🚨 Colab submissions, additional baselines, and more!

Dear participants,

We hope you are having fun with the Music Demixing challenge! :handshake:

We have some exciting updates for you today!

If you were looking for a way to get started, or were facing any issue with your submissions, let us know here! Today’s updates should hopefully make things much easier for you too. :wink:


1. Four new high-quality baselines(XUMXPredictor, Spleeter, Demucs & Lightsaft)

4 new high-quality baselines are now added to the challenge, bringing the total baselines count to 6! :rocket:

Demucs48-HQ (SDR: 6.20)
Demucs is based on U-Net convolutional architecture inspired by Wave-U-Net and SING, with GLUs, a BiLSTM between the encoder and decoder, specific initialization of weights and transposed convolutions in the decoder.

Spleeter – 4stem (SDR: 5.98)
Spleeter is Deezer source separation library with pretrained models written in Python and uses Tensorflow. It makes it easy to train source separation model, and provides already trained state of the art models.

Lightsaft (SDR: 5.73)
This implementation is a very light version of the LaSAFT+GPoCM model proposed in LaSAFT paper.

XUMXPredictor (SDR: 5.83)
This uses CrossNet-UMX (X-UMX) for music demixing.

:point_right: Access Baselines


2. Submission via Colab

We have now enabled Colab notebook submissions for the challenge! Check out the beta version of the Notebook - you can explore different baselines right from inside Colab and also make your submission!

If you are resource-constrained or would not like to setup everything in your system, we hope this Notebook helps you make your submission easily! If you have any queries or issues, let us know! :smile:

:point_right: Make your submissions through Colab


3. Too overwhelming, where to start?

You can visit a curated list of getting started resources and Ideas for improving the baselines.

:point_right: Deep dive into Music Demixing Domain


Nudge:
Please remember to update the “external_dataset_used ” parameter in your aicrowd.json for separating your submissions between Leaderboard A & Leaderboard B. If you’ve added wrong value in your past submissions, let us know to get it changed.


2 Likes