File size: 11,844 Bytes
9c6594c |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 |
# Copyright The Lightning team.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import os
from functools import lru_cache
from typing import Any, Optional
import numpy as np
import torch
from torch import Tensor
from torchmetrics.utilities import rank_zero_info, rank_zero_warn
from torchmetrics.utilities.imports import _LIBROSA_AVAILABLE, _ONNXRUNTIME_AVAILABLE, _REQUESTS_AVAILABLE
if _LIBROSA_AVAILABLE and _ONNXRUNTIME_AVAILABLE and _REQUESTS_AVAILABLE:
import librosa
import onnxruntime as ort
import requests
from onnxruntime import InferenceSession
else:
librosa, ort, requests = None, None, None # type:ignore
class InferenceSession: # type:ignore
"""Dummy InferenceSession."""
def __init__(self, **kwargs: dict[str, Any]) -> None: ...
__doctest_requires__ = {
("deep_noise_suppression_mean_opinion_score", "_load_session"): ["requests", "librosa", "onnxruntime"]
}
SAMPLING_RATE = 16000
INPUT_LENGTH = 9.01
DNSMOS_DIR = "~/.torchmetrics/DNSMOS"
def _prepare_dnsmos(dnsmos_dir: str) -> None:
"""Download required DNSMOS files.
Args:
dnsmos_dir: a dir to save the downloaded files. Defaults to "~/.torchmetrics".
"""
# https://raw.githubusercontent.com/microsoft/DNS-Challenge/master/DNSMOS/DNSMOS/model_v8.onnx
# https://raw.githubusercontent.com/microsoft/DNS-Challenge/master/DNSMOS/DNSMOS/sig_bak_ovr.onnx
# https://raw.githubusercontent.com/microsoft/DNS-Challenge/master/DNSMOS/pDNSMOS/sig_bak_ovr.onnx
url = "https://raw.githubusercontent.com/microsoft/DNS-Challenge/master"
dnsmos_dir = os.path.expanduser(dnsmos_dir)
# save to or load from ~/torchmetrics/dnsmos/.
for file in ["DNSMOS/DNSMOS/model_v8.onnx", "DNSMOS/DNSMOS/sig_bak_ovr.onnx", "DNSMOS/pDNSMOS/sig_bak_ovr.onnx"]:
saveto = os.path.join(dnsmos_dir, file[7:])
os.makedirs(os.path.dirname(saveto), exist_ok=True)
if os.path.exists(saveto):
# try loading onnx
try:
_ = InferenceSession(saveto)
continue # skip downloading if succeeded
except Exception as _:
os.remove(saveto)
urlf = f"{url}/{file}"
rank_zero_info(f"downloading {urlf} to {saveto}")
myfile = requests.get(urlf)
with open(saveto, "wb") as f:
f.write(myfile.content)
def _load_session(
path: str,
device: torch.device,
num_threads: Optional[int] = None,
) -> InferenceSession:
"""Load onnxruntime session.
Args:
path: the model path
device: the device used
num_threads: the number of threads to use. Defaults to None.
Returns:
onnxruntime session
"""
path = os.path.expanduser(path)
if not os.path.exists(path):
_prepare_dnsmos(DNSMOS_DIR)
opts = ort.SessionOptions()
if num_threads is not None:
opts.inter_op_num_threads = num_threads
opts.intra_op_num_threads = num_threads
if device.type == "cpu":
infs = InferenceSession(path, providers=["CPUExecutionProvider"], sess_options=opts)
elif "CUDAExecutionProvider" in ort.get_available_providers(): # win or linux with cuda
providers = ["CUDAExecutionProvider", "CPUExecutionProvider"]
provider_options = [{"device_id": device.index}, {}]
infs = InferenceSession(path, providers=providers, provider_options=provider_options, sess_options=opts)
elif "CoreMLExecutionProvider" in ort.get_available_providers(): # macos with coreml
providers = ["CoreMLExecutionProvider", "CPUExecutionProvider"]
provider_options = [{"device_id": device.index}, {}]
infs = InferenceSession(path, providers=providers, provider_options=provider_options, sess_options=opts)
else:
infs = InferenceSession(path, providers=["CPUExecutionProvider"], sess_options=opts)
return infs
_cached_load_session = lru_cache()(_load_session)
def _audio_melspec(
audio: np.ndarray,
n_mels: int = 120,
frame_size: int = 320,
hop_length: int = 160,
sr: int = 16000,
to_db: bool = True,
) -> np.ndarray:
"""Calculate the mel-spectrogram of an audio.
