From 787d5345609459f21fd65d2d8b4fcd55201e21a1 Mon Sep 17 00:00:00 2001
From: Joseph Redmon <pjreddie@gmail.com>
Date: Mon, 13 Oct 2014 07:31:10 +0000
Subject: [PATCH] Convolutional working on GPU
---
src/cnn.c | 191 ++++++++++++++++++++++++++++++++++++++++++-----
1 files changed, 170 insertions(+), 21 deletions(-)
diff --git a/src/cnn.c b/src/cnn.c
index f866194..472aa03 100644
--- a/src/cnn.c
+++ b/src/cnn.c
@@ -32,6 +32,113 @@
show_image_layers(edge, "Test Convolve");
}
+#ifdef GPU
+
+void test_convolutional_layer()
+{
+ int i;
+ image dog = load_image("data/dog.jpg",224,224);
+ network net = parse_network_cfg("cfg/convolutional.cfg");
+// data test = load_cifar10_data("data/cifar10/test_batch.bin");
+// float *X = calloc(net.batch*test.X.cols, sizeof(float));
+// float *y = calloc(net.batch*test.y.cols, sizeof(float));
+ int in_size = get_network_input_size(net)*net.batch;
+ int del_size = get_network_output_size_layer(net, 0)*net.batch;
+ int size = get_network_output_size(net)*net.batch;
+ float *X = calloc(in_size, sizeof(float));
+ float *y = calloc(size, sizeof(float));
+ for(i = 0; i < in_size; ++i){
+ X[i] = dog.data[i%get_network_input_size(net)];
+ }
+// get_batch(test, net.batch, X, y);
+ clock_t start, end;
+ cl_mem input_cl = cl_make_array(X, in_size);
+ cl_mem truth_cl = cl_make_array(y, size);
+
+ forward_network_gpu(net, input_cl, truth_cl, 1);
+ start = clock();
+ forward_network_gpu(net, input_cl, truth_cl, 1);
+ end = clock();
+ float gpu_sec = (float)(end-start)/CLOCKS_PER_SEC;
+ printf("forward gpu: %f sec\n", gpu_sec);
+ start = clock();
+ backward_network_gpu(net, input_cl);
+ end = clock();
+ gpu_sec = (float)(end-start)/CLOCKS_PER_SEC;
+ printf("backward gpu: %f sec\n", gpu_sec);
+ //float gpu_cost = get_network_cost(net);
+ float *gpu_out = calloc(size, sizeof(float));
+ memcpy(gpu_out, get_network_output(net), size*sizeof(float));
+
+ float *gpu_del = calloc(del_size, sizeof(float));
+ memcpy(gpu_del, get_network_delta_layer(net, 0), del_size*sizeof(float));
+
+/*
+ start = clock();
+ forward_network(net, X, y, 1);
+ backward_network(net, X);
+ float cpu_cost = get_network_cost(net);
+ end = clock();
+ float cpu_sec = (float)(end-start)/CLOCKS_PER_SEC;
+ float *cpu_out = calloc(size, sizeof(float));
+ memcpy(cpu_out, get_network_output(net), size*sizeof(float));
+ float *cpu_del = calloc(del_size, sizeof(float));
+ memcpy(cpu_del, get_network_delta_layer(net, 0), del_size*sizeof(float));
+
+ float sum = 0;
+ float del_sum = 0;
+ for(i = 0; i < size; ++i) sum += pow(gpu_out[i] - cpu_out[i], 2);
+ for(i = 0; i < del_size; ++i) {
+ //printf("%f %f\n", cpu_del[i], gpu_del[i]);
+ del_sum += pow(cpu_del[i] - gpu_del[i], 2);
+ }
+ printf("GPU cost: %f, CPU cost: %f\n", gpu_cost, cpu_cost);
+ printf("gpu: %f sec, cpu: %f sec, diff: %f, delta diff: %f, size: %d\n", gpu_sec, cpu_sec, sum, del_sum, size);
+ */
+}
+
+void test_col2im()
+{
+ float col[] = {1,2,1,2,
+ 1,2,1,2,
+ 1,2,1,2,
+ 1,2,1,2,
+ 1,2,1,2,
+ 1,2,1,2,
+ 1,2,1,2,
+ 1,2,1,2,
+ 1,2,1,2};
+ float im[16] = {0};
+ int batch = 1;
+ int channels = 1;
+ int height=4;
+ int width=4;
+ int ksize = 3;
+ int stride = 1;
+ int pad = 0;
+ col2im_gpu(col, batch,
+ channels, height, width,
+ ksize, stride, pad, im);
+ int i;
+ for(i = 0; i < 16; ++i)printf("%f,", im[i]);
+ printf("\n");
+ /*
+ float data_im[] = {
+ 1,2,3,4,
+ 5,6,7,8,
+ 9,10,11,12
+ };
+ float data_col[18] = {0};
