Go

Makefile

@@ -1,8 +1,8 @@
-GPU=1
-OPENCV=1
+GPU=0
+OPENCV=0
 DEBUG=0
 
-ARCH= --gpu-architecture=compute_20 --gpu-code=compute_20
+ARCH= --gpu-architecture=compute_20 --gpu-code=compute_20 
 
 VPATH=./src/
 EXEC=darknet
@@ -34,9 +34,9 @@ CFLAGS+= -DGPU
 LDFLAGS+= -L/usr/local/cuda/lib64 -lcuda -lcudart -lcublas -lcurand
 endif
 
-OBJ=gemm.o utils.o cuda.o deconvolutional_layer.o convolutional_layer.o list.o image.o activations.o im2col.o col2im.o blas.o crop_layer.o dropout_layer.o maxpool_layer.o softmax_layer.o data.o matrix.o network.o connected_layer.o cost_layer.o parser.o option_list.o darknet.o detection_layer.o imagenet.o captcha.o route_layer.o writing.o box.o nightmare.o normalization_layer.o avgpool_layer.o coco.o dice.o yolo.o layer.o compare.o classifier.o local_layer.o swag.o shortcut_layer.o activation_layer.o rnn_layer.o rnn.o rnn_vid.o crnn_layer.o coco_demo.o tag.o cifar.o
+OBJ=gemm.o utils.o cuda.o deconvolutional_layer.o convolutional_layer.o list.o image.o activations.o im2col.o col2im.o blas.o crop_layer.o dropout_layer.o maxpool_layer.o softmax_layer.o data.o matrix.o network.o connected_layer.o cost_layer.o parser.o option_list.o darknet.o detection_layer.o imagenet.o captcha.o route_layer.o writing.o box.o nightmare.o normalization_layer.o avgpool_layer.o coco.o dice.o yolo.o layer.o compare.o classifier.o local_layer.o swag.o shortcut_layer.o activation_layer.o rnn_layer.o rnn.o rnn_vid.o crnn_layer.o coco_demo.o tag.o cifar.o yolo_demo.o go.o
 ifeq ($(GPU), 1) 
-OBJ+=convolutional_kernels.o deconvolutional_kernels.o activation_kernels.o im2col_kernels.o col2im_kernels.o blas_kernels.o crop_layer_kernels.o dropout_layer_kernels.o maxpool_layer_kernels.o softmax_layer_kernels.o network_kernels.o avgpool_layer_kernels.o yolo_kernels.o
+OBJ+=convolutional_kernels.o deconvolutional_kernels.o activation_kernels.o im2col_kernels.o col2im_kernels.o blas_kernels.o crop_layer_kernels.o dropout_layer_kernels.o maxpool_layer_kernels.o softmax_layer_kernels.o network_kernels.o avgpool_layer_kernels.o
 endif
 
 OBJS = $(addprefix $(OBJDIR), $(OBJ))

cfg/go.test.cfg

@@ -0,0 +1,67 @@
+[net]
+batch=1
+subdivisions=1
+height=19
+width=19
+channels=1
+momentum=0.9
+decay=0.0005
+
+learning_rate=0.1
+max_batches = 0
+policy=steps
+steps=50000, 90000
+scales=.1, .1
+
+[convolutional]
+filters=256
+size=3
+stride=1
+pad=1
+activation=leaky
+batch_normalize=1
+
+[convolutional]
+filters=256
+size=3
+stride=1
+pad=1
+activation=leaky
+batch_normalize=1
+
+[convolutional]
+filters=256
+size=3
+stride=1
+pad=1
+activation=leaky
+batch_normalize=1
+
+[convolutional]
+filters=256
+size=3
+stride=1
+pad=1
+activation=leaky
+batch_normalize=1
+
+[convolutional]
+filters=256
+size=3
+stride=1
+pad=1
+activation=leaky
+batch_normalize=1
+
+[convolutional]
+filters=1
+size=1
+stride=1
+pad=1
+activation=leaky
+
+[softmax]
+
+[cost]
+type=sse
+

src/blas.c

@@ -46,7 +46,7 @@ void mean_cpu(float *x, int batch, int filters, int spatial, float *mean)
 
 void variance_cpu(float *x, float *mean, int batch, int filters, int spatial, float *variance)
 {
-    float scale = 1./(batch * spatial);
+    float scale = 1./(batch * spatial - 1);
     int i,j,k;
     for(i = 0; i < filters; ++i){
         variance[i] = 0;
@@ -67,7 +67,7 @@ void normalize_cpu(float *x, float *mean, float *variance, int batch, int filter
         for(f = 0; f < filters; ++f){
             for(i = 0; i < spatial; ++i){
                 int index = b*filters*spatial + f*spatial + i;
-                x[index] = (x[index] - mean[f])/(sqrt(variance[f]) + .00001f);
+                x[index] = (x[index] - mean[f])/(sqrt(variance[f]) + .000001f);
             }
         }
     }

src/blas_kernels.cu

@@ -15,7 +15,7 @@ __global__ void normalize_kernel(int N, float *x, float *mean, float *variance,
     if (index >= N) return;
     int f = (index/spatial)%filters;
     
