From cc06817efa24f20811ef6b32143c6700a91c5f2a Mon Sep 17 00:00:00 2001
From: Joseph Redmon <pjreddie@gmail.com>
Date: Fri, 11 Apr 2014 08:00:27 +0000
Subject: [PATCH] Attempt at visualizing ImageNet Features
---
src/tests.c | 374 +++++++++++++++++++++++++++++++++++++++++++++++------
1 files changed, 331 insertions(+), 43 deletions(-)
diff --git a/src/tests.c b/src/tests.c
index 09ec7b2..5d9136d 100644
--- a/src/tests.c
+++ b/src/tests.c
@@ -77,7 +77,7 @@
int size = 3;
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);
+ convolutional_layer layer = *make_convolutional_layer(1,test.h,test.w,test.c, n, size, stride, RELU);
image out = get_convolutional_image(layer);
float **jacobian = calloc(test.h*test.w*test.c, sizeof(float));
@@ -188,70 +188,151 @@
free_data(train);
}
-void test_full()
+void train_full()
{
- network net = parse_network_cfg("full.cfg");
+ network net = parse_network_cfg("cfg/imagenet.cfg");
srand(2222222);
- int i = 800;
+ int i = 0;
char *labels[] = {"cat","dog"};
float lr = .00001;
float momentum = .9;
float decay = 0.01;
- while(i++ < 1000 || 1){
- 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);
+ while(1){
+ i += 1000;
+ data train = load_data_image_pathfile_random("images/assira/train.list", 1000, labels, 2, 256, 256);
+ //image im = float_to_image(256, 256, 3,train.X.vals[0]);
+ //visualize_network(net);
+ //cvWaitKey(100);
+ //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);
+ 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%100==0){
+ if(i%10000==0){
char buff[256];
- sprintf(buff, "backup_%d.cfg", i);
- //save_network(net, buff);
+ sprintf(buff, "cfg/assira_backup_%d.cfg", i);
+ save_network(net, buff);
}
//lr *= .99;
}
}
+void test_visualize()
+{
+ network net = parse_network_cfg("cfg/imagenet.cfg");
+ srand(2222222);
+ visualize_network(net);
+ cvWaitKey(0);
+}
+void test_full()
+{
+ network net = parse_network_cfg("cfg/backup_1300.cfg");
+ srand(2222222);
+ int i,j;
+ int total = 100;
+ char *labels[] = {"cat","dog"};
+ FILE *fp = fopen("preds.txt","w");
+ for(i = 0; i < total; ++i){
+ visualize_network(net);
+ cvWaitKey(100);
+ data test = load_data_image_pathfile_part("images/assira/test.list", i, total, labels, 2, 256, 256);
+ image im = float_to_image(256, 256, 3,test.X.vals[0]);
+ show_image(im, "input");
+ cvWaitKey(100);
+ normalize_data_rows(test);
+ for(j = 0; j < test.X.rows; ++j){
+ float *x = test.X.vals[j];
+ forward_network(net, x);
+ int class = get_predicted_class_network(net);
+ fprintf(fp, "%d\n", class);
+ }
+ free_data(test);
+ }
+ fclose(fp);
+}
+
+void test_cifar10()
+{
+ data test = load_cifar10_data("images/cifar10/test_batch.bin");
+ scale_data_rows(test, 1./255);
+ network net = parse_network_cfg("cfg/cifar10.cfg");
+ int count = 0;
+ float lr = .000005;
+ float momentum = .99;
+ float decay = 0.001;
+ decay = 0;
+ int batch = 10000;
+ while(++count <= 10000){
+ char buff[256];
+ sprintf(buff, "images/cifar10/data_batch_%d.bin", rand()%5+1);
+ data train = load_cifar10_data(buff);
+ scale_data_rows(train, 1./255);
+ train_network_sgd(net, train, batch, lr, momentum, decay);
+ //printf("%5f %5f\n",(double)count*batch/train.X.rows, loss);
+
+ float test_acc = network_accuracy(net, test);
+ printf("%5f %5f\n",(double)count*batch/train.X.rows/5, 1-test_acc);
+ free_data(train);
+ }
+
+}
+
+void test_vince()
+{
+ network net = parse_network_cfg("cfg/vince.cfg");
+ data train = load_categorical_data_csv("images/vince.txt", 144, 2);
+ normalize_data_rows(train);
+
+ int count = 0;
+ float lr = .00005;
+ float momentum = .9;
+ float decay = 0.0001;
+ decay = 0;
+ int batch = 10000;
+ while(++count <= 10000){
+ float loss = train_network_sgd(net, train, batch, lr, momentum, decay);
+ printf("%5f %5f\n",(double)count*batch/train.X.rows, loss);
