In Caffe, Does TanHLayer Backward compute error gradient w.r.t. the sigmoid inputs?

Question

I'm new to Caffe.

TanHLayer's Doxygen documentation says that,

Backward_cpu() Computes the error gradient w.r.t. the sigmoid inputs.

http://caffe.berkeleyvision.org/doxygen/classcaffe_1_1TanHLayer.html

Why does it takes sigmoid inputs, not TanH? It confuses me. Please give me some hints.

I asked on Caffe Users Group, but got no answers. https://groups.google.com/forum/#!topic/caffe-users/oxRDDOy_xN8

Thanks.