Add ESC-10 / 50 / US dataset

[lcances]

🚀 Feature

The ESC-10/50 dataset is widely used and no yet available in torchaudio.datasets. More Information

Motivation

• This dataset is often used as a reference in the same way as its computer vision equivalent: CIFAR or MNIST
• It can be useful for tutorial purposes.
• Its 'US' variant makes it suitable for semi-supervised, self-supervised, and unsupervised learning.
• The ESC-50 and ESC-10 are rather small and may be useful to perform experiments quickly.

Pitch

I am currently using this dataset in my research and I tend to follow the PyTorch way to implement datasets. Therefore, it is almost ready to be submitted here.

• The structure is similar to the already available datasets.
• Downloading the dataset is possible.

I should be able to deliver the code for the end of the week.

[mthrok]

Hi @leocances

Thanks for the suggestion.

It seems to me adding this dataset is straightforward and won't be an issue. @vincentqb What do you think?

For the clarification, is this the reference paper? Could you give us an idea of what are the common use-cases / application / research field of this dataset?

[vincentqb]

@leocances -- yes, please feel free to open a pull request for this! it was also mentioned here :)

[lcances]

Hello, Sorry for the late reply, I had something else on hand and I had to fix some error. (And I need to learn how to do a pull request)

For the clarification, is this the reference paper?

It is indeed the reference paper

Could you give us an idea of what are the common use-cases / application / research field of this dataset?

• ESC-10 is a small dataset with a limited amount of classes. Training on it is fast and allows quick experimentation.
• ESC-50 is still a small dataset but with much fewer examples for each class (only 40 samples). It is therefore a more challenging dataset.
• ESC-US is a large dataset of 250k audio samples and can be used for unsupervised learning. Coupled with ESC-10/50 it can be used for semi-supervised learning.

These datasets are mainly used for sound event classification and sound event detection (when coupled with an activity detector).

The following github is interesting to see the diversity of the system that are tested on the ESC datasets