bonejay

Yeah it would be a great help. Kind of tiresome that none of the document transformers (TILT, StructuralLM, layoutlm,StrucTexT) have release their pretraining code, although it is the most important...

Here is my image_lib folder where there is only the .dll file. 

I changed the cmake file to your suggestion and this came out. Only more conflicts with dynamic libraries. 

So during inference in order to generate rk we attend r1 til r(k-1). Is that really needed. Can't we simply just attend r(k-1) and mask everything else for the cross...