This is an unfinished test model of CNN, based on cnn.h5 Keras pretrained model https://github.com/EN10/KerasMNIST/blob/4ef71d6332e12376df93c96b39cf01ffbe48d02a/cnn.h5 .
This model serves as a testing ground for making Keras models in SystemC and contains the following layers:
tb_driver.h contains image of number 6 in form of grayscale values matix and arbitrary kernel. All interfaces are 1d vectors, so on inputs/output ports of some modules 1d vectors transform into 2d matrices and vice versa.
Convolution is realised with element-wise, for-loop multiplication of kernel and image instead of multiplication of matricies using Toepliz matricies. This is the issue to be solved.
The model does not contain kernel and bias values from the original model. upd: description is outdated