Try without imagenet normalization about audiotagging

[JoshVarty]

Using Classifier architecture

Transform + ImageNet Normalization

97	0.085756	0.028508	0.674055	00:54
98	0.085725	0.028905	0.672101	00:54
99	0.085808	0.028740	0.676140	00:54

Transform + Normalization over batch

97	0.085892	0.028007	0.682177	00:54
98	0.084816	0.028324	0.677291	00:53
99	0.085583	0.027846	0.681515	00:54

Transform

97	0.086221	0.029027	0.665488	00:54
98	0.086549	0.028947	0.666236	00:53
99	0.086307	0.028842	0.668887	00:53

Nothing

97	0.073415	0.034966	0.546998	00:53
98	0.072962	0.035036	0.547777	00:53
99	0.072679	0.034979	0.549354	00:53

Transform + Our Stats

stats = [[0.38895802705981103, 0.38895802705981103, 0.38895802705981103], [0.2345895250966713, 0.2345895250966713, 0.2345895250966713]]

97	0.086323	0.029946	0.652789	00:53
98	0.086527	0.055577	0.656116	00:53
99	0.086919	0.038345	0.657061	00:53

[JoshVarty]

Normalization over batch seems to be at least as good as any other approach, so we'll use it.