From 228d3663f871d0e4bdee468572eb80141cb4fe3f Mon Sep 17 00:00:00 2001
From: Joseph Redmon <pjreddie@gmail.com>
Date: Sat, 15 Feb 2014 00:09:07 +0000
Subject: [PATCH] Extracting features from VOC with temp filters
---
src/tests.c | 457 ++++++++++++++++++++++++++++++++++----------------------
1 files changed, 275 insertions(+), 182 deletions(-)
diff --git a/src/tests.c b/src/tests.c
index 722de1a..47c9787 100644
--- a/src/tests.c
+++ b/src/tests.c
@@ -1,4 +1,5 @@
#include "connected_layer.h"
+//#include "old_conv.h"
#include "convolutional_layer.h"
#include "maxpool_layer.h"
#include "network.h"
@@ -7,15 +8,18 @@
#include "data.h"
#include "matrix.h"
#include "utils.h"
+#include "mini_blas.h"
#include <time.h>
#include <stdlib.h>
#include <stdio.h>
+#define _GNU_SOURCE
+#include <fenv.h>
+
void test_convolve()
{
- image dog = load_image("dog.jpg");
- //show_image_layers(dog, "Dog");
+ image dog = load_image("dog.jpg",300,400);
printf("dog channels %d\n", dog.c);
image kernel = make_random_image(3,3,dog.c);
image edge = make_image(dog.h, dog.w, 1);
@@ -25,39 +29,45 @@
convolve(dog, kernel, 1, 0, edge, 1);
}
end = clock();
- printf("Convolutions: %lf seconds\n", (double)(end-start)/CLOCKS_PER_SEC);
+ printf("Convolutions: %lf seconds\n", (float)(end-start)/CLOCKS_PER_SEC);
show_image_layers(edge, "Test Convolve");
}
+void test_convolve_matrix()
+{
+ image dog = load_image("dog.jpg",300,400);
+ printf("dog channels %d\n", dog.c);
+
+ int size = 11;
+ int stride = 4;
+ int n = 40;
+ float *filters = make_random_image(size, size, dog.c*n).data;
+
+ int mw = ((dog.h-size)/stride+1)*((dog.w-size)/stride+1);
+ int mh = (size*size*dog.c);
+ float *matrix = calloc(mh*mw, sizeof(float));
+
+ image edge = make_image((dog.h-size)/stride+1, (dog.w-size)/stride+1, n);
+
+
+ int i;
+ clock_t start = clock(), end;
+ for(i = 0; i < 1000; ++i){
+ im2col_cpu(dog.data, dog.c, dog.h, dog.w, size, stride, matrix);
+ gemm(0,0,n,mw,mh,1,filters,mh,matrix,mw,1,edge.data,mw);
+ }
+ end = clock();
+ printf("Convolutions: %lf seconds\n", (float)(end-start)/CLOCKS_PER_SEC);
+ show_image_layers(edge, "Test Convolve");
+ cvWaitKey(0);
+}
+
void test_color()
{
- image dog = load_image("test_color.png");
+ image dog = load_image("test_color.png", 300, 400);
show_image_layers(dog, "Test Color");
}
-void test_convolutional_layer()
-{
- srand(0);
- image dog = load_image("dog.jpg");
- int i;
- int n = 3;
- int stride = 1;
- int size = 3;
- convolutional_layer layer = *make_convolutional_layer(dog.h, dog.w, dog.c, n, size, stride, RELU);
- char buff[256];
- for(i = 0; i < n; ++i) {
- sprintf(buff, "Kernel %d", i);
- show_image(layer.kernels[i], buff);
- }
- forward_convolutional_layer(layer, dog.data);
-
- image output = get_convolutional_image(layer);
- maxpool_layer mlayer = *make_maxpool_layer(output.h, output.w, output.c, 2);
- forward_maxpool_layer(mlayer, layer.output);
-
- show_image_layers(get_maxpool_image(mlayer), "Test Maxpool Layer");
-}
-
void verify_convolutional_layer()
{
srand(0);
@@ -65,11 +75,11 @@
int n = 1;
int stride = 1;
int size = 3;
- double eps = .00000001;
+ float eps = .00000001;
image test = make_random_image(5,5, 1);
convolutional_layer layer = *make_convolutional_layer(test.h,test.w,test.c, n, size, stride, RELU);
