Some details about the baselines.

Hello,
It is very nice for you to open this code, and i have some questions about the baselines and other details.
Firstly, the key "value" in json file is derived from the evaluation step or training step?
Secondly, i find it spend 7 hours to train only 0.6M frames using my Tesla p100 gpu, it is too slow and have big difference from the reported speed . Is there any tricks to speed up the training process?
Thirdly, is this code has api so that we can produce the .jspn file as the baselines do?
Looking forward to your reply.