Truncated BPTT

[Fernando Quivira]

Hi

I'm reading a bit more on backpropagation throught time (BPTT) and its variations, and I came across the following blog post: http://r2rt.com/styles-of-truncated-backpropagation.html

If I understand correctly, when we use unchain_backward(),  we are using the following paradigm, right? (graphic from the blog post)

The unchaining would occur right when we compute the "final state." Then, after unchaining, we use the previous state and continue computing with the RNN.

The blog post mentions this other approach: 

This is another variation on BPTT (different parameters k1 and k2 as explained in http://www.cs.utoronto.ca/~ilya/pubs/ilya_sutskever_phd_thesis.pdf)

My question is: is there a clean way of implementing these BPTT variations in Chainer? 

Thanks!

[Fernando Quivira]

Does anybody have an idea how to do this? or am I incorrect in my understanding of BPTT?

Thanks