Figure out best label smoothing parameters about audiotagging

[JoshVarty]

Trying with xresnet-18 because this is taking so long.

Min	Max	lwlrap
0.000	1.000	0.8395
0.000	0.95	0.8432

Min	Max	lwlrap
0.001	0.990	0.8403
0.001	0.975	0.8393
0.001	0.950	0.8415
0.001	0.935	0.8422
0.001	0.900	0.8414
0.001	0.800	0.8421
0.001	0.700	0.8402
0.001	0.500	0.8433
0.001	0.250	0.8405
0.001	0.100	0.8374 (LB 0.669)

Min	Max	lwlrap
0.010	0.990	0.8398
0.010	0.975	0.8412
0.010	0.950	0.8394
0.010	0.935	0.8393
0.010	0.900	0.8368
0.001	0.800	0.8408
0.010	0.700	0.8353
0.010	0.500	0.8400
0.010	0.250	0.8353

Min	Max	lwlrap
0.05	0.990	0.8367

Min	Max	lwlrap
0.4	0.6	0.5722

Min	Max	lwlrap
0.2	0.99	0.7057
0.2	0.8	0.7268
0.2	0.5	0.6673

[JoshVarty]

For the curated data I don't think it matters what we use so long as the minimum isn't too large. Something like 0.001 and 0.990 seems to work reasonably well.

No clue how we should treat the noisy data, though.