something

Joseph Redmon

2015-05-12 7399dd1af577cbba5f4b9124c917364fd7613dcb

 src/detection.c

@@ -1,5 +1,6 @@
#include "network.h"
#include "detection_layer.h"
#include "cost_layer.h"
#include "utils.h"
#include "parser.h"

@@ -7,7 +8,7 @@
char *class_names[] = {"aeroplane", "bicycle", "bird", "boat", "bottle", "bus", "car", "cat", "chair", "cow", "diningtable", "dog", "horse", "motorbike", "person", "pottedplant", "sheep", "sofa", "train", "tvmonitor"};
char *inet_class_names[] = {"bg", "accordion", "airplane", "ant", "antelope", "apple", "armadillo", "artichoke", "axe", "baby bed", "backpack", "bagel", "balance beam", "banana", "band aid", "banjo", "baseball", "basketball", "bathing cap", "beaker", "bear", "bee", "bell pepper", "bench", "bicycle", "binder", "bird", "bookshelf", "bow tie", "bow", "bowl", "brassiere", "burrito", "bus", "butterfly", "camel", "can opener", "car", "cart", "cattle", "cello", "centipede", "chain saw", "chair", "chime", "cocktail shaker", "coffee maker", "computer keyboard", "computer mouse", "corkscrew", "cream", "croquet ball", "crutch", "cucumber", "cup or mug", "diaper", "digital clock", "dishwasher", "dog", "domestic cat", "dragonfly", "drum", "dumbbell", "electric fan", "elephant", "face powder", "fig", "filing cabinet", "flower pot", "flute", "fox", "french horn", "frog", "frying pan", "giant panda", "goldfish", "golf ball", "golfcart", "guacamole", "guitar", "hair dryer", "hair spray", "hamburger", "hammer", "hamster", "harmonica", "harp", "hat with a wide brim", "head cabbage", "helmet", "hippopotamus", "horizontal bar", "horse", "hotdog", "iPod", "isopod", "jellyfish", "koala bear", "ladle", "ladybug", "lamp", "laptop", "lemon", "lion", "lipstick", "lizard", "lobster", "maillot", "maraca", "microphone", "microwave", "milk can", "miniskirt", "monkey", "motorcycle", "mushroom", "nail", "neck brace", "oboe", "orange", "otter", "pencil box", "pencil sharpener", "perfume", "person", "piano", "pineapple", "ping-pong ball", "pitcher", "pizza", "plastic bag", "plate rack", "pomegranate", "popsicle", "porcupine", "power drill", "pretzel", "printer", "puck", "punching bag", "purse", "rabbit", "racket", "ray", "red panda", "refrigerator", "remote control", "rubber eraser", "rugby ball", "ruler", "salt or pepper shaker", "saxophone", "scorpion", "screwdriver", "seal", "sheep", "ski", "skunk", "snail", "snake", "snowmobile", "snowplow", "soap dispenser", "soccer ball", "sofa", "spatula", "squirrel", "starfish", "stethoscope", "stove", "strainer", "strawberry", "stretcher", "sunglasses", "swimming trunks", "swine", "syringe", "table", "tape player", "tennis ball", "tick", "tie", "tiger", "toaster", "traffic light", "train", "trombone", "trumpet", "turtle", "tv or monitor", "unicycle", "vacuum", "violin", "volleyball", "waffle iron", "washer", "water bottle", "watercraft", "whale", "wine bottle", "zebra"};
#define AMNT 3
void draw_detection(image im, float *box, int side)
void draw_detection(image im, float *box, int side, char *label)
{
    int classes = 20;
    int elems = 4+classes;
@@ -20,7 +21,7 @@
            //printf("%d\n", j);
            //printf("Prob: %f\n", box[j]);
            int class = max_index(box+j, classes);
            if(box[j+class] > .2){
            if(box[j+class] > .4){
                //int z;
                //for(z = 0; z < classes; ++z) printf("%f %s\n", box[j+z], class_names[z]);
                printf("%f %s\n", box[j+class], class_names[class]);
@@ -35,8 +36,8 @@
                float x = box[j+1];
                x = (x+c)/side;
                y = (y+r)/side;
                float h = box[j+2]; //*maxheight;
                float w = box[j+3]; //*maxwidth;
                float w = box[j+2]; //*maxwidth;
                float h = box[j+3]; //*maxheight;
                h = h*h;
                w = w*w;
                //printf("coords %f %f %f %f\n", x, y, w, h);
@@ -50,8 +51,177 @@
        }
    }
    //printf("Done\n");
    show_image(im, "box");
    cvWaitKey(0);
    show_image(im, label);
}

