the required chunk size
|
| 198 |
+
|
| 199 |
+
Args:
|
| 200 |
+
wave: (np.array) Wave array to be padded
|
| 201 |
+
|
| 202 |
+
Returns:
|
| 203 |
+
tuple: (padded_wave, pad, trim)
|
| 204 |
+
- padded_wave: Padded wave array
|
| 205 |
+
- pad: Number of samples that were padded
|
| 206 |
+
- trim: Number of samples that were trimmed
|
| 207 |
+
"""
|
| 208 |
+
n_sample = wave.shape[1]
|
| 209 |
+
trim = self.model.n_fft // 2
|
| 210 |
+
gen_size = self.model.chunk_size - 2 * trim
|
| 211 |
+
pad = gen_size - n_sample % gen_size
|
| 212 |
+
|
| 213 |
+
# Padded wave
|
| 214 |
+
wave_p = np.concatenate(
|
| 215 |
+
(
|
| 216 |
+
np.zeros((2, trim)),
|
| 217 |
+
wave,
|
| 218 |
+
np.zeros((2, pad)),
|
| 219 |
+
np.zeros((2, trim)),
|
| 220 |
+
),
|
| 221 |
+
1,
|
| 222 |
+
)
|
| 223 |
+
|
| 224 |
+
mix_waves = []
|
| 225 |
+
for i in range(0, n_sample + pad, gen_size):
|
| 226 |
+
waves = np.array(wave_p[:, i:i + self.model.chunk_size])
|
| 227 |
+
mix_waves.append(waves)
|
| 228 |
+
|
| 229 |
+
mix_waves = torch.tensor(mix_waves, dtype=torch.float32).to(
|
| 230 |
+
self.device
|
| 231 |
+
)
|
| 232 |
+
|
| 233 |
+
return mix_waves, pad, trim
|
| 234 |
+
|
| 235 |
+
def _process_wave(self, mix_waves, trim, pad, q: queue.Queue, _id: int):
|
| 236 |
+
"""
|
| 237 |
+
Process each wave segment in a multi-threaded environment
|
| 238 |
+
|
| 239 |
+
Args:
|
| 240 |
+
mix_waves: (torch.Tensor) Wave segments to be processed
|
| 241 |
+
trim: (int) Number of samples trimmed during padding
|
| 242 |
+
pad: (int) Number of samples padded during padding
|
| 243 |
+
q: (queue.Queue) Queue to hold the processed wave segments
|
| 244 |
+
_id: (int) Identifier of the processed wave segment
|
| 245 |
+
|
| 246 |
+
Returns:
|
| 247 |
+
numpy array: Processed wave segment
|
| 248 |
+
"""
|
| 249 |
+
mix_waves = mix_waves.split(1)
|
| 250 |
+
with torch.no_grad():
|
| 251 |
+
pw = []
|
| 252 |
+
for mix_wave in mix_waves:
|
| 253 |
+
self.prog.update()
|
| 254 |
+
spec = self.model.stft(mix_wave)
|
| 255 |
+
processed_spec = torch.tensor(self.process(spec))
|
| 256 |
+
processed_wav = self.model.istft(
|
| 257 |
+
processed_spec.to(self.device)
|
| 258 |
+
)
|
| 259 |
+
processed_wav = (
|
| 260 |
+
processed_wav[:, :, trim:-trim]
|
| 261 |
+
.transpose(0, 1)
|
| 262 |
+
.reshape(2, -1)
|
| 263 |
+
.cpu()
|
| 264 |
+
.numpy()
|
| 265 |
+
)
|
| 266 |
+
pw.append(processed_wav)
|
| 267 |
+
processed_signal = np.concatenate(pw, axis=-1)[:, :-pad]
|
| 268 |
+
q.put({_id: processed_signal})
|
| 269 |
+
return processed_signal
|
| 270 |
+
|
| 271 |
+
def process_wave(self, wave: np.array, mt_threads=1):
|
| 272 |
+
"""
|
| 273 |
+
Process the wave array in a multi-threaded environment
|
| 274 |
+
|
| 275 |
+
Args:
|
| 276 |
+
wave: (np.array) Wave array to be processed
|
| 277 |
+
mt_threads: (int) Number of threads to be used for processing
|
| 278 |
+
|
| 279 |
+
Returns:
|
| 280 |
+
numpy array: Processed wave array
|
| 281 |
+
"""
|
| 282 |
+
self.prog = tqdm(total=0)
|
| 283 |
+
chunk = wave.shape[-1] // mt_threads
|
| 284 |
+
waves = self.segment(wave, False, chunk)
|
| 285 |
+
|
| 286 |
+
# Create a queue to hold the processed wave segments
|
| 287 |
+
q = queue.Queue()
|
| 288 |
+
threads = []
|
| 289 |
+
for c, batch in enumerate(waves):
|
| 290 |
+
mix_waves, pad, trim = self.pad_wave(batch)
|
| 291 |
+
self.prog.total = len(mix_waves) * mt_threads
|
| 292 |
+
thread = threading.Thread(
|
| 293 |
+
target=self._process_wave, args=(mix_waves, trim, pad, q, c)
|
| 294 |
+
)
|
| 295 |
+
thread.start()
|
| 296 |
+
threads.append(thread)
|
| 297 |
+
for thread in threads:
|
| 298 |
+
thread.join()
|
| 299 |
+
self.prog.close()
|
| 300 |
+
|
| 301 |
+
processed_batches = []
|
| 302 |
+
while not q.empty():
|
| 303 |
+
processed_batches.append(q.get())
|
| 304 |
+
processed_batches = [
|
| 305 |
+
list(wave.values())[0]
|
| 306 |
+
for wave in sorted(
|
| 307 |
+
processed_batches, key=lambda d: list(d.keys())[0]
|
| 308 |
+
)
|
| 309 |
+
]
|
| 310 |
+
assert len(processed_batches) == len(
|
| 311 |
+
waves
|
| 312 |
+
), "Incomplete processed batches, please reduce batch size!"
|
| 313 |
+
return self.segment(processed_batches, True, chunk)
|