TODO

[Howe-Young]

1. Batch All
2. Batch All(not equal zero)
3. Batch Hard
4. network architecture
5. soft margin
6. distance(cosine distance)
7. device B, C

[hackerekcah]

network arch

consider residual block as in deep speaker, but with less channels

[hackerekcah]

mean pooling over time axis instead of last frame of LSTM