Michaël Defferrard

You're welcome ! Thanks for considering it. :) Yep the current implementation assumes the graph is fixed, i.e. the same graph Laplacian is used across all samples / signals. You...

Thanks for your interest. To take the structure into account, you can add a constant signal on the graph (while keeping or not the feature signals). Then: * If the...

Hi, thanks for your interest. The current code was not developed with this application in mind, so I don't have any example ready. It might in the future, but for...

Hi halwai, thanks for your interest. Your reasoning seems right. At the input and hidden layers you should not need to care which node in a graph corresponds to which...

Hi @anthony123, thanks for your interest. In such a case you'll have to normalize the size of your data at some point. Graphs of different sizes are not a problem...

I didn't compute tempo features. It should be easy to modify the `compute_features` function in `features.py` to compute them with `librosa` (or any other library) and add them to the...

Branches: * The [`master`](https://github.com/mdeff/fma/tree/master) branch contains the code that produced the **latest released data**. The usage code and documentation can be updated but should work with the released data. *...

Potential todos for a dataset update: * [ ] schema with type and explanation of every field (#14) * [ ] Data format: relational tables (`tracks.csv`, `artists.csv`, `albums.csv`) instead of...

Ouch. Thanks @mgrankin. In the meantime, direct download from should work. What are your reasons to prefer torrents over direct downloads? We considered to setup a seeding server at EPFL...

Thanks for the investigation @ejhumphrey. If I understood correctly, it meas that `105042` MP3s are unique, but there are `105637-105042=595` unique bytestreams out of the remaining `106574-105042=1532` MP3s. Are we...