image out = get_convolutional_image(layer);
- double **jacobian = calloc(test.h*test.w*test.c, sizeof(double));
+ float **jacobian = calloc(test.h*test.w*test.c, sizeof(float));
forward_convolutional_layer(layer, test.data);
image base = copy_image(out);
@@ -83,19 +93,19 @@
jacobian[i] = partial.data;
test.data[i] -= eps;
}
- double **jacobian2 = calloc(out.h*out.w*out.c, sizeof(double));
+ float **jacobian2 = calloc(out.h*out.w*out.c, sizeof(float));
image in_delta = make_image(test.h, test.w, test.c);
image out_delta = get_convolutional_delta(layer);
for(i = 0; i < out.h*out.w*out.c; ++i){
out_delta.data[i] = 1;
- backward_convolutional_layer2(layer, test.data, in_delta.data);
+ backward_convolutional_layer(layer, in_delta.data);
image partial = copy_image(in_delta);
jacobian2[i] = partial.data;
out_delta.data[i] = 0;
}
int j;
- double *j1 = calloc(test.h*test.w*test.c*out.h*out.w*out.c, sizeof(double));
- double *j2 = calloc(test.h*test.w*test.c*out.h*out.w*out.c, sizeof(double));
+ float *j1 = calloc(test.h*test.w*test.c*out.h*out.w*out.c, sizeof(float));
+ float *j2 = calloc(test.h*test.w*test.c*out.h*out.w*out.c, sizeof(float));
for(i = 0; i < test.h*test.w*test.c; ++i){
for(j =0 ; j < out.h*out.w*out.c; ++j){
j1[i*out.h*out.w*out.c + j] = jacobian[i][j];
@@ -105,23 +115,22 @@
}
- image mj1 = double_to_image(test.w*test.h*test.c, out.w*out.h*out.c, 1, j1);
- image mj2 = double_to_image(test.w*test.h*test.c, out.w*out.h*out.c, 1, j2);
+ image mj1 = float_to_image(test.w*test.h*test.c, out.w*out.h*out.c, 1, j1);
+ image mj2 = float_to_image(test.w*test.h*test.c, out.w*out.h*out.c, 1, j2);
printf("%f %f\n", avg_image_layer(mj1,0), avg_image_layer(mj2,0));
show_image(mj1, "forward jacobian");
show_image(mj2, "backward jacobian");
-
}
void test_load()
{
- image dog = load_image("dog.jpg");
+ image dog = load_image("dog.jpg", 300, 400);
show_image(dog, "Test Load");
show_image_layers(dog, "Test Load");
}
void test_upsample()
{
- image dog = load_image("dog.jpg");
+ image dog = load_image("dog.jpg", 300, 400);
int n = 3;
image up = make_image(n*dog.h, n*dog.w, dog.c);
upsample_image(dog, n, up);
@@ -132,13 +141,13 @@
void test_rotate()
{
int i;
- image dog = load_image("dog.jpg");
+ image dog = load_image("dog.jpg",300,400);
clock_t start = clock(), end;
for(i = 0; i < 1001; ++i){
rotate_image(dog);
}
end = clock();
- printf("Rotations: %lf seconds\n", (double)(end-start)/CLOCKS_PER_SEC);
+ printf("Rotations: %lf seconds\n", (float)(end-start)/CLOCKS_PER_SEC);
show_image(dog, "Test Rotate");
image random = make_random_image(3,3,3);
@@ -152,174 +161,166 @@
void test_parser()
{
network net = parse_network_cfg("test_parser.cfg");
- double input[1];
+ float input[1];
int count = 0;
- double avgerr = 0;
- while(1){
- double v = ((double)rand()/RAND_MAX);
- double truth = v*v;
+ float avgerr = 0;
+ while(++count < 100000000){
+ float v = ((float)rand()/RAND_MAX);
+ float truth = v*v;
input[0] = v;
forward_network(net, input);
- double *out = get_network_output(net);
- double *delta = get_network_delta(net);
- double err = pow((out[0]-truth),2.);
+ float *out = get_network_output(net);
+ float *delta = get_network_delta(net);
+ float err = pow((out[0]-truth),2.);
