Consider other models about audiotagging

[JoshVarty]

Conv 64-128-256-512

97	0.086184	0.028476	0.678350	00:52
98	0.085402	0.028466	0.679489	00:52
99	0.085432	0.028146	0.682341	00:52

[JoshVarty]

xresnet50

97	0.078883	0.025930	0.718647	00:51
98	0.078497	0.026046	0.718951	00:49
99	0.078972	0.025905	0.716066	00:51

Seems to be a bit better

[JoshVarty]

xresnet18

97	0.074626	0.028025	0.690300	00:36
98	0.075188	0.027592	0.690504	00:36
99	0.074820	0.027749	0.687696	00:36

[JoshVarty]

xresnet101

97	0.080427	0.025303	0.717841	01:13
98	0.079735	0.025446	0.712157	01:13
99	0.080249	0.025226	0.715997	01:13

*I should mention that there were a few iterations in which it reached an f-score >0.72

[JoshVarty]

xresnet101 with stacked random crops

97	0.076211	0.026887	0.702940	01:15
98	0.076367	0.026963	0.685226	01:16
99	0.076069	0.027496	0.682742	01:14

[JoshVarty]

xresnet101 with stacked sequential crops

97	0.074217	0.027665	0.680261	01:16
98	0.074828	0.027505	0.690851	01:14
99	0.074364	0.027654	0.696997	01:12

[JoshVarty]

xresnet101 with 256x256 images*

97	0.084554	0.342670	0.635889	04:06
98	0.084848	0.505083	0.661808	04:05
99	0.084444	0.454247	0.670786	04:05

*One thing I noticed is that validation loss is still fluctuating. This may be a sign that we just have to train for more epochs.