void draw_localization(image im, float *box)
{
    int classes = 20;
    int class;
    for(class = 0; class < classes; ++class){
        //int z;
        //for(z = 0; z < classes; ++z) printf("%f %s\n", box[j+z], class_names[z]);
        float red = get_color(0,class,classes);
        float green = get_color(1,class,classes);
        float blue = get_color(2,class,classes);

        int j = class*4;
        float x = box[j+0];
        float y = box[j+1];
        float w = box[j+2]; //*maxheight;
        float h = box[j+3]; //*maxwidth;
        //printf("coords %f %f %f %f\n", x, y, w, h);

        int left  = (x-w/2)*im.w;
        int right = (x+w/2)*im.w;
        int top   = (y-h/2)*im.h;
        int bot   = (y+h/2)*im.h;
        draw_box(im, left, top, right, bot, red, green, blue);
    }
    //printf("Done\n");
}

void train_localization(char *cfgfile, char *weightfile)
{
    srand(time(0));
    data_seed = time(0);
    char *base = basecfg(cfgfile);
    printf("%s\n", base);
    float avg_loss = -1;
    network net = parse_network_cfg(cfgfile);
    if(weightfile){
        load_weights(&net, weightfile);
    }
    printf("Learning Rate: %g, Momentum: %g, Decay: %g\n", net.learning_rate, net.momentum, net.decay);
    int imgs = 128;
    int classes = 20;
    int i = net.seen/imgs;
    data train, buffer;

    char **paths;
    list *plist;
    plist = get_paths("/home/pjreddie/data/voc/loc.2012val.txt");
    paths = (char **)list_to_array(plist);
    pthread_t load_thread = load_data_localization_thread(imgs, paths, plist->size, classes, net.w, net.h, &buffer);
    clock_t time;
    while(1){
        i += 1;
        time=clock();
        pthread_join(load_thread, 0);
        train = buffer;
        load_thread = load_data_localization_thread(imgs, paths, plist->size, classes, net.w, net.h, &buffer);

        printf("Loaded: %lf seconds\n", sec(clock()-time));
        time=clock();
        float loss = train_network(net, train);

        //TODO
        float *out = get_network_output_gpu(net);
        image im = float_to_image(net.w, net.h, 3, train.X.vals[127]);
        image copy = copy_image(im);
        draw_localization(copy, &(out[63*80]));
        draw_localization(copy, train.y.vals[127]);
        show_image(copy, "box");
        cvWaitKey(0);
        free_image(copy);

        net.seen += imgs;
        if (avg_loss < 0) avg_loss = loss;
        avg_loss = avg_loss*.9 + loss*.1;
        printf("%d: %f, %f avg, %lf seconds, %d images\n", i, loss, avg_loss, sec(clock()-time), i*imgs);
        if(i%100==0){
            char buff[256];
            sprintf(buff, "/home/pjreddie/imagenet_backup/%s_%d.weights",base, i);
            save_weights(net, buff);
        }
        free_data(train);
    }
}

void train_detection_teststuff(char *cfgfile, char *weightfile)
{
    srand(time(0));
    data_seed = time(0);
    int imgnet = 0;
    char *base = basecfg(cfgfile);
    printf("%s\n", base);
    float avg_loss = -1;
    network net = parse_network_cfg(cfgfile);
    if(weightfile){
        load_weights(&net, weightfile);
    }
    detection_layer layer = get_network_detection_layer(net);
    net.learning_rate = 0;
    net.decay = 0;
    printf("Learning Rate: %g, Momentum: %g, Decay: %g\n", net.learning_rate, net.momentum, net.decay);
    int imgs = 128;
    int i = net.seen/imgs;
    data train, buffer;

    int classes = layer.classes;
    int background = layer.background;
    int side = sqrt(get_detection_layer_locations(layer));

