Fix loading weights even if --no-weights is set.

Allow to specify the learning rate

Fix #840

keras sequence methods for generator

Float16 fixes

Avoid importing numpy in setup.py

Small questioning ...

Wouldnt it be better to print both mAP values since the evaluation takes some time already for some datasets? It will take twice this time if we want to compute these values separately... (this is happening too me right now and I really start to find this annoying :) ).

OID annotation file parse bug?

Update ReduceLROnPlateau callback parameter name

keras warning: `epsilon` argument is deprecated and will be removed, use `min_delta` instead.

Replace printing with progressbar.