The training code for the 4th place model at MDX 2021 leaderboard A
This repository contains the training code of our winning model at Music Demixing Challenge 2021, which got the 4th place on leaderboard A (6th in overall), and help us (Kazane Ryo no Danna) winned the bronze prize.
Model Summary
Our final winning approach blends the outputs from three models, which are:
-
model 1: A X-UMX model [1] which is initialized with the weights of the official baseline, and is fine-tuned with a modified Combinational Multi-Domain Loss from [1]. In particular, we implement and apply a differentiable Multichannel Wiener Filter (MWF) [2] before the loss calculation, and compute the frequency-domain L2 loss with raw complex values.
-
model 2: A U-Net