mithgroth / music-genre-recognition-gtzan

MGR on GTZAN

Music genre recognition on GTZAN dataset

• Music genre recognition on GTZAN dataset
  • Understanding GTZAN dataset
  • Training results
    • 30 sec samples (features_30_sec.csv)
    • 3 sec samples (features_3_sec.csv)
    • Image-based training

• Dataset: https://www.kaggle.com/datasets/andradaolteanu/gtzan-dataset-music-genre-classification

Understanding GTZAN dataset

• There are samples of audio files, classified according to their genre.
• All genres have 100 samples.

Training results

In GTZAN dataset there are two CSV files with features:

1. features_30_sec.csv
2. features_3_sec.csv

We'll train on these two sets to observe differences:

30 sec samples (features_30_sec.csv)

First test was done on 30 seconds of samples.

Training details

Epoch	Train Loss	Valid Loss	Accuracy	Time
0	2.410511	2.301348	0.100000	00:00
1	2.358546	2.294634	0.150000	00:00
2	2.308661	2.264426	0.200000	00:00
3	2.234084	2.203711	0.237500	00:00
4	2.171810	2.121932	0.243750	00:00
5	2.107840	2.050719	0.256250	00:00
6	2.039669	1.963892	0.281250	00:00
7	1.977348	1.895904	0.306250	00:00
8	1.910779	1.822734	0.331250	00:00
9	1.841186	1.764186	0.337500	00:00
10	1.774352	1.701079	0.356250	00:00
11	1.703704	1.629614	0.356250	00:00
12	1.633385	1.574726	0.387500	00:00
13	1.564862	1.509628	0.412500	00:00
14	1.501113	1.446676	0.475000	00:00
15	1.449700	1.387999	0.500000	00:00
16	1.374840	1.333856	0.575000	00:00
17	1.309132	1.269413	0.606250	00:00
18	1.251293	1.207156	0.625000	00:00
19	1.191447	1.157811	0.643750	00:00
20	1.127759	1.115797	0.650000	00:00
21	1.067127	1.049386	0.687500	00:00
22	1.009375	1.008271	0.706250	00:00
23	0.951300	0.977570	0.718750	00:00
24	0.893850	0.922465	0.731250	00:00
25	0.838244	0.878590	0.712500	00:00
26	0.787135	0.868914	0.725000	00:00
27	0.732795	0.853740	0.737500	00:00
28	0.678135	0.883555	0.693750	00:00
29	0.625145	0.845650	0.706250	00:00
30	0.583684	0.802656	0.725000	00:00
31	0.532835	0.794608	0.731250	00:00
32	0.489864	0.795410	0.743750	00:00
33	0.443835	0.774997	0.743750	00:00
34	0.409310	0.803072	0.725000	00:00
35	0.375299	0.759579	0.731250	00:00
36	0.340224	0.802695	0.743750	00:00
37	0.306562	0.796328	0.743750	00:00
38	0.280238	0.809968	0.750000	00:00
39	0.254191	0.783211	0.743750	00:00
40	0.228860	0.777353	0.762500	00:00
41	0.208046	0.811578	0.725000	00:00
42	0.186007	0.822938	0.731250	00:00
43	0.170087	0.735695	0.737500	00:00
44	0.159519	0.791559	0.743750	00:00
45	0.149130	0.859067	0.718750	00:00
46	0.136737	0.820262	0.762500	00:00
47	0.122899	0.873859	0.743750	00:00
48	0.116056	0.791122	0.750000	00:00
49	0.107419	0.854898	0.706250	00:00
50	0.100285	0.879887	0.718750	00:00
51	0.089521	0.862034	0.743750	00:00
52	0.081394	0.828892	0.737500	00:00
53	0.073154	0.887935	0.743750	00:00

And classification report results:

Class	Precision	Recall	F1-Score	Support
0 (blues)	0.80	0.80	0.80	20
1 (classical)	0.86	1.00	0.92	12
2 (country)	0.60	0.75	0.67	16
3 (disco)	0.67	0.67	0.67	18
4 (hiphop)	0.92	0.55	0.69	22
5 (jazz)	0.94	0.71	0.81	21
6 (metal)	0.87	0.93	0.90	14
7 (pop)	0.65	0.85	0.73	13
8 (reggae)	0.46	0.67	0.55	9
9 (rock)	0.71	0.67	0.69	15
Accuracy	0.74			160
Macro Avg	0.75	0.76	0.74	160
Weighted Avg	0.77	0.74	0.74	160

Running the training a few more times yielded results between accuracies of 70% to 74%.

3 sec samples (features_3_sec.csv)

Second test was done on 3 seconds of samples.

