keras_spmv_rt.py

class keras_spmv_rt.KerasSpmvRT(keras_model, batch_sz, xclbin_opts)

runtime class to use SPMV in the keras example

It will send all the weights matrices loaded from the model to kernel for future usage

Parameters

• keras_model (model) – pre-trained model

• xclbin_opts (dictionary) – information read from config_info.dat used to build the xclbin

• batch_sz (shape) – output prediction matrix size

format_for_sparse_fpga(weight_mat, shape)

return padded sizes, row index, col index and non-zero elements array of given sparse matrix to fit the format of spmv engine

Parameters

• weight_mat (ndarray) – transpose of the weight sparse matrix

• shape (shape) – shape of the sparse matrix

Returns

• int – padded row size

• int – padded col size

• int – padded number of non-zero elements

• array – row indices

• array – col indices

• array – non-zero elements

predict(inp)

prepare input matrix for the engine, send all the matrices and instructions to kernel and get the result prediction matrix

Parameters

inp (array) – input matrix

Returns

result prediction matrix

Return type

array