+ im2col_cpu(data_im, batch,
+ channels, height, width,
+ ksize, stride, pad, data_col) ;
+ for(i = 0; i < 18; ++i)printf("%f,", data_col[i]);
+ printf("\n");
+ */
+}
+
+#endif
+
void test_convolve_matrix()
{
image dog = load_image("dog.jpg",300,400);
@@ -229,7 +336,7 @@
normalize_data_rows(test);
for(j = 0; j < test.X.rows; ++j){
float *x = test.X.vals[j];
- forward_network(net, x, 0);
+ forward_network(net, x, 0, 0);
int class = get_predicted_class_network(net);
fprintf(fp, "%d\n", class);
}
@@ -240,9 +347,21 @@
void test_cifar10()
{
- srand(222222);
+ network net = parse_network_cfg("cfg/cifar10_part5.cfg");
+ data test = load_cifar10_data("data/cifar10/test_batch.bin");
+ clock_t start = clock(), end;
+ float test_acc = network_accuracy(net, test);
+ end = clock();
+ printf("%f in %f Sec\n", test_acc, (float)(end-start)/CLOCKS_PER_SEC);
+ visualize_network(net);
+ cvWaitKey(0);
+}
+
+void train_cifar10()
+{
+ srand(555555);
network net = parse_network_cfg("cfg/cifar10.cfg");
- //data test = load_cifar10_data("data/cifar10/test_batch.bin");
+ data test = load_cifar10_data("data/cifar10/test_batch.bin");
int count = 0;
int iters = 10000/net.batch;
data train = load_all_cifar10();
@@ -250,12 +369,20 @@
clock_t start = clock(), end;
float loss = train_network_sgd(net, train, iters);
end = clock();
- //visualize_network(net);
- //cvWaitKey(1000);
+ visualize_network(net);
+ cvWaitKey(5000);
//float test_acc = network_accuracy(net, test);
//printf("%d: Loss: %f, Test Acc: %f, Time: %lf seconds, LR: %f, Momentum: %f, Decay: %f\n", count, loss, test_acc,(float)(end-start)/CLOCKS_PER_SEC, net.learning_rate, net.momentum, net.decay);
- printf("%d: Loss: %f, Time: %lf seconds, LR: %f, Momentum: %f, Decay: %f\n", count, loss, (float)(end-start)/CLOCKS_PER_SEC, net.learning_rate, net.momentum, net.decay);
+ if(count%10 == 0){
+ float test_acc = network_accuracy(net, test);
+ printf("%d: Loss: %f, Test Acc: %f, Time: %lf seconds, LR: %f, Momentum: %f, Decay: %f\n", count, loss, test_acc,(float)(end-start)/CLOCKS_PER_SEC, net.learning_rate, net.momentum, net.decay);
+ char buff[256];
+ sprintf(buff, "/home/pjreddie/cifar/cifar2_%d.cfg", count);
+ save_network(net, buff);
+ }else{
+ printf("%d: Loss: %f, Time: %lf seconds, LR: %f, Momentum: %f, Decay: %f\n", count, loss, (float)(end-start)/CLOCKS_PER_SEC, net.learning_rate, net.momentum, net.decay);
+ }
}
free_data(train);
}
@@ -281,10 +408,10 @@
void test_nist_single()
{
srand(222222);
- network net = parse_network_cfg("cfg/nist.cfg");
+ network net = parse_network_cfg("cfg/nist_single.cfg");
data train = load_categorical_data_csv("data/mnist/mnist_tiny.csv", 0, 10);
normalize_data_rows(train);
- float loss = train_network_sgd(net, train, 5);
+ float loss = train_network_sgd(net, train, 1);
printf("Loss: %f, LR: %f, Momentum: %f, Decay: %f\n", loss, net.learning_rate, net.momentum, net.decay);
}
@@ -292,25 +419,43 @@
void test_nist()
{
srand(222222);
+ network net = parse_network_cfg("cfg/nist_final.cfg");
+ data test = load_categorical_data_csv("data/mnist/mnist_test.csv",0,10);
+ translate_data_rows(test, -144);
+ clock_t start = clock(), end;
+ float test_acc = network_accuracy_multi(net, test,16);
+ end = clock();
+ printf("Accuracy: %f, Time: %lf seconds\n", test_acc,(float)(end-start)/CLOCKS_PER_SEC);
+}
+
+void train_nist()
+{
+ srand(222222);
network net = parse_network_cfg("cfg/nist.cfg");