Training details

Epoch	Train Loss	Valid Loss	Accuracy	Time
0	2.032246	1.873680	0.360451	00:01
1	1.751166	1.637501	0.448686	00:01
2	1.570411	1.483692	0.508761	00:01
3	1.430684	1.365427	0.556320	00:01
4	1.307044	1.269431	0.596370	00:01
5	1.219212	1.175792	0.637672	00:01
6	1.138508	1.086589	0.667710	00:01
7	1.026570	1.008204	0.693367	00:01
8	0.954476	0.930218	0.705882	00:01
9	0.889936	0.864621	0.731539	00:01
10	0.815705	0.815270	0.740926	00:01
11	0.761109	0.770075	0.759700	00:01
12	0.719386	0.718789	0.777847	00:01
13	0.663675	0.683248	0.783479	00:01
14	0.603862	0.653833	0.793492	00:01
15	0.555469	0.619850	0.801627	00:01
16	0.506831	0.575025	0.820400	00:01
17	0.471152	0.554522	0.825407	00:01
18	0.433732	0.531214	0.836045	00:01
19	0.416024	0.520105	0.828536	00:01
20	0.362679	0.504946	0.831039	00:01
21	0.333553	0.481875	0.839800	00:01
22	0.313346	0.454316	0.848561	00:01
23	0.295104	0.459885	0.836671	00:01
24	0.260602	0.424484	0.864205	00:01
25	0.246882	0.449828	0.852941	00:01
26	0.212048	0.420795	0.857322	00:01
27	0.194751	0.419301	0.862328	00:01
28	0.193910	0.427121	0.860451	00:01
29	0.177843	0.416969	0.860451	00:01
30	0.161258	0.408608	0.875469	00:01
31	0.168606	0.395215	0.866083	00:01
32	0.144902	0.402985	0.866083	00:01
33	0.127127	0.395520	0.867334	00:01
34	0.130257	0.389912	0.882979	00:01
35	0.136157	0.402041	0.869837	00:01
36	0.128909	0.403417	0.877347	00:01
37	0.124474	0.409458	0.867334	00:01
38	0.112614	0.412239	0.881727	00:01
39	0.120688	0.421026	0.873592	00:01
40	0.101785	0.377680	0.886733	00:01
41	0.107204	0.402054	0.880476	00:01
42	0.104357	0.389439	0.877347	00:01
43	0.101453	0.407585	0.881101	00:01
44	0.088406	0.386060	0.884230	00:01
45	0.090783	0.396249	0.880476	00:01
46	0.082592	0.372472	0.879850	00:01
47	0.086191	0.373930	0.894243	00:01
48	0.081868	0.362440	0.889862	00:01
49	0.065329	0.384066	0.882353	00:01
50	0.083963	0.380745	0.889862	00:01
51	0.082650	0.417959	0.889862	00:01
52	0.075654	0.388021	0.892365	00:01
53	0.080464	0.403859	0.891740	00:01
54	0.063071	0.371227	0.891114	00:01
55	0.073778	0.353450	0.899249	00:01
56	0.073829	0.391176	0.884230	00:01
57	0.064231	0.365528	0.894243	00:01
58	0.062055	0.373422	0.895494	00:01
59	0.063687	0.391158	0.893617	00:01
60	0.061825	0.417160	0.882979	00:01
61	0.052974	0.394985	0.903004	00:01
62	0.063245	0.395809	0.887985	00:01
63	0.054693	0.358832	0.903004	00:01
64	0.052507	0.388410	0.895494	00:01
65	0.052654	0.352484	0.906758	00:01
66	0.051661	0.413555	0.892991	00:01
67	0.056109	0.394127	0.895494	00:01
68	0.056794	0.372550	0.903004	00:01
69	0.058300	0.374250	0.896746	00:01
70	0.046908	0.352735	0.909262	00:01
71	0.054862	0.368875	0.904881	00:01
72	0.042489	0.400974	0.891740	00:01
73	0.045906	0.384763	0.897372	00:01
74	0.053988	0.396204	0.896746	00:01
75	0.042612	0.384841	0.898623	00:01

And classification report results:

Class	Precision	Recall	F1-Score	Support
0 (blues)	0.94	0.89	0.91	163
1 (classical)	0.91	0.93	0.92	146
2 (country)	0.82	0.82	0.82	161
3 (disco)	0.84	0.93	0.88	168
4 (hiphop)	0.94	0.91	0.93	162
5 (jazz)	0.91	0.91	0.91	151
6 (metal)	0.95	0.94	0.94	172
7 (pop)	0.93	0.90	0.92	150
8 (reggae)	0.93	0.90	0.91	163
9 (rock)	0.82	0.86	0.84	162
Accuracy	0.90			1598
Macro Avg	0.90	0.90	0.90	1598
Weighted Avg	0.90	0.90	0.90	1598

Running the training a few more times yielded results between accuracies around 90%.

Image-based training

TODO

music-genre-recognition-gtzan's Issues

Can I analyse genre of my own audio file using this?

Can I analyse genre of my own audio file using this?

Thanks!