    char **paths;
    list *plist;
    if (imgnet){
        plist = get_paths("/home/pjreddie/data/imagenet/det.train.list");
    }else{
        plist = get_paths("/home/pjreddie/data/voc/val_2012.txt");
        //plist = get_paths("/home/pjreddie/data/voc/no_2007_test.txt");
        //plist = get_paths("/home/pjreddie/data/coco/trainval.txt");
        //plist = get_paths("/home/pjreddie/data/voc/all2007-2012.txt");
    }
    paths = (char **)list_to_array(plist);
    pthread_t load_thread = load_data_detection_thread(imgs, paths, plist->size, classes, net.w, net.h, side, side, background, &buffer);
    clock_t time;
    cost_layer clayer = net.layers[net.n-1];
    while(1){
        i += 1;
        time=clock();
        pthread_join(load_thread, 0);
        train = buffer;
        load_thread = load_data_detection_thread(imgs, paths, plist->size, classes, net.w, net.h, side, side, background, &buffer);

        /*
           image im = float_to_image(net.w, net.h, 3, train.X.vals[114]);
           image copy = copy_image(im);
           draw_detection(copy, train.y.vals[114], 7);
           free_image(copy);
         */

        int z;
        int count = 0;
        float sx, sy, sw, sh;
        sx = sy = sw = sh = 0;
        for(z = 0; z < clayer.batch*clayer.inputs; z += 24){
            if(clayer.delta[z+20]){
                ++count;
                sx += fabs(clayer.delta[z+20])*64;
                sy += fabs(clayer.delta[z+21])*64;
                sw += fabs(clayer.delta[z+22])*448;
                sh += fabs(clayer.delta[z+23])*448;
            }
        }
        printf("Avg error: %f, %f, %f x %f\n", sx/count, sy/count, sw/count, sh/count);

        printf("Loaded: %lf seconds\n", sec(clock()-time));
        time=clock();
        float loss = train_network(net, train);
        net.seen += imgs;
        if (avg_loss < 0) avg_loss = loss;
        avg_loss = avg_loss*.9 + loss*.1;
        printf("%d: %f, %f avg, %lf seconds, %d images\n", i, loss, avg_loss, sec(clock()-time), i*imgs);
        if(i == 100){
            net.learning_rate *= 10;
        }
        if(i%100==0){
            char buff[256];
            sprintf(buff, "/home/pjreddie/imagenet_backup/%s_%d.weights",base, i);
            save_weights(net, buff);
        }
        free_data(train);
    }
}

void train_detection(char *cfgfile, char *weightfile)
@@ -66,22 +236,23 @@
    if(weightfile){
        load_weights(&net, weightfile);
    }
    detection_layer *layer = get_network_detection_layer(net);
    detection_layer layer = get_network_detection_layer(net);
    printf("Learning Rate: %g, Momentum: %g, Decay: %g\n", net.learning_rate, net.momentum, net.decay);
    int imgs = 128;
    int i = net.seen/imgs;
    data train, buffer;

    int classes = layer->classes;
    int background = layer->background;
    int side = sqrt(get_detection_layer_locations(*layer));
    int classes = layer.classes;
    int background = layer.background;
    int side = sqrt(get_detection_layer_locations(layer));

