How to process variable timesteps input with LstmLayer?

[diany...@gmail.com]

Recent Caffe have merged the recurrent branch by jeffdonahue.

I look at the code in RecurrentLayer which LstmLayer inherit from, the RecurrentLayer::LayerSetUp makes the input timesteps fixed.

So how to process with variable timesteps input?Does it need to call LayerSetUp in every Reshape function?

 

And I am planning to implement a recurrent in convolutional layer what there is connection between adjacent feature map cells.

Shoud I do this by inheriting both ConvolutionLayer(BaseConvolutionLayer) and RecurrentLayer?