avgerr = .99 * avgerr + .01 * err;
- //if(++count % 100000 == 0) printf("%f\n", avgerr);
- if(++count % 1000000 == 0) printf("%f %f :%f AVG %f \n", truth, out[0], err, avgerr);
+ if(count % 1000000 == 0) printf("%f %f :%f AVG %f \n", truth, out[0], err, avgerr);
delta[0] = truth - out[0];
- learn_network(net, input);
- update_network(net, .001);
+ backward_network(net, input, &truth);
+ update_network(net, .001,0,0);
}
}
void test_data()
{
char *labels[] = {"cat","dog"};
- batch train = random_batch("train_paths.txt", 101,labels, 2);
- show_image(train.images[0], "Test Data Loading");
- show_image(train.images[100], "Test Data Loading");
- show_image(train.images[10], "Test Data Loading");
- free_batch(train);
+ data train = load_data_image_pathfile_random("train_paths.txt", 101,labels, 2, 300, 400);
+ free_data(train);
}
void test_full()
{
network net = parse_network_cfg("full.cfg");
- srand(0);
- int i = 0;
+ srand(2222222);
+ int i = 800;
char *labels[] = {"cat","dog"};
+ float lr = .00001;
+ float momentum = .9;
+ float decay = 0.01;
while(i++ < 1000 || 1){
- batch train = random_batch("train_paths.txt", 1000, labels, 2);
- train_network_batch(net, train);
- free_batch(train);
- printf("Round %d\n", i);
+ visualize_network(net);
+ cvWaitKey(100);
+ data train = load_data_image_pathfile_random("train_paths.txt", 1000, labels, 2, 256, 256);
+ image im = float_to_image(256, 256, 3,train.X.vals[0]);
+ show_image(im, "input");
+ cvWaitKey(100);
+ //scale_data_rows(train, 1./255.);
+ normalize_data_rows(train);
+ clock_t start = clock(), end;
+ float loss = train_network_sgd(net, train, 100, lr, momentum, decay);
+ end = clock();
+ printf("%d: %f, Time: %lf seconds, LR: %f, Momentum: %f, Decay: %f\n", i, loss, (float)(end-start)/CLOCKS_PER_SEC, lr, momentum, decay);
+ free_data(train);
+ if(i%100==0){
+ char buff[256];
+ sprintf(buff, "backup_%d.cfg", i);
+ //save_network(net, buff);
+ }
+ //lr *= .99;
}
}
-double error_network(network net, matrix m, double *truth)
-{
- int i;
- int correct = 0;
- for(i = 0; i < m.rows; ++i){
- forward_network(net, m.vals[i]);
- double *out = get_network_output(net);
- double err = truth[i] - out[0];
- if(fabs(err) < .5) ++correct;
- }
- return (double)correct/m.rows;
-}
-
-double **one_hot(double *a, int n, int k)
-{
- int i;
- double **t = calloc(n, sizeof(double*));
- for(i = 0; i < n; ++i){
- t[i] = calloc(k, sizeof(double));
- int index = (int)a[i];
- t[i][index] = 1;
- }
- return t;
-}
-
void test_nist()
{
+ srand(444444);
+ srand(888888);
network net = parse_network_cfg("nist.cfg");
- matrix m = csv_to_matrix("images/nist_train.csv");
- matrix ho = hold_out_matrix(&m, 3000);
- double *truth_1d = pop_column(&m, 0);
- double **truth = one_hot(truth_1d, m.rows, 10);
- double *ho_truth_1d = pop_column(&ho, 0);
- double **ho_truth = one_hot(ho_truth_1d, ho.rows, 10);
- int i,j;
- clock_t start = clock(), end;
+ data train = load_categorical_data_csv("mnist/mnist_train.csv", 0, 10);
+ data test = load_categorical_data_csv("mnist/mnist_test.csv",0,10);
+ normalize_data_rows(train);
+ normalize_data_rows(test);
+ //randomize_data(train);
int count = 0;
- double lr = .0001;
- while(++count <= 3000000){
- //lr *= .99;
- int index = 0;
- int correct = 0;
