Keunwoo Choi, May 2017

A repo to host convnet features for FMA. Wanted to do more experiment than baseline method, but don't think I have time to do so, so just releasing it.

A set of features that is computed from transfer learning repo

Free Music Archieve dataset: A large collection of audio and its genre annotation released in 2017.

• In general, it achieved better performance than some popular audio features as in this paper
• For FMA dataset, it's bit better than the provided audio features -- showed 63.94% of accuracy. +1% improvement over baseline. Although I think this convnet feature also should be considered as one of baseline features.

# Load the features
import numpy as np
feat1 = np.load('fma_large_layer1.npy')
feat2 = np.load('fma_large_layer2.npy')
feat3 = np.load('fma_large_layer3.npy')
feat4 = np.load('fma_large_layer4.npy')
feat5 = np.load('fma_large_layer5.npy')

# concatenate the features
features = np.concatenate((feat1, feat2, feat3, feat4, feat5), axis=1)
features.shape
# (106574, 160)
# This is matched to the order of metadata.csv provided in FMA.
# Now use it for your task!

• This is computed from FMA large, which is 30-second previews clips