julius-richter/beat_tracker

Style-specific beat tracking with deep neural networks

beat_tracker's Issues

Squeeze function called to reduce tensor dimensions on an input with non singleton dimensions.

Hi! Thank you for this nice collection of code to work on this interesting problem.

Unfortunately I'm having some difficulties reproducing your results. In particular I am not sure how you preprocess your audio data. In some places you use multiple frame sizes. This is important because depending on the preprocessing method you get different shapes for the input data for the model. The TCN_Davies model uses a series of 2d convolutions before squeezing away the last dimension of the input and feeding the result to a 1d CNN. However, my input data has the wrong shape and thus I feed a 4 dimensional tensor to the TCN because the squeeze function doesn't do anything.

beat_tracker/python/models.py

Line 45 in b5b9971

I am trying to use the model according to instructions here https://github.com/julius-richter/beat_tracker/blob/master/jupyter/process.ipynb

Hope you can help,

Best,

Thor

Not receiving same values as in post_processing.ipynb

Hi,
Thank you for sharing the bar pointer model for beat tracker.

I was using your beat tracker post_processing.ipynb and I had a small problem decoding the path and unable to understand where the problem might be.
I see there are two implementations of Viterbi algorithm one in python and onother in Cpython.

When I tried on my data (beat activation from neural network) on Viterbi implemted in the python I do not get the path corrent. I get an array of zeros where as Cpython implementation works fine. The same happens when I tried to re run with the same example mentioned observations = np.sin(np.linspace(-0.2, 10, 51)).clip(0.0, 1.0)

Do you know why this is happening?
Thank you,
Christon