-    x[index] = (x[index] - mean[f])/(sqrt(variance[f]) + .00001f);
+    x[index] = (x[index] - mean[f])/(sqrt(variance[f]) + .000001f);
 }
 
 __global__ void normalize_delta_kernel(int N, float *x, float *mean, float *variance, float *mean_delta, float *variance_delta, int batch, int filters, int spatial, float *delta)
@@ -24,7 +24,7 @@ __global__ void normalize_delta_kernel(int N, float *x, float *mean, float *vari
     if (index >= N) return;
     int f = (index/spatial)%filters;
     
-    delta[index] = delta[index] * 1./(sqrt(variance[f]) + .00001f) + variance_delta[f] * 2. * (x[index] - mean[f]) / (spatial * batch) + mean_delta[f]/(spatial*batch);
+    delta[index] = delta[index] * 1./(sqrt(variance[f]) + .000001f) + variance_delta[f] * 2. * (x[index] - mean[f]) / (spatial * batch) + mean_delta[f]/(spatial*batch);
 }
 
 extern "C" void normalize_delta_gpu(float *x, float *mean, float *variance, float *mean_delta, float *variance_delta, int batch, int filters, int spatial, float *delta)
@@ -46,7 +46,7 @@ __global__ void  variance_delta_kernel(float *x, float *delta, float *mean, floa
             variance_delta[i] += delta[index]*(x[index] - mean[i]);
         }
     }
-    variance_delta[i] *= -.5 * pow(variance[i] + .00001f, (float)(-3./2.));
+    variance_delta[i] *= -.5 * pow(variance[i] + .000001f, (float)(-3./2.));
 }
 
 __global__ void accumulate_kernel(float *x, int n, int groups, float *sum)
@@ -83,7 +83,7 @@ __global__ void fast_mean_delta_kernel(float *delta, float *variance, int batch,
         for(i = 0; i < threads; ++i){
             mean_delta[filter] += local[i];
         }
-        mean_delta[filter] *= (-1./sqrt(variance[filter] + .00001f));
+        mean_delta[filter] *= (-1./sqrt(variance[filter] + .000001f));
     }
 }
 
@@ -111,7 +111,7 @@ __global__ void  fast_variance_delta_kernel(float *x, float *delta, float *mean,
         for(i = 0; i < threads; ++i){
             variance_delta[filter] += local[i];
         }
-        variance_delta[filter] *= -.5 * pow(variance[filter] + .00001f, (float)(-3./2.));
+        variance_delta[filter] *= -.5 * pow(variance[filter] + .000001f, (float)(-3./2.));
     }
 }
 
@@ -128,7 +128,7 @@ __global__ void mean_delta_kernel(float *delta, float *variance, int batch, int
             mean_delta[i] += delta[index];
         }
     }
-    mean_delta[i] *= (-1./sqrt(variance[i] + .00001f));
+    mean_delta[i] *= (-1./sqrt(variance[i] + .000001f));
 }
 
 extern "C" void mean_delta_gpu(float *delta, float *variance, int batch, int filters, int spatial, float *mean_delta)
@@ -167,7 +167,7 @@ __global__ void  mean_kernel(float *x, int batch, int filters, int spatial, floa
 
 __global__ void variance_kernel(float *x, float *mean, int batch, int filters, int spatial, float *variance)
 {
-    float scale = 1./(batch * spatial);
+    float scale = 1./(batch * spatial - 1);
     int j,k;
     int i = (blockIdx.x + blockIdx.y*gridDim.x) * blockDim.x + threadIdx.x;
     if (i >= filters) return;
@@ -288,7 +288,7 @@ __global__ void  fast_variance_kernel(float *x, float *mean, int batch, int filt
         for(i = 0; i < threads; ++i){
             variance[filter] += local[i];
         }
-        variance[filter] /= spatial * batch;
+        variance[filter] /= (spatial * batch - 1);
     }
 }
 

src/cifar.c

@@ -33,7 +33,7 @@ void train_cifar(char *cfgfile, char *weightfile)
 
         float loss = train_network_sgd(net, train, 1);
         if(avg_loss == -1) avg_loss = loss;
-        avg_loss = avg_loss*.9 + loss*.1;
+        avg_loss = avg_loss*.95 + loss*.05;
         printf("%d, %.3f: %f, %f avg, %f rate, %lf seconds, %d images\n", get_current_batch(net), (float)(*net.seen)/N, loss, avg_loss, get_current_rate(net), sec(clock()-time), *net.seen);
         if(*net.seen/N > epoch){
             epoch = *net.seen/N;
@@ -57,6 +57,95 @@ void train_cifar(char *cfgfile, char *weightfile)
     free_data(train);
 }
 