Args:
audio: [..., T]
n_mels: the number of mel-frequencies
frame_size: stft length
hop_length: stft hop length
sr: sample rate of audio
to_db: convert to dB scale if `True` is given
Returns:
mel-spectrogram: [..., num_mel, T']
"""
shape = audio.shape
audio = audio.reshape(-1, shape[-1])
mel_spec = librosa.feature.melspectrogram(
y=audio, sr=sr, n_fft=frame_size + 1, hop_length=hop_length, n_mels=n_mels
)
mel_spec = mel_spec.transpose(0, 2, 1)
mel_spec = mel_spec.reshape(shape[:-1] + mel_spec.shape[1:])
if to_db:
for b in range(mel_spec.shape[0]):
mel_spec[b, ...] = (librosa.power_to_db(mel_spec[b], ref=np.max) + 40) / 40
return mel_spec
def _polyfit_val(mos: np.ndarray, personalized: bool) -> np.ndarray:
"""Use polyfit to convert raw mos values to DNSMOS values.
Args:
mos: the raw mos values, [..., 4]
personalized: whether interfering speaker is penalized
Returns:
DNSMOS: [..., 4]
"""
if personalized:
p_ovr = np.poly1d([-0.00533021, 0.005101, 1.18058466, -0.11236046])
p_sig = np.poly1d([-0.01019296, 0.02751166, 1.19576786, -0.24348726])
p_bak = np.poly1d([-0.04976499, 0.44276479, -0.1644611, 0.96883132])
else:
p_ovr = np.poly1d([-0.06766283, 1.11546468, 0.04602535])
p_sig = np.poly1d([-0.08397278, 1.22083953, 0.0052439]) # x**2*v0 + x**1*v1+ v2
p_bak = np.poly1d([-0.13166888, 1.60915514, -0.39604546])
mos[..., 1] = p_sig(mos[..., 1])
mos[..., 2] = p_bak(mos[..., 2])
mos[..., 3] = p_ovr(mos[..., 3])
return mos
def deep_noise_suppression_mean_opinion_score(
preds: Tensor,
fs: int,
personalized: bool,
device: Optional[str] = None,
num_threads: Optional[int] = None,
cache_session: bool = True,
) -> Tensor:
"""Calculate `Deep Noise Suppression performance evaluation based on Mean Opinion Score`_ (DNSMOS).
Human subjective evaluation is the ”gold standard” to evaluate speech quality optimized for human perception.
Perceptual objective metrics serve as a proxy for subjective scores. The conventional and widely used metrics
require a reference clean speech signal, which is unavailable in real recordings. The no-reference approaches
correlate poorly with human ratings and are not widely adopted in the research community. One of the biggest
use cases of these perceptual objective metrics is to evaluate noise suppression algorithms. DNSMOS generalizes
well in challenging test conditions with a high correlation to human ratings in stack ranking noise suppression
methods. More details can be found in `DNSMOS paper <https://arxiv.org/abs/2010.15258>`_ and
`DNSMOS P.835 paper <https://arxiv.org/abs/2110.01763>`_.
.. hint::
Using this metric requires you to have ``librosa``, ``onnxruntime`` and ``requests`` installed. Install
as ``pip install torchmetrics['audio']`` or alternatively ``pip install librosa onnxruntime-gpu requests``
(if you do not have GPU enabled machine install ``onnxruntime`` instead of ``onnxruntime-gpu``)
Args:
preds: [..., time]
fs: sampling frequency
personalized: whether interfering speaker is penalized
device: the device used for calculating DNSMOS, can be cpu or cuda:n, where n is the index of gpu.
