Visualizations?

Joseph Redmon

2014-04-17 738cd4c2d7abcf62b85b759a5765b08380ee90e8

 src/tests.c

@@ -1,4 +1,5 @@
#include "connected_layer.h"

//#include "old_conv.h"
#include "convolutional_layer.h"
#include "maxpool_layer.h"
@@ -220,6 +221,14 @@
        //lr *= .99;
    }
}

void test_visualize()
{
    network net = parse_network_cfg("cfg/voc_imagenet.cfg");
    srand(2222222);
    visualize_network(net);
    cvWaitKey(0);
}
void test_full()
{
    network net = parse_network_cfg("cfg/backup_1300.cfg");
@@ -265,7 +274,7 @@
        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);
@@ -316,15 +325,15 @@
        //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;
           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;
        }
        */
         */
    }
}

@@ -437,6 +446,12 @@
    }
}

void flip_network()
{
    network net = parse_network_cfg("cfg/voc_imagenet_orig.cfg");
    save_network(net, "cfg/voc_imagenet_rev.cfg");
}

void train_VOC()
{
    network net = parse_network_cfg("cfg/voc_start.cfg");
@@ -490,6 +505,7 @@
    IplImage *sized = cvCreateImage(cvSize(w,h), src->depth, src->nChannels);
    cvResize(src, sized, CV_INTER_LINEAR);
    image im = ipl_to_image(sized);
    normalize_array(im.data, im.h*im.w*im.c);
    resize_network(net, im.h, im.w, im.c);
    forward_network(net, im.data);
    image out = get_network_image_layer(net, 6);
@@ -515,6 +531,125 @@
    free_image(out);
    cvReleaseImage(&src);
}
void visualize_imagenet_topk(char *filename)
{
    int i,j,k,l;
    int topk = 10;
    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*));
    float **score = calloc(num, sizeof(float *));
    for(i = 0; i < num; ++i){
        vizs[i] = calloc(topk, sizeof(image));
        for(j = 0; j < topk; ++j) vizs[i][j] = make_image(h,w,3);
        score[i] = calloc(topk, sizeof(float));
    }

    while(n){
        char *image_path = (char *)n->val;
        image im = load_image(image_path, 0, 0);
        n = n->next;
        if(im.h < 200 || im.w < 200) continue;
        printf("Processing %dx%d image\n", im.h, im.w);
        resize_network(net, im.h, im.w, im.c);
        //scale_image(im, 1./255);
        translate_image(im, -144);
        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);
                    for(l = 0; l < topk; ++l){
                        if(val > score[k][l]){
                            float swap = score[k][l];
                            score[k][l] = val;
                            val = swap;

                            image swapi = vizs[k][l];
                            vizs[k][l] = sub_c;
                            sub_c = swapi;
                        }
                    }
                    free_image(sub_c);
                }
                free_image(sub);
            }
        }
        free_image(im);
        //printf("\n");
        image grid = grid_images(vizs, num, topk);
        show_image(grid, "IMAGENET Visualization");
        save_image(grid, "IMAGENET Grid");
        free_image(grid);
    }
    //cvWaitKey(0);
}

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 visualize_cat()
{
    network net = parse_network_cfg("cfg/voc_imagenet.cfg");
    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);
    
    image out = get_network_image(net);
    visualize_network(net);
    cvWaitKey(1000);
    cvWaitKey(0);
}

void features_VOC_image(char *image_file, char *image_dir, char *out_dir)
{
@@ -628,6 +763,9 @@
    //feenableexcept(FE_DIVBYZERO | FE_INVALID | FE_OVERFLOW);

    //test_blas();
    //test_visualize();
    //test_gpu_blas();
    //test_blas();
    //test_convolve_matrix();
    //    test_im2row();
    //test_split();
@@ -638,7 +776,12 @@
    //test_full();
    //train_VOC();
    //features_VOC_image(argv[1], argv[2], argv[3]);
    features_VOC_image_size(argv[1], atoi(argv[2]), atoi(argv[3]));
    //features_VOC_image_size(argv[1], atoi(argv[2]), atoi(argv[3]));
    //visualize_imagenet_features("data/assira/train.list");
    visualize_imagenet_topk("data/VOC2011.list");
    //visualize_cat();
    //flip_network();
    //test_visualize();
    fprintf(stderr, "Success!\n");
    //test_random_preprocess();
    //test_random_classify();