- for(i = 0; i < 1000; ++i){
- index = rand()%m.rows;
- normalize_array(m.vals[index], 28*28);
- forward_network(net, m.vals[index]);
- double *out = get_network_output(net);
- double *delta = get_network_delta(net);
- int max_i = 0;
- double max = out[0];
- for(j = 0; j < 10; ++j){
- delta[j] = truth[index][j]-out[j];
- if(out[j] > max){
- max = out[j];
- max_i = j;
- }
- }
- if(truth[index][max_i]) ++correct;
- learn_network(net, m.vals[index]);
- update_network(net, lr);
+ float lr = .0005;
+ float momentum = .9;
+ float decay = 0.001;
+ clock_t start = clock(), end;
+ while(++count <= 100){
+ //visualize_network(net);
+ float loss = train_network_sgd(net, train, 1000, lr, momentum, decay);
+ printf("%5d Training Loss: %lf, Params: %f %f %f, ",count*100, loss, lr, momentum, decay);
+ end = clock();
+ printf("Time: %lf seconds\n", (float)(end-start)/CLOCKS_PER_SEC);
+ start=end;
+ //cvWaitKey(100);
+ //lr /= 2;
+ if(count%5 == 0){
+ float train_acc = network_accuracy(net, train);
+ fprintf(stderr, "\nTRAIN: %f\n", train_acc);
+ float test_acc = network_accuracy(net, test);
+ fprintf(stderr, "TEST: %f\n\n", test_acc);
+ printf("%d, %f, %f\n", count, train_acc, test_acc);
+ //lr *= .5;
}
- print_network(net);
- image input = double_to_image(28,28,1, m.vals[index]);
- show_image(input, "Input");
- image o = get_network_image(net);
- show_image_collapsed(o, "Output");
- visualize_network(net);
- cvWaitKey(100);
- //double test_acc = error_network(net, m, truth);
- //double valid_acc = error_network(net, ho, ho_truth);
- //printf("%f, %f\n", test_acc, valid_acc);
- fprintf(stderr, "%5d: %f %f\n",count, (double)correct/1000, lr);
- //if(valid_acc > .70) break;
}
- end = clock();
- printf("Neural Net Learning: %lf seconds\n", (double)(end-start)/CLOCKS_PER_SEC);
}
-void test_kernel_update()
+void test_ensemble()
{
- srand(0);
- double delta[] = {.1};
- double input[] = {.3, .5, .3, .5, .5, .5, .5, .0, .5};
- double kernel[] = {1,2,3,4,5,6,7,8,9};
- convolutional_layer layer = *make_convolutional_layer(3, 3, 1, 1, 3, 1, IDENTITY);
- layer.kernels[0].data = kernel;
- layer.delta = delta;
- learn_convolutional_layer(layer, input);
- print_image(layer.kernels[0]);
- print_image(get_convolutional_delta(layer));
- print_image(layer.kernel_updates[0]);
-
+ int i;
+ srand(888888);
+ data d = load_categorical_data_csv("mnist/mnist_train.csv", 0, 10);
+ normalize_data_rows(d);
+ data test = load_categorical_data_csv("mnist/mnist_test.csv", 0,10);
+ normalize_data_rows(test);
+ data train = d;
+ // data *split = split_data(d, 1, 10);
+ // data train = split[0];
+ // data test = split[1];
+ matrix prediction = make_matrix(test.y.rows, test.y.cols);
+ int n = 30;
+ for(i = 0; i < n; ++i){
+ int count = 0;
+ float lr = .0005;
+ float momentum = .9;
+ float decay = .01;
+ network net = parse_network_cfg("nist.cfg");
+ while(++count <= 15){
+ float acc = train_network_sgd(net, train, train.X.rows, lr, momentum, decay);
+ printf("Training Accuracy: %lf Learning Rate: %f Momentum: %f Decay: %f\n", acc, lr, momentum, decay );
+ lr /= 2;
+ }
+ matrix partial = network_predict_data(net, test);
+ float acc = matrix_accuracy(test.y, partial);
+ printf("Model Accuracy: %lf\n", acc);
+ matrix_add_matrix(partial, prediction);
