How to tune Honglak Lee's CRBM Sparsity learning rate ?

I'm implementing the sparsity technique used by Honglak Lee for CRBM. There are two parameters:

• pbias: The target sparsity
• pbias_lambda: The learning rate for sparsity

I have some issues for pbias_lambda. On my case (working on MNIST), I have to set it very high (around 1000) in order to reach the target sparsity sometime. When running Honglak Lee's code, the target sparsity is reached very quickly and the pbias_lambda is only 5.

Is that normal behaviour ? How should I go about tuning this ?

Thanks