data train = load_categorical_data_csv("data/mnist/mnist_train.csv", 0, 10);
data test = load_categorical_data_csv("data/mnist/mnist_test.csv",0,10);
- translate_data_rows(train, -144);
- scale_data_rows(train, 1./128);
- translate_data_rows(test, -144);
- scale_data_rows(test, 1./128);
+ translate_data_rows(train, -144);
+ //scale_data_rows(train, 1./128);
+ translate_data_rows(test, -144);
+ //scale_data_rows(test, 1./128);
//randomize_data(train);
int count = 0;
//clock_t start = clock(), end;
int iters = 10000/net.batch;
- while(++count <= 100){
+ while(++count <= 2000){
clock_t start = clock(), end;
float loss = train_network_sgd(net, train, iters);
end = clock();
float test_acc = network_accuracy(net, test);
//float test_acc = 0;
printf("%d: Loss: %f, Test Acc: %f, Time: %lf seconds, LR: %f, Momentum: %f, Decay: %f\n", count, loss, test_acc,(float)(end-start)/CLOCKS_PER_SEC, net.learning_rate, net.momentum, net.decay);
- //save_network(net, "cfg/nist_basic_trained.cfg");
+ /*printf("%f %f %f %f %f\n", mean_array(get_network_output_layer(net,0), 100),
+ mean_array(get_network_output_layer(net,1), 100),
+ mean_array(get_network_output_layer(net,2), 100),
+ mean_array(get_network_output_layer(net,3), 100),
+ mean_array(get_network_output_layer(net,4), 100));
+ */
+ //save_network(net, "cfg/nist_final2.cfg");
//printf("%5d Training Loss: %lf, Params: %f %f %f, ",count*1000, loss, lr, momentum, decay);
//end = clock();
@@ -373,7 +518,7 @@
int index = rand()%m.rows;
//image p = float_to_image(1690,1,1,m.vals[index]);
//normalize_image(p);
- forward_network(net, m.vals[index], 1);
+ forward_network(net, m.vals[index], 0, 1);
float *out = get_network_output(net);
float *delta = get_network_delta(net);
//printf("%f\n", out[0]);
@@ -394,7 +539,7 @@
matrix test = csv_to_matrix("test.csv");
truth = pop_column(&test, 0);
for(i = 0; i < test.rows; ++i){
- forward_network(net, test.vals[i], 0);
+ forward_network(net, test.vals[i],0, 0);
float *out = get_network_output(net);
if(fabs(out[0]) < .5) fprintf(fp, "0\n");
else fprintf(fp, "1\n");
@@ -494,7 +639,7 @@
//normalize_array(im.data, im.h*im.w*im.c);
translate_image(im, -144);
resize_network(net, im.h, im.w, im.c);
- forward_network(net, im.data, 0);
+ forward_network(net, im.data, 0, 0);
image out = get_network_image(net);
free_image(im);
cvReleaseImage(&sized);
@@ -546,7 +691,7 @@
resize_network(net, im.h, im.w, im.c);
//scale_image(im, 1./255);
translate_image(im, -144);
- forward_network(net, im.data, 0);
+ forward_network(net, im.data, 0, 0);
image out = get_network_image(net);
int dh = (im.h - h)/(out.h-1);
@@ -608,7 +753,7 @@
image im = load_image(image_path, 0, 0);
printf("Processing %dx%d image\n", im.h, im.w);
resize_network(net, im.h, im.w, im.c);
- forward_network(net, im.data, 0);
+ forward_network(net, im.data, 0, 0);
image out = get_network_image(net);
int dh = (im.h - h)/h;
@@ -641,7 +786,7 @@
image im = load_image("data/cat.png", 0, 0);
printf("Processing %dx%d image\n", im.h, im.w);
resize_network(net, im.h, im.w, im.c);
- forward_network(net, im.data, 0);
+ forward_network(net, im.data, 0, 0);
visualize_network(net);
cvWaitKey(0);
@@ -770,8 +915,12 @@
//test_split();
//test_ensemble();
//test_nist_single();
- test_nist();
+ //test_nist();
+ //train_nist();
+ test_convolutional_layer();
+ //test_col2im();
//test_cifar10();
+ //train_cifar10();
//test_vince();
//test_full();
//tune_VOC();
--
Gitblit v1.10.0