Hello everyone.<\/p>\n
Recently I embarked on the adventure of implementing a Guitar Chord classifier. My goal is to classify 14 guitar chords (A, Am, B, Bm, C, Cm, D, Dm, E, Em, F, Fm, G, Gm). Unfortunately, I found no publicly available datasets containing all of the chords as most contained only the non-minor chords. Therefore, I decided to build my dataset.<\/p>\n
I shoot three different videos for each chord in three different locations of two different people playing the chords. The videos were shot in 4k 60fps so I had plenty of HQ frames to choose from.<\/p>\n
I sampled 250 images for training, 100 for valid validating, and 100 for testing. After that, using Robflow I augmented my data, which is now publically available<\/a>.<\/p>\n I started training my classifier with this data (MobileNetv2) and I was surprised to find that the model achieved almost 100% in the validation set (97%). I’m trying to figure out why that happened because it seems very suspicious to me but till now I cannot pinpoint the problem. I generated a PCA plot for each of the splits which can be found here: PCA<\/a><\/p>\n Any feedback would be appreciated \ud83d\ude42<\/p>\n<\/div>\n submitted by \/u\/dduka99 <\/a>
[link]<\/a><\/span> [comments]<\/a><\/span><\/p>
