How can we train these models on our own data?

[Peter Mitrano]

I'm interested in training (a modifed version of) the DBNDownBeatTracker, but I'm unsure where to begin. It's quite easy to run them, but I don't see any hints on how to train one myself.

Thanks in advance!

[Sebastian Böck]

Hi,

unfortunately we don't provide code to train the models and I am not sure if this will ever happen.

Having this said, training a model is (more or less) straight forward. You can use the same signal pre-processing as in our code. Simply add some targets and train a model with whatever library/toolkit. Training parameters are published in the respective paper. Then, only conversion from the trained model into madmom is needed. There's an example in the madmom paper (ACMMM version, Listing 4).

HTH

[Peter Mitrano]

Ah, what a shame. May I ask what framework you use for training the networks?

[Sebastian Böck]

For historical reasons I trained most RNNs with Alex Graves' RNNLIB, whereas all newer networks were trained with Lasagne.