cpp-taskflow/benchmark/parallel_dnn/seq.cpp at cpp14 · walker2010cs/cpp-taskflow · GitHub

History

90 lines (72 loc) · 2.02 KB

Raw

#include "dnn.hpp"

void run_sequential(MNIST& D, unsigned num_threads) {

const auto iter_num = D.images.rows()/D.batch_size;

for(auto e=0u; e<D.epoch; e++) {

for(auto it=0u; it<iter_num; it++) {

// Foward propagation

for(size_t i=0; i<D.acts.size(); i++) {

if(i == 0){

D.forward(i, D.images.middleRows(D.beg_row, D.batch_size));

}

else {

D.forward(i, D.Ys[i-1]);

}

// Calculate loss

D.loss(D.labels);

// Backward propagation

for(int i=D.acts.size()-1; i>=0; i--) {

if(i > 0) {

//D.backward(i, D.Ys[i-1].transpose());

D.backward(i, D.Ys[i-1]);

}

else {

//D.backward(i, D.images.middleRows(D.beg_row, D.batch_size).transpose());

D.backward(i, D.images.middleRows(D.beg_row, D.batch_size));

}

// Update parameters

for(int i=D.acts.size()-1; i>=0; i--) {

D.update(i);

}

// Get next batch

D.beg_row += D.batch_size;

if(D.beg_row >= D.images.rows()) {

D.beg_row = 0;

}

} // End of iterations

D.validate();

// Shuffle input

D.shuffle(D.images, D.labels, D.images.rows());

} // End of epoch

}

void run_sequential(unsigned num_epochs, unsigned num_threads) {

MNIST_DNN D;

init_dnn(D, rand_rate());

for(auto e=0u; e<100; e++) {

for(auto it=0u; it<NUM_ITERATIONS; it++) {

// Foward propagation

forward_task(D, IMAGES, LABELS);

// Calculate loss

D.loss(LABELS);

// Backward propagation

for(int i=D.acts.size()-1; i>=0; i--) {

backward_task(D, i, IMAGES);

}

// Update parameters

for(int i=D.acts.size()-1; i>=0; i--) {

D.update(i);

}

// Get next batch

D.beg_row += D.batch_size;

if(D.beg_row >= IMAGES.rows()) {

D.beg_row = 0;

}

} // End of iterations

// Shuffle input

shuffle(IMAGES, LABELS);

D.validate(TEST_IMAGES, TEST_LABELS);

} // End of epoch

}

cpp-taskflow/benchmark/parallel_dnn/seq.cpp at cpp14 · walker2010cs/cpp-taskflow · GitHub

Expand file tree

Breadcrumbs

seq.cpp

Latest commit

History

Breadcrumbs

seq.cpp

File metadata and controls