+ }
+}
+
void test_nist()
{
srand(444444);
srand(888888);
- network net = parse_network_cfg("nist.cfg");
+ network net = parse_network_cfg("cfg/nist_basic.cfg");
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;
- float lr = .0005;
+ float lr = .00005;
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;
+ float decay = 0.0001;
+ decay = 0;
+ //clock_t start = clock(), end;
+ int batch = 10000;
+ while(++count <= 10000){
+ float loss = train_network_sgd(net, train, batch, lr, momentum, decay);
+ printf("%5f %5f\n",(double)count*batch/train.X.rows, loss);
+ //printf("%5d Training Loss: %lf, Params: %f %f %f, ",count*1000, loss, lr, momentum, decay);
+ //end = clock();
+ //printf("Time: %lf seconds\n", (float)(end-start)/CLOCKS_PER_SEC);
+ //start=end;
+ /*
+ 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;
}
+ */
}
}
@@ -366,20 +447,21 @@
void train_VOC()
{
- network net = parse_network_cfg("cfg/voc_backup_ramp_80.cfg");
+ network net = parse_network_cfg("cfg/voc_start.cfg");
srand(2222222);
- int i = 0;
+ 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){
- visualize_network(net);
- cvWaitKey(100);
- data train = load_data_image_pathfile_random("images/VOC2012/train_paths.txt", 1000, labels, 20, 300, 400);
+ data train = load_data_image_pathfile_random("images/VOC2012/val_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);
@@ -388,25 +470,231 @@
free_data(train);
if(i%10==0){
char buff[256];
- sprintf(buff, "cfg/voc_backup_ramp_%d.cfg", i);
+ sprintf(buff, "cfg/voc_clean_ramp_%d.cfg", i);
save_network(net, buff);
}
//lr *= .99;
}
}
-int main()
+int voc_size(int x)
{
+ x = x-1+3;
+ x = x-1+3;
+ x = x-1+3;
+ x = (x-1)*2+1;
+ x = x-1+5;
+ x = (x-1)*2+1;
+ x = (x-1)*4+11;
+ return x;
+}
+
+image features_output_size(network net, IplImage *src, int outh, int outw)
+{
+ int h = voc_size(outh);
+ int w = voc_size(outw);
+ fprintf(stderr, "%d %d\n", h, w);
+
+ IplImage *sized = cvCreateImage(cvSize(w,h), src->depth, src->nChannels);
+ cvResize(src, sized, CV_INTER_LINEAR);
+ image im = ipl_to_image(sized);
+ resize_network(net, im.h, im.w, im.c);
+ forward_network(net, im.data);
+ image out = get_network_image_layer(net, 6);
+ free_image(im);
+ cvReleaseImage(&sized);
+ return copy_image(out);
+}
+
+void features_VOC_image_size(char *image_path, int h, int w)
+{
+ int j;
+ network net = parse_network_cfg("cfg/voc_imagenet.cfg");
+ fprintf(stderr, "%s\n", image_path);
+
+ IplImage* src = 0;
+ if( (src = cvLoadImage(image_path,-1)) == 0 ) file_error(image_path);
+ image out = features_output_size(net, src, h, w);
+ for(j = 0; j < out.c*out.h*out.w; ++j){
+ if(j != 0) printf(",");
+ printf("%g", out.data[j]);
+ }
+ printf("\n");
+ free_image(out);
+ cvReleaseImage(&src);
+}
+
+void visualize_imagenet_features(char *filename)
+{
+ int i,j,k;
+ network net = parse_network_cfg("cfg/voc_imagenet.cfg");
+ list *plist = get_paths(filename);
+ node *n = plist->front;
+ int h = voc_size(1), w = voc_size(1);
+ int num = get_network_image(net).c;
+ image *vizs = calloc(num, sizeof(image));
+ for(i = 0; i < num; ++i) vizs[i] = make_image(h, w, 3);
+ while(n){
+ char *image_path = (char *)n->val;
+ 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);
+ image out = get_network_image(net);
+
+ int dh = (im.h - h)/h;
+ int dw = (im.w - w)/w;
+ for(i = 0; i < out.h; ++i){
+ for(j = 0; j < out.w; ++j){
+ image sub = get_sub_image(im, dh*i, dw*j, h, w);
+ for(k = 0; k < out.c; ++k){
+ float val = get_pixel(out, i, j, k);
+ //printf("%f, ", val);
+ image sub_c = copy_image(sub);
+ scale_image(sub_c, val);
+ add_into_image(sub_c, vizs[k], 0, 0);
+ free_image(sub_c);
+ }
+ free_image(sub);
+ }
+ }
+ //printf("\n");
+ show_images(vizs, 10, "IMAGENET Visualization");