If None is given, then the device of input is used.
num_threads: the number of threads to use for cpu inference. Defaults to None.
cache_session: whether to cache the onnx session. By default this is true, meaning that repeated calls to this
method is faster than if this was set to False, the consequence is that the session will be cached in
memory until the process is terminated.
Returns:
Float tensor with shape ``(...,4)`` of DNSMOS values per sample, i.e. [p808_mos, mos_sig, mos_bak, mos_ovr]
Raises:
ModuleNotFoundError:
If ``librosa``, ``onnxruntime`` or ``requests`` packages are not installed
Example:
>>> from torch import randn
>>> from torchmetrics.functional.audio.dnsmos import deep_noise_suppression_mean_opinion_score
>>> preds = randn(8000)
>>> deep_noise_suppression_mean_opinion_score(preds, 8000, False)
tensor([2.2..., 2.0..., 1.1..., 1.2...], dtype=torch.float64)
"""
if not _LIBROSA_AVAILABLE or not _ONNXRUNTIME_AVAILABLE or not _REQUESTS_AVAILABLE:
raise ModuleNotFoundError(
"DNSMOS metric requires that librosa, onnxruntime and requests are installed."
" Install as `pip install librosa onnxruntime-gpu requests`."
)
device = torch.device(device) if device is not None else preds.device
_load_session_function = _cached_load_session if cache_session else _load_session
onnx_sess = _load_session_function(
f"{DNSMOS_DIR}/{'p' if personalized else ''}DNSMOS/sig_bak_ovr.onnx", device, num_threads
)
p808_onnx_sess = _load_session_function(f"{DNSMOS_DIR}/DNSMOS/model_v8.onnx", device, num_threads)
desired_fs = SAMPLING_RATE
if fs != desired_fs:
audio = librosa.resample(preds.cpu().numpy(), orig_sr=fs, target_sr=desired_fs)
else:
audio = preds.cpu().numpy()
len_samples = int(INPUT_LENGTH * desired_fs)
while audio.shape[-1] < len_samples:
audio = np.concatenate([audio, audio], axis=-1)
num_hops = int(np.floor(audio.shape[-1] / desired_fs) - INPUT_LENGTH) + 1
moss = []
hop_len_samples = desired_fs
for idx in range(num_hops):
audio_seg = audio[..., int(idx * hop_len_samples) : int((idx + INPUT_LENGTH) * hop_len_samples)]
if audio_seg.shape[-1] < len_samples:
continue
shape = audio_seg.shape
audio_seg = audio_seg.reshape((-1, shape[-1]))
input_features = np.array(audio_seg).astype("float32")
p808_input_features = np.array(_audio_melspec(audio=audio_seg[..., :-160])).astype("float32")
if device.type != "cpu" and (
"CUDAExecutionProvider" in ort.get_available_providers()
or "CoreMLExecutionProvider" in ort.get_available_providers()
):
try:
input_features = ort.OrtValue.ortvalue_from_numpy(input_features, device.type, device.index)
p808_input_features = ort.OrtValue.ortvalue_from_numpy(p808_input_features, device.type, device.index)
except Exception as e:
rank_zero_warn(f"Failed to use GPU for DNSMOS, reverting to CPU. Error: {e}")
oi = {"input_1": input_features}
p808_oi = {"input_1": p808_input_features}
mos_np = np.concatenate(
[p808_onnx_sess.run(None, p808_oi)[0], onnx_sess.run(None, oi)[0]], axis=-1, dtype="float64"
)
mos_np = _polyfit_val(mos_np, personalized)
mos_np = mos_np.reshape(shape[:-1] + (4,))
moss.append(mos_np)
return torch.from_numpy(np.mean(np.stack(moss, axis=-1), axis=-1)) # [p808_mos, mos_sig, mos_bak, mos_ovr]
|