Metadata-Version: 2.1
Name: auraloss
Version: 0.4.0
Summary: Collection of audio-focused loss functions in PyTorch.
Home-page: https://github.com/csteinmetz1/auraloss
Author: Christian Steinmetz
Author-email: c.j.steinmetz@qmul.ac.uk
License: Apache License 2.0
Classifier: License :: OSI Approved :: Apache Software License
Classifier: Topic :: Multimedia :: Sound/Audio
Classifier: Topic :: Scientific/Engineering
Requires-Python: >=3.6.0
Description-Content-Type: text/markdown
Provides-Extra: test
Provides-Extra: all
License-File: LICENSE


<div  align="center">

# auraloss

<img width="200px" src="docs/auraloss-logo.svg">

A collection of audio-focused loss functions in PyTorch. 

[[PDF](https://www.christiansteinmetz.com/s/DMRN15__auraloss__Audio_focused_loss_functions_in_PyTorch.pdf)]

</div>

## Setup

```
pip install auraloss
```

If you want to use `MelSTFTLoss()` or `FIRFilter()` you will need to specify the extra install (librosa and scipy).

```
pip install auraloss[all]
```

## Usage

```python
import torch
import auraloss

mrstft = auraloss.freq.MultiResolutionSTFTLoss()

input = torch.rand(8,1,44100)
target = torch.rand(8,1,44100)

loss = mrstft(input, target)
```

**NEW**: Perceptual weighting with mel scaled spectrograms.

```python

bs = 8
chs = 1
seq_len = 131072
sample_rate = 44100

# some audio you want to compare
target = torch.rand(bs, chs, seq_len)
pred = torch.rand(bs, chs, seq_len)

# define the loss function
loss_fn = auraloss.freq.MultiResolutionSTFTLoss(
    fft_sizes=[1024, 2048, 8192],
    hop_sizes=[256, 512, 2048],
    win_lengths=[1024, 2048, 8192],
    scale="mel",
    n_bins=128,
    sample_rate=sample_rate,
    perceptual_weighting=True,
)

# compute
loss = loss_fn(pred, target)

```

## Citation
If you use this code in your work please consider citing us.
```bibtex
@inproceedings{steinmetz2020auraloss,
    title={auraloss: {A}udio focused loss functions in {PyTorch}},
    author={Steinmetz, Christian J. and Reiss, Joshua D.},
    booktitle={Digital Music Research Network One-day Workshop (DMRN+15)},
    year={2020}
}
```


# Loss functions

We categorize the loss functions as either time-domain or frequency-domain approaches. 
Additionally, we include perceptual transforms.

<table>
    <tr>
        <th>Loss function</th>
        <th>Interface</th>
        <th>Reference</th>
    </tr>
    <tr>
        <td colspan="3" align="center"><b>Time domain</b></td>
    </tr>
    <tr>
        <td>Error-to-signal ratio (ESR)</td>
        <td><code>auraloss.time.ESRLoss()</code></td>
        <td><a href=https://arxiv.org/abs/1911.08922>Wright & Välimäki, 2019</a></td>
    </tr>
    <tr>
        <td>DC error (DC)</td>
        <td><code>auraloss.time.DCLoss()</code></td>
        <td><a href=https://arxiv.org/abs/1911.08922>Wright & Välimäki, 2019</a></td>
    </tr>
    <tr>
        <td>Log hyperbolic cosine (Log-cosh)</td>
        <td><code>auraloss.time.LogCoshLoss()</code></td>
        <td><a href=https://openreview.net/forum?id=rkglvsC9Ym>Chen et al., 2019</a></td>
    </tr>
    <tr>
        <td>Signal-to-noise ratio (SNR)</td>
        <td><code>auraloss.time.SNRLoss()</code></td>
        <td></td>
    </tr>
    <tr>
        <td>Scale-invariant signal-to-distortion <br>  ratio (SI-SDR)</td>
        <td><code>auraloss.time.SISDRLoss()</code></td>
        <td><a href=https://arxiv.org/abs/1811.02508>Le Roux et al., 2018</a></td>
    </tr>
    <tr>
        <td>Scale-dependent signal-to-distortion <br>  ratio (SD-SDR)</td>
        <td><code>auraloss.time.SDSDRLoss()</code></td>
        <td><a href=https://arxiv.org/abs/1811.02508>Le Roux et al., 2018</a></td>
    </tr>
    <tr>
        <td colspan="3" align="center"><b>Frequency domain</b></td>
    </tr>
    <tr>
        <td>Aggregate STFT</td>
        <td><code>auraloss.freq.STFTLoss()</code></td>
        <td><a href=https://arxiv.org/abs/1808.06719>Arik et al., 2018</a></td>
    </tr>
    <tr>
        <td>Aggregate Mel-scaled STFT</td>
        <td><code>auraloss.freq.MelSTFTLoss(sample_rate)</code></td>
        <td></td>
    </tr>
    <tr>
        <td>Multi-resolution STFT</td>
        <td><code>auraloss.freq.MultiResolutionSTFTLoss()</code></td>
        <td><a href=https://arxiv.org/abs/1910.11480>Yamamoto et al., 2019*</a></td>
    </tr>
    <tr>
        <td>Random-resolution STFT</td>
        <td><code>auraloss.freq.RandomResolutionSTFTLoss()</code></td>
        <td><a href=https://www.christiansteinmetz.com/s/DMRN15__auraloss__Audio_focused_loss_functions_in_PyTorch.pdf>Steinmetz & Reiss, 2020</a></td>
    </tr>
    <tr>
        <td>Sum and difference STFT loss</td>
        <td><code>auraloss.freq.SumAndDifferenceSTFTLoss()</code></td>
        <td><a href=https://arxiv.org/abs/2010.10291>Steinmetz et al., 2020</a></td>
    </tr>
    <tr>
        <td colspan="3" align="center"><b>Perceptual transforms</b></td>
    </tr>
    <tr>
        <td>Sum and difference signal transform</td>
        <td><code>auraloss.perceptual.SumAndDifference()</code></td>
        <td><a href=#></a></td>
    </tr>
    <tr>
        <td>FIR pre-emphasis filters</td>
        <td><code>auraloss.perceptual.FIRFilter()</code></td>
        <td><a href=https://arxiv.org/abs/1911.08922>Wright & Välimäki, 2019</a></td>
    </tr>
</table>

