Hey! I'm pariticipating to the kaggle competition of free sound (https://www.kaggle.com/c/freesound-audio-tagging)
It's competition without prizes but just wanted to start without the pressure and try to apply the ML knowledge to specific problems.
The way I'm working is with iterations, make something working and after each step think which is gonna to be my next iteration (wich thinks I'm gonna try)
This is the first attemp, on which basically the focus in:
- Loading the data
- Split train/validation
- Having the mapk metric (is how the LeaderBoard is calculated)
- Creating the result csv file
Basically I use two super simple features that librosa give you and use SVC from scikit-learn
Just like the MVP but:
- Uses MFCC (from librosa)
- Does a search of the best parameters for SVM
- ADD 3 CNN models (two with only 1D CNN and another with 2D CNN)
- Since CNN needs a fixed size (for the dense layer)
- Crop the MFCC to 5 seconds, if the audio is shorter, repeat until 5 seconds
- Each clip is splitted in various 5 seconds split with a window that overlaps 1.5 seconds
- Also change the way the csv is computed