-
Notifications
You must be signed in to change notification settings - Fork 388
/
si_sdr.py
64 lines (52 loc) · 2.2 KB
/
si_sdr.py
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
# Copyright The PyTorch 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 torch
from torch import Tensor
from torchmetrics.utilities.checks import _check_same_shape
def si_sdr(preds: Tensor, target: Tensor, zero_mean: bool = False) -> Tensor:
"""Calculates Scale-invariant signal-to-distortion ratio (SI-SDR) metric. The SI-SDR value is in general
considered an overall measure of how good a source sound.
Args:
preds:
shape ``[...,time]``
target:
shape ``[...,time]``
zero_mean:
If to zero mean target and preds or not
Returns:
si-sdr value of shape [...]
Example:
>>> from torchmetrics.functional.audio import si_sdr
>>> target = torch.tensor([3.0, -0.5, 2.0, 7.0])
>>> preds = torch.tensor([2.5, 0.0, 2.0, 8.0])
>>> si_sdr_val = si_sdr(preds, target)
>>> si_sdr_val
tensor(18.4030)
References:
[1] Le Roux, Jonathan, et al. "SDR half-baked or well done." IEEE International Conference on Acoustics, Speech
and Signal Processing (ICASSP) 2019.
"""
_check_same_shape(preds, target)
EPS = torch.finfo(preds.dtype).eps
if zero_mean:
target = target - torch.mean(target, dim=-1, keepdim=True)
preds = preds - torch.mean(preds, dim=-1, keepdim=True)
alpha = (torch.sum(preds * target, dim=-1, keepdim=True) + EPS) / (
torch.sum(target ** 2, dim=-1, keepdim=True) + EPS
)
target_scaled = alpha * target
noise = target_scaled - preds
si_sdr_value = (torch.sum(target_scaled ** 2, dim=-1) + EPS) / (torch.sum(noise ** 2, dim=-1) + EPS)
si_sdr_value = 10 * torch.log10(si_sdr_value)
return si_sdr_value