what about the Initial state of the network?

[nightinwhite]

we used the weight of your network's weight in our network as initial state and modify the fc layer to recognize 4225 chinese characters,but its accuracy is only around 55%.Should I try to initialize the weight to a random number ?or Is there any tricks to train the model?Thanks for any help!