\* [Wang et al., 2019](https://arxiv.org/abs/1904.12088) also propose a multi-resolution spectral loss (that [Engel et al., 2020](https://arxiv.org/abs/2001.04643) follow), 
but they do not include both the log magnitude (L1 distance) and spectral convergence terms, introduced in [Arik et al., 2018](https://arxiv.org/abs/1808.0671), and then extended for the multi-resolution case in [Yamamoto et al., 2019](https://arxiv.org/abs/1910.11480).

## Examples

Currently we include an example using a set of the loss functions to train a TCN for modeling an analog dynamic range compressor. 
For details please refer to the details in [`examples/compressor`](examples/compressor). 
We provide pre-trained models, evaluation scripts to compute the metrics in the [paper](https://www.christiansteinmetz.com/s/DMRN15__auraloss__Audio_focused_loss_functions_in_PyTorch.pdf), as well as scripts to retrain models. 

There are some more advanced things you can do based upon the `STFTLoss` class. 
For example, you can compute both linear and log scaled STFT errors as in [Engel et al., 2020](https://arxiv.org/abs/2001.04643).
In this case we do not include the spectral convergence term. 
```python
stft_loss = auraloss.freq.STFTLoss(
    w_log_mag=1.0, 
    w_lin_mag=1.0, 
    w_sc=0.0,
)
```

There is also a Mel-scaled STFT loss, which has some special requirements. 
This loss requires you set the sample rate as well as specify the correct device. 
```python
sample_rate = 44100
melstft_loss = auraloss.freq.MelSTFTLoss(sample_rate, device="cuda")
```

You can also build a multi-resolution Mel-scaled STFT loss with 64 bins easily. 
Make sure you pass the correct device where the tensors you are comparing will be. 
```python
loss_fn = auraloss.freq.MultiResolutionSTFTLoss(
    scale="mel", 
    n_bins=64,
    sample_rate=sample_rate,
    device="cuda"
)
```

If you are computing a loss on stereo audio you may want to consider the sum and difference (mid/side) loss. 
Below we have shown an example of using this loss function with the perceptual weighting and mel scaling for 
further perceptual relevance. 

```python

target = torch.rand(8, 2, 44100)
pred = torch.rand(8, 2, 44100)

loss_fn = auraloss.freq.SumAndDifferenceSTFTLoss(
    fft_sizes=[1024, 2048, 8192],
    hop_sizes=[256, 512, 2048],
    win_lengths=[1024, 2048, 8192],
    perceptual_weighting=True,
    sample_rate=44100,
    scale="mel",
    n_bins=128,
)

loss = loss_fn(pred, target)
```

# Development

Run tests locally with pytest. 

```python -m pytest```