+ acc = matrix_accuracy(test.y, prediction);
+ printf("Current Ensemble Accuracy: %lf\n", acc);
+ free_matrix(partial);
+ }
+ float acc = matrix_accuracy(test.y, prediction);
+ printf("Full Ensemble Accuracy: %lf\n", acc);
}
void test_random_classify()
{
network net = parse_network_cfg("connected.cfg");
matrix m = csv_to_matrix("train.csv");
- matrix ho = hold_out_matrix(&m, 2500);
- double *truth = pop_column(&m, 0);
- double *ho_truth = pop_column(&ho, 0);
+ //matrix ho = hold_out_matrix(&m, 2500);
+ float *truth = pop_column(&m, 0);
+ //float *ho_truth = pop_column(&ho, 0);
int i;
clock_t start = clock(), end;
int count = 0;
while(++count <= 300){
for(i = 0; i < m.rows; ++i){
int index = rand()%m.rows;
- //image p = double_to_image(1690,1,1,m.vals[index]);
+ //image p = float_to_image(1690,1,1,m.vals[index]);
//normalize_image(p);
forward_network(net, m.vals[index]);
- double *out = get_network_output(net);
- double *delta = get_network_delta(net);
+ float *out = get_network_output(net);
+ float *delta = get_network_delta(net);
//printf("%f\n", out[0]);
delta[0] = truth[index] - out[0];
- // printf("%f\n", delta[0]);
+ // printf("%f\n", delta[0]);
//printf("%f %f\n", truth[index], out[0]);
- learn_network(net, m.vals[index]);
- update_network(net, .00001);
+ //backward_network(net, m.vals[index], );
+ update_network(net, .00001, 0,0);
}
- double test_acc = error_network(net, m, truth);
- double valid_acc = error_network(net, ho, ho_truth);
- printf("%f, %f\n", test_acc, valid_acc);
- fprintf(stderr, "%5d: %f Valid: %f\n",count, test_acc, valid_acc);
+ //float test_acc = error_network(net, m, truth);
+ //float valid_acc = error_network(net, ho, ho_truth);
+ //printf("%f, %f\n", test_acc, valid_acc);
+ //fprintf(stderr, "%5d: %f Valid: %f\n",count, test_acc, valid_acc);
//if(valid_acc > .70) break;
}
end = clock();
@@ -328,42 +329,134 @@
truth = pop_column(&test, 0);
for(i = 0; i < test.rows; ++i){
forward_network(net, test.vals[i]);
- double *out = get_network_output(net);
+ float *out = get_network_output(net);
if(fabs(out[0]) < .5) fprintf(fp, "0\n");
else fprintf(fp, "1\n");
}
fclose(fp);
- printf("Neural Net Learning: %lf seconds\n", (double)(end-start)/CLOCKS_PER_SEC);
+ printf("Neural Net Learning: %lf seconds\n", (float)(end-start)/CLOCKS_PER_SEC);
}
-void test_random_preprocess()
+void test_split()
{
- FILE *file = fopen("train.csv", "w");
- char *labels[] = {"cat","dog"};
- int i,j,k;
- srand(0);
- network net = parse_network_cfg("convolutional.cfg");
- for(i = 0; i < 100; ++i){
- printf("%d\n", i);
- batch part = get_batch("train_paths.txt", i, 100, labels, 2);
- for(j = 0; j < part.n; ++j){
- forward_network(net, part.images[j].data);
- double *out = get_network_output(net);
- fprintf(file, "%f", part.truth[j][0]);
- for(k = 0; k < get_network_output_size(net); ++k){
- fprintf(file, ",%f", out[k]);
- }
- fprintf(file, "\n");
+ data train = load_categorical_data_csv("mnist/mnist_train.csv", 0, 10);
+ data *split = split_data(train, 0, 13);
+ printf("%d, %d, %d\n", train.X.rows, split[0].X.rows, split[1].X.rows);
+}
+
+void test_im2row()
+{
+ int h = 20;
+ int w = 20;
+ int c = 3;
+ int stride = 1;
+ int size = 11;
+ image test = make_random_image(h,w,c);
