training loss is nan

[heichang12138]

i noticed that lib/networks/network.py have been modified. use the previous version is fine, but use modified version will lead to gradient explosion. by the way, the dataset i use is caltech pedestrian detection. apologize for my poor English:)