+ cvWaitKey(1000);
+ n = n->next;
+ }
+ cvWaitKey(0);
+}
+
+void features_VOC_image(char *image_file, char *image_dir, char *out_dir)
+{
+ int i,j;
+ network net = parse_network_cfg("cfg/voc_imagenet.cfg");
+ char image_path[1024];
+ sprintf(image_path, "%s/%s",image_dir, image_file);
+ char out_path[1024];
+ sprintf(out_path, "%s/%s.txt",out_dir, image_file);
+ printf("%s\n", image_file);
+ FILE *fp = fopen(out_path, "w");
+ if(fp == 0) file_error(out_path);
+
+ IplImage* src = 0;
+ if( (src = cvLoadImage(image_path,-1)) == 0 ) file_error(image_path);
+ int w = src->width;
+ int h = src->height;
+ int sbin = 8;
+ int interval = 4;
+ double scale = pow(2., 1./interval);
+ int m = (w<h)?w:h;
+ int max_scale = 1+floor((double)log((double)m/(5.*sbin))/log(scale));
+ if(max_scale < interval) error("max_scale must be >= interval");
+ image *ims = calloc(max_scale+interval, sizeof(image));
+
+ for(i = 0; i < interval; ++i){
+ double factor = 1./pow(scale, i);
+ double ih = round(h*factor);
+ double iw = round(w*factor);
+ int ex_h = round(ih/4.) - 2;
+ int ex_w = round(iw/4.) - 2;
+ ims[i] = features_output_size(net, src, ex_h, ex_w);
+
+ ih = round(h*factor);
+ iw = round(w*factor);
+ ex_h = round(ih/8.) - 2;
+ ex_w = round(iw/8.) - 2;
+ ims[i+interval] = features_output_size(net, src, ex_h, ex_w);
+ for(j = i+interval; j < max_scale; j += interval){
+ factor /= 2.;
+ ih = round(h*factor);
+ iw = round(w*factor);
+ ex_h = round(ih/8.) - 2;
+ ex_w = round(iw/8.) - 2;
+ ims[j+interval] = features_output_size(net, src, ex_h, ex_w);
+ }
+ }
+ for(i = 0; i < max_scale+interval; ++i){
+ image out = ims[i];
+ 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");
+ free_image(out);
+ }
+ free(ims);
+ fclose(fp);
+ cvReleaseImage(&src);
+}
+
+void test_distribution()
+{
+ IplImage* img = 0;
+ if( (img = cvLoadImage("im_small.jpg",-1)) == 0 ) file_error("im_small.jpg");
+ network net = parse_network_cfg("cfg/voc_features.cfg");
+ int h = img->height/8-2;
+ int w = img->width/8-2;
+ image out = features_output_size(net, img, h, w);
+ int c = out.c;
+ out.c = 1;
+ show_image(out, "output");
+ out.c = c;
+ image input = ipl_to_image(img);
+ show_image(input, "input");
+ CvScalar s;
+ int i,j;
+ image affects = make_image(input.h, input.w, 1);
+ int count = 0;
+ for(i = 0; i<img->height; i += 1){
+ for(j = 0; j < img->width; j += 1){
+ IplImage *copy = cvCloneImage(img);
+ s=cvGet2D(copy,i,j); // get the (i,j) pixel value
+ printf("%d/%d\n", count++, img->height*img->width);
+ s.val[0]=0;
+ s.val[1]=0;
+ s.val[2]=0;
+ cvSet2D(copy,i,j,s); // set the (i,j) pixel value
+ image mod = features_output_size(net, copy, h, w);
+ image dist = image_distance(out, mod);
+ show_image(affects, "affects");
+ cvWaitKey(1);
+ cvReleaseImage(©);
+ //affects.data[i*affects.w + j] += dist.data[3*dist.w+5];
+ affects.data[i*affects.w + j] += dist.data[1*dist.w+1];
+ free_image(mod);
+ free_image(dist);
+ }
+ }
+ show_image(affects, "Origins");
+ cvWaitKey(0);
+ cvWaitKey(0);
+}
+
+
+int main(int argc, char *argv[])
+{
+ //train_full();
+ //test_distribution();
//feenableexcept(FE_DIVBYZERO | FE_INVALID | FE_OVERFLOW);
//test_blas();
+ //test_visualize();
+ //test_gpu_blas();
+ //test_blas();
//test_convolve_matrix();
// test_im2row();
//test_split();
//test_ensemble();
//test_nist();
+ //test_cifar10();
+ //test_vince();
//test_full();
- train_VOC();
+ //train_VOC();
+ //features_VOC_image(argv[1], argv[2], argv[3]);
+ //features_VOC_image_size(argv[1], atoi(argv[2]), atoi(argv[3]));
+ //visualize_imagenet_features("data/assira/train.list");
+ visualize_imagenet_features("data/VOC2011.list");
+ fprintf(stderr, "Success!\n");
//test_random_preprocess();
//test_random_classify();
//test_parser();
--
Gitblit v1.10.0