+ int mc = 1;
+ int mw = ((h-size)/stride+1)*((w-size)/stride+1);
+ int mh = (size*size*c);
+ int msize = mc*mw*mh;
+ float *matrix = calloc(msize, sizeof(float));
+ int i;
+ for(i = 0; i < 1000; ++i){
+ im2col_cpu(test.data, c, h, w, size, stride, matrix);
+ //image render = float_to_image(mh, mw, mc, matrix);
+ }
+}
+
+void train_VOC()
+{
+ network net = parse_network_cfg("cfg/voc_backup_sig_20.cfg");
+ srand(2222222);
+ int i = 20;
+ char *labels[] = {"aeroplane","bicycle","bird","boat","bottle","bus","car","cat","chair","cow","diningtable","dog","horse","motorbike","person","pottedplant","sheep","sofa","train","tvmonitor"};
+ float lr = .00001;
+ float momentum = .9;
+ float decay = 0.01;
+ while(i++ < 1000 || 1){
+ data train = load_data_image_pathfile_random("images/VOC2012/train_paths.txt", 1000, labels, 20, 300, 400);
+
+ image im = float_to_image(300, 400, 3,train.X.vals[0]);
+ show_image(im, "input");
+ visualize_network(net);
+ cvWaitKey(100);
+
+ normalize_data_rows(train);
+ clock_t start = clock(), end;
+ float loss = train_network_sgd(net, train, 1000, lr, momentum, decay);
+ end = clock();
+ printf("%d: %f, Time: %lf seconds, LR: %f, Momentum: %f, Decay: %f\n", i, loss, (float)(end-start)/CLOCKS_PER_SEC, lr, momentum, decay);
+ free_data(train);
+ if(i%10==0){
+ char buff[256];
+ sprintf(buff, "cfg/voc_backup_sig_%d.cfg", i);
+ save_network(net, buff);
}
- free_batch(part);
+ //lr *= .99;
+ }
+}
+
+void features_VOC()
+{
+ int i,j;
+ network net = parse_network_cfg("cfg/voc_features.cfg");
+ char *path_file = "images/VOC2012/all_paths.txt";
+ char *out_dir = "voc_features/";
+ list *paths = get_paths(path_file);
+ node *n = paths->front;
+ while(n){
+ char *path = (char *)n->val;
+ char buff[1024];
+ sprintf(buff, "%s%s.txt",out_dir, path);
+ FILE *fp = fopen(buff, "w");
+ if(fp == 0) file_error(buff);
+
+ IplImage* src = 0;
+ if( (src = cvLoadImage(path,-1)) == 0 )
+ {
+ printf("Cannot load file image %s\n", path);
+ exit(0);
+ }
+
+ for(i = 0; i < 10; ++i){
+ int w = 1024 - 90*i; //PICKED WITH CAREFUL CROSS-VALIDATION!!!!
+ int h = (int)((double)w/src->width * src->height);
+ IplImage *sized = cvCreateImage(cvSize(w,h), src->depth, src->nChannels);
+ cvResize(src, sized, CV_INTER_LINEAR);
+ image im = ipl_to_image(sized);
+ reset_network_size(net, im.h, im.w, im.c);
+ forward_network(net, im.data);
+ free_image(im);
+ image out = get_network_image_layer(net, 5);
+ fprintf(fp, "%d, %d, %d\n",out.c, out.h, out.w);
+ for(j = 0; j < out.c*out.h*out.w; ++j){
+ if(j != 0)fprintf(fp, ",");
+ fprintf(fp, "%g", out.data[j]);
+ }
+ fprintf(fp, "\n");
+ out.c = 1;
+ show_image(out, "output");
+ cvWaitKey(10);
+ cvReleaseImage(&sized);
+ }
+ fclose(fp);
+ n = n->next;
}
}
int main()
{
- //test_kernel_update();
+ //feenableexcept(FE_DIVBYZERO | FE_INVALID | FE_OVERFLOW);
+
+ //test_blas();
+ //test_convolve_matrix();
+ // test_im2row();
+ //test_split();
+ //test_ensemble();
//test_nist();
- test_full();
+ //test_full();
+ //train_VOC();
+ features_VOC();
//test_random_preprocess();
//test_random_classify();
//test_parser();
@@ -377,6 +470,6 @@
//test_convolutional_layer();
//verify_convolutional_layer();
//test_color();
- cvWaitKey(0);
+ //cvWaitKey(0);
return 0;
}
--
Gitblit v1.10.0