ECE699

Downloading Code for First Time

You may need to set up an ssh certificate for git.

https://docs.github.com/en/authentication/connecting-to-github-with-ssh/generating-a-new-ssh-key-and-adding-it-to-the-ssh-agent

How to run

Make the file using make
Get the output using ./output
If needed, clear the object files (*.o) using make clean

How to run different examples

There are three different examples for you to choose:

Very simple integer BP: example_fc_int_bp_very_simple()
Fully connected (fc) mnist integer BP: example_fc_int_bp_mnist()
Fully connected (fc) mnist integer DFA: example_fc_int_dfa_mnist()
Fully connected (fc) fashion mnist integer DFA: example_fc_int_dfa_fashion_mnist()

You can comment out the current function and call the new function in the main.cpp file

TODO

For mnist dataset training,use classification loss function batchCrossEntropyLoss() and test the accuracy

Debug

Error: output: pktnn_mat.cpp:566: pktnn::pktmat& pktnn::pktmat::matElemDivMat(pktnn::pktmat&, pktnn::pktmat&): Assertion `mat1.dimsEqual(mat2)' failed

Cause: backward() --> computeDeltas() --> matElemDivMat()
Training Accuracy low, for example 9.8% Cause: useDfa(false)

Training Accuracy 100%, maybe not correct either Try to modify the layer setting to match with BP_simple_training

//Modified the fc layers to do BP training
fc1.useDfa(false).initWeightBias().setActv(a).setNextLayer(fc2);         //Use BP for training, set useDfa(false)
fc2.useDfa(false).initWeightBias().setActv(a).setPrevLayer(fc1).setNextLayer(fcLast);
fcLast.useDfa(false).setActv(b).setPrevLayer(fc2);

Reference

Neural Networks and Deep Learning - Michael Nielsen http://neuralnetworksanddeeplearning.com/
Different activation functions and initialization functions https://shahaab-co.com/mag/en-articles/weight-initialization-in-deep-learning/

joshmeiyb/ECE699