(infinity) Extend the results of the paper: Beat them!

redouane-dziri / deep-music-classification

Classify music into genres using GLCM on mel-maps and CNNs

0 stars 0 forks source link

Closed redouane-dziri closed 4 years ago

redouane-dziri commented 5 years ago

List of ideas to have better results than the paper's with a similar setting

[x] Normalize input into [0, 1] before feeding to the network
[x] Experiment with other similar neural network architectures. Ideas:
- relu instead of tanh
- more conv filters of smaller size
- more conv layers
- dropout
- max instead of average pooling
[x] Instead of majority voting to get the class of a track, sum up the softmax of each piece and take the max of that
[x] Train for less epochs

arnaudstiegler commented 5 years ago

I would add to this to check for RNN models for that (if it does not require too much work)

redouane-dziri commented 5 years ago

I would add to this to check for RNN models for that (if it does not require too much work)

add it to the checklist!

arnaudstiegler commented 4 years ago

Results are in the evalute_GLCM.ipynb, the concatenated model is much better than the stacked one, especially on the track level