Hi,
thanks for the cool work!
Are there any plans to add the code necessary to evaluate with MV2H as in the paper?
Thanks!

In the forward()-method of the quantisation model, the data is already encoded when the forward()-method of the beat model is called. The data then gets encoded a second time in the beat models forward()-method, leading to a different beat detection in the quantisation model.

$ pip install -r requirements.txt
Collecting torch==1.12.1 (from -r requirements.txt (line 1))
Using cached torch-1.12.1-cp38-cp38-manylinux1_x86_64.whl.metadata (22 kB)
Collecting pretty-midi==0.2.9 (from -r requirements.txt (line 2))
Using cached pretty_midi-0.2.9-py3-none-any.whl
Collecting numpy==1.19.1 (from -r requirements.txt (line 3))
Using cached numpy-1.19.1-cp38-cp38-manylinux2010_x86_64.whl.metadata (2.0 kB)
Collecting mir_eval==0.6 (from -r requirements.txt (line 4))
Using cached mir_eval-0.6-py3-none-any.whl
Collecting mido==1.2.9 (from -r requirements.txt (line 5))
Using cached mido-1.2.9-py2.py3-none-any.whl.metadata (4.1 kB)
Requirement already satisfied: typing-extensions in /root/miniconda3/envs/py38/lib/python3.8/site-packages (from torch==1.12.1->-r requirements.txt (line 1)) (4.10.0)
Requirement already satisfied: six in /root/miniconda3/envs/py38/lib/python3.8/site-packages (from pretty-midi==0.2.9->-r requirements.txt (line 2)) (1.16.0)
Requirement already satisfied: scipy>=1.0.0 in /root/miniconda3/envs/py38/lib/python3.8/site-packages (from mir_eval==0.6->-r requirements.txt (line 4)) (1.9.3)
Requirement already satisfied: future in /root/miniconda3/envs/py38/lib/python3.8/site-packages (from mir_eval==0.6->-r requirements.txt (line 4)) (1.0.0)
Using cached torch-1.12.1-cp38-cp38-manylinux1_x86_64.whl (776.3 MB)
Using cached numpy-1.19.1-cp38-cp38-manylinux2010_x86_64.whl (14.5 MB)
Using cached mido-1.2.9-py2.py3-none-any.whl (52 kB)
Installing collected packages: mido, torch, numpy, pretty-midi, mir_eval

ERROR: pip's dependency resolver does not currently take into account all the packages that are installed. This behaviour is the source of the following dependency conflicts.
pandas 2.0.0 requires numpy>=1.20.3; python_version < "3.10", but you have numpy 1.19.1 which is incompatible.

Successfully installed mido-1.2.9 mir_eval-0.6 numpy-1.19.1 pretty-midi-0.2.9 torch-1.12.1
WARNING: Running pip as the 'root' user can result in broken permissions and conflicting behaviour with the system package manager. It is recommended to use a virtual environment instead: https://pip.pypa.io/warnings/venv

The time signature processor does not seem to work properly. It seems to detect 3/8 or 3/4 almost exclusively, regardless of the processed piece.
Example Code to reproduce the Issue:
from pm2s.features.time_signature import RNNTimeSignatureProcessor
ts_processor = RNNTimeSignatureProcessor()
midi_file = 'Datasets/A-MAPS_1.1/MAPS_MUS-scn15_12_StbgTGd2.mid'
time_sig = ts_processor.process(midi_file)
print(f'Time Signature:{time_sig}')
Example Output:
Time Signature:[(0.500004, '3/8')]

When the checkpoint of the Quantization Model is loaded in the Quantization Processor, it overwrites the beat model checkpoint that has been previously loaded in the Quantization Model. This leads to the beat detection not working properly, because even though the weights and biases stay the same, the parameter num_batches_tracked, which is used for the initialisation of the running mean and running variance of the batch normalisation changes (at least for the supplied model checkpoints), leading the beat detection to behave differently than expected.