    char **paths;
    list *plist;
    if (imgnet){
        plist = get_paths("/home/pjreddie/data/imagenet/det.train.list");
    }else{
        plist = get_paths("/home/pjreddie/data/voc/trainall.txt");
        plist = get_paths("/home/pjreddie/data/voc/no_2012_val.txt");
        //plist = get_paths("/home/pjreddie/data/voc/no_2007_test.txt");
        //plist = get_paths("/home/pjreddie/data/coco/trainval.txt");
        //plist = get_paths("/home/pjreddie/data/voc/all2007-2012.txt");
    }
@@ -109,6 +280,9 @@
        if (avg_loss < 0) avg_loss = loss;
        avg_loss = avg_loss*.9 + loss*.1;
        printf("%d: %f, %f avg, %lf seconds, %d images\n", i, loss, avg_loss, sec(clock()-time), i*imgs);
        if(i == 100){
            net.learning_rate *= 10;
        }
        if(i%100==0){
            char buff[256];
            sprintf(buff, "/home/pjreddie/imagenet_backup/%s_%d.weights",base, i);
@@ -131,15 +305,15 @@
            if (nuisance) scale = 1.-pred.vals[j][k];
            for (class = 0; class < classes; ++class){
                int ci = k+classes+background+nuisance;
                float y = (pred.vals[j][ci + 0] + row)/num_boxes;
                float x = (pred.vals[j][ci + 1] + col)/num_boxes;
                float h = pred.vals[j][ci + 2]; //* distance_from_edge(row, num_boxes);
                h = h*h;
                float w = pred.vals[j][ci + 3]; //* distance_from_edge(col, num_boxes);
                w = w*w;
                float x = (pred.vals[j][ci + 0] + col)/num_boxes;
                float y = (pred.vals[j][ci + 1] + row)/num_boxes;
                float w = pred.vals[j][ci + 2]; //* distance_from_edge(row, num_boxes);
                float h = pred.vals[j][ci + 3]; //* distance_from_edge(col, num_boxes);
                w = pow(w, 2);
                h = pow(h, 2);
                float prob = scale*pred.vals[j][k+class+background+nuisance];
                if(prob < threshold) continue;
                printf("%d %d %f %f %f %f %f\n", offset +  j, class, prob, y, x, h, w);
                printf("%d %d %f %f %f %f %f\n", offset +  j, class, prob, x, y, w, h);
            }
        }
    }
@@ -152,20 +326,21 @@
    if(weightfile){
        load_weights(&net, weightfile);
    }
    detection_layer *layer = get_network_detection_layer(net);
    detection_layer layer = get_network_detection_layer(net);
    fprintf(stderr, "Learning Rate: %g, Momentum: %g, Decay: %g\n", net.learning_rate, net.momentum, net.decay);
    srand(time(0));

    list *plist = get_paths("/home/pjreddie/data/voc/val.txt");
    //list *plist = get_paths("/home/pjreddie/data/voc/test_2007.txt");
    list *plist = get_paths("/home/pjreddie/data/voc/val_2012.txt");
    //list *plist = get_paths("/home/pjreddie/data/voc/test.txt");
    //list *plist = get_paths("/home/pjreddie/data/voc/val.expanded.txt");
    //list *plist = get_paths("/home/pjreddie/data/voc/train.txt");
    char **paths = (char **)list_to_array(plist);

    int classes = layer->classes;
    int nuisance = layer->nuisance;
    int background = (layer->background && !nuisance);
    int num_boxes = sqrt(get_detection_layer_locations(*layer));
    int classes = layer.classes;
    int nuisance = layer.nuisance;
    int background = (layer.background && !nuisance);
    int num_boxes = sqrt(get_detection_layer_locations(layer));

    int per_box = 4+classes+background+nuisance;
    int num_output = num_boxes*num_boxes*per_box;
@@ -207,6 +382,130 @@
    }
}

void validate_detection_post(char *cfgfile, char *weightfile)
{
    network net = parse_network_cfg(cfgfile);
    if(weightfile){
        load_weights(&net, weightfile);
    }
    set_batch_network(&net, 1);

    network post = parse_network_cfg("cfg/localize.cfg");
    load_weights(&post, "/home/pjreddie/imagenet_backup/localize_1000.weights");
    set_batch_network(&post, 1);

    detection_layer layer = get_network_detection_layer(net);
    fprintf(stderr, "Learning Rate: %g, Momentum: %g, Decay: %g\n", net.learning_rate, net.momentum, net.decay);
    srand(time(0));

    //list *plist = get_paths("/home/pjreddie/data/voc/test_2007.txt");
    list *plist = get_paths("/home/pjreddie/data/voc/val_2012.txt");
    //list *plist = get_paths("/home/pjreddie/data/voc/test.txt");
    //list *plist = get_paths("/home/pjreddie/data/voc/val.expanded.txt");
    //list *plist = get_paths("/home/pjreddie/data/voc/train.txt");
    char **paths = (char **)list_to_array(plist);

    int classes = layer.classes;
    int nuisance = layer.nuisance;
    int background = (layer.background && !nuisance);
    int num_boxes = sqrt(get_detection_layer_locations(layer));

    int per_box = 4+classes+background+nuisance;

    int m = plist->size;
    int i = 0;
    float threshold = .01;

