How to fix the parameters in lstm layer?

[吴一鸣]

I trained a network contain the lstm layer. And I want to fix the parameters' value in lstm layer for finetuning. However, I found that parameter in lstm layer could not fixed  using

param{lr_mult: 0}

In the attachment, i set

loss weight: 0

and the loss is actually 0, but the Test loss is change every test step (the test data is same each time). Then I found lstm parameters is need to be backward computation as the second picture shown. But the other layers is not.