    clock_t time = clock();
    for(i = 0; i < m; ++i){
        image im = load_image_color(paths[i], 0, 0);
        if(i % 100 == 0) {
            fprintf(stderr, "%d: Loaded: %lf seconds\n", i, sec(clock()-time));
            time = clock();
        }
        image sized = resize_image(im, net.w, net.h);
        float *out = network_predict(net, sized.data);
        free_image(sized);
        int k, class;
        //show_image(im, "original");
        int num_output = num_boxes*num_boxes*per_box;
        //image cp1 = copy_image(im);
        //draw_detection(cp1, out, 7, "before");
        for(k = 0; k < num_output; k += per_box){
            float *post_out = 0;
            float scale = 1.;
            int index = k/per_box;
            int row = index / num_boxes;
            int col = index % num_boxes;
            if (nuisance) scale = 1.-out[k];
            for (class = 0; class < classes; ++class){
                int ci = k+classes+background+nuisance;
                float x = (out[ci + 0] + col)/num_boxes;
                float y = (out[ci + 1] + row)/num_boxes;
                float w = out[ci + 2]; //* distance_from_edge(row, num_boxes);
                float h = out[ci + 3]; //* distance_from_edge(col, num_boxes);
                w = w*w;
                h = h*h;
                float prob = scale*out[k+class+background+nuisance];
                if (prob >= threshold) {
                    x *= im.w;
                    y *= im.h;
                    w *= im.w;
                    h *= im.h;
                    w += 32;
                    h += 32;
                    int left = (x - w/2);
                    int top = (y - h/2);
                    int right = (x + w/2);
                    int bot = (y+h/2);
                    if (left < 0) left = 0;
                    if (right > im.w) right = im.w;
                    if (top < 0) top = 0;
                    if (bot > im.h) bot = im.h;

                    image crop = crop_image(im, left, top, right-left, bot-top);
                    image resize = resize_image(crop, post.w, post.h);
                    if (!post_out){
                        post_out = network_predict(post, resize.data);
                    }
                    /*
                    draw_localization(resize, post_out);
                    show_image(resize, "second");
                    fprintf(stderr, "%s\n", class_names[class]);
                    cvWaitKey(0);
                    */
                    int index = 4*class;
                    float px = post_out[index+0];
                    float py = post_out[index+1];
                    float pw = post_out[index+2];
                    float ph = post_out[index+3];
                    px = (px * crop.w + left) / im.w;
                    py = (py * crop.h + top) / im.h;
                    pw = (pw * crop.w) / im.w;
                    ph = (ph * crop.h) / im.h;

                    out[ci + 0] = px*num_boxes - col;
                    out[ci + 1] = py*num_boxes - row;
                    out[ci + 2] = sqrt(pw);
                    out[ci + 3] = sqrt(ph);
                    /*
                       show_image(crop, "cropped");
                       cvWaitKey(0);
                     */
                    free_image(crop);
                    free_image(resize);
                    printf("%d %d %f %f %f %f %f\n", i, class, prob, px, py, pw, ph);
                }
            }
        }
        /*
        image cp2 = copy_image(im);
        draw_detection(cp2, out, 7, "after");
        cvWaitKey(0);
        */
    }
}

void test_detection(char *cfgfile, char *weightfile)
{
    network net = parse_network_cfg(cfgfile);
@@ -227,7 +526,7 @@
        time=clock();
        float *predictions = network_predict(net, X);
        printf("%s: Predicted in %f seconds.\n", filename, sec(clock()-time));
        draw_detection(im, predictions, 7);
        draw_detection(im, predictions, 7, "detections");
        free_image(im);
    }
}
@@ -243,5 +542,8 @@
    char *weights = (argc > 4) ? argv[4] : 0;
    if(0==strcmp(argv[2], "test")) test_detection(cfg, weights);
    else if(0==strcmp(argv[2], "train")) train_detection(cfg, weights);
    else if(0==strcmp(argv[2], "teststuff")) train_detection_teststuff(cfg, weights);
    else if(0==strcmp(argv[2], "trainloc")) train_localization(cfg, weights);
    else if(0==strcmp(argv[2], "valid")) validate_detection(cfg, weights);
    else if(0==strcmp(argv[2], "validpost")) validate_detection_post(cfg, weights);
}