+void train_cifar_distill(char *cfgfile, char *weightfile)
+{
+    data_seed = time(0);
+    srand(time(0));
+    float avg_loss = -1;
+    char *base = basecfg(cfgfile);
+    printf("%s\n", base);
+    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);
+
+    char *backup_directory = "/home/pjreddie/backup/";
+    int classes = 10;
+    int N = 50000;
+
+    char **labels = get_labels("data/cifar/labels.txt");
+    int epoch = (*net.seen)/N;
+
+    data train = load_all_cifar10();
+    matrix soft = csv_to_matrix("results/ensemble.csv");
+
+    float weight = .9;
+    scale_matrix(soft, weight);
+    scale_matrix(train.y, 1. - weight);
+    matrix_add_matrix(soft, train.y);
+
+    while(get_current_batch(net) < net.max_batches || net.max_batches == 0){
+        clock_t time=clock();
+
+        float loss = train_network_sgd(net, train, 1);
+        if(avg_loss == -1) avg_loss = loss;
+        avg_loss = avg_loss*.95 + loss*.05;
+        printf("%d, %.3f: %f, %f avg, %f rate, %lf seconds, %d images\n", get_current_batch(net), (float)(*net.seen)/N, loss, avg_loss, get_current_rate(net), sec(clock()-time), *net.seen);
+        if(*net.seen/N > epoch){
+            epoch = *net.seen/N;
+            char buff[256];
+            sprintf(buff, "%s/%s_%d.weights",backup_directory,base, epoch);
+            save_weights(net, buff);
+        }
+        if(get_current_batch(net)%100 == 0){
+            char buff[256];
+            sprintf(buff, "%s/%s.backup",backup_directory,base);
+            save_weights(net, buff);
+        }
+    }
+    char buff[256];
+    sprintf(buff, "%s/%s.weights", backup_directory, base);
+    save_weights(net, buff);
+
+    free_network(net);
+    free_ptrs((void**)labels, classes);
+    free(base);
+    free_data(train);
+}
+
+void test_cifar_multi(char *filename, char *weightfile)
+{
+    network net = parse_network_cfg(filename);
+    if(weightfile){
+        load_weights(&net, weightfile);
+    }
+    set_batch_network(&net, 1);
+    srand(time(0));
+
+    float avg_acc = 0;
+    data test = load_cifar10_data("data/cifar/cifar-10-batches-bin/test_batch.bin");
+
+    int i;
+    for(i = 0; i < test.X.rows; ++i){
+        image im = float_to_image(32, 32, 3, test.X.vals[i]);
+
+        float pred[10] = {0};
+
+        float *p = network_predict(net, im.data);
+        axpy_cpu(10, 1, p, 1, pred, 1);
+        flip_image(im);
+        p = network_predict(net, im.data);
+        axpy_cpu(10, 1, p, 1, pred, 1);
+
+        int index = max_index(pred, 10);
+        int class = max_index(test.y.vals[i], 10);
+        if(index == class) avg_acc += 1;
+        free_image(im);
+        printf("%4d: %.2f%%\n", i, 100.*avg_acc/(i+1));
+    }
+}
+
 void test_cifar(char *filename, char *weightfile)
 {
     network net = parse_network_cfg(filename);
@@ -79,6 +168,73 @@ void test_cifar(char *filename, char *weightfile)
     free_data(test);
 }
 
+void test_cifar_csv(char *filename, char *weightfile)
+{
+    network net = parse_network_cfg(filename);
+    if(weightfile){
+        load_weights(&net, weightfile);
+    }
+    srand(time(0));
+
+    data test = load_cifar10_data("data/cifar/cifar-10-batches-bin/test_batch.bin");
+
+    matrix pred = network_predict_data(net, test);
+
+    int i;
+    for(i = 0; i < test.X.rows; ++i){
+        image im = float_to_image(32, 32, 3, test.X.vals[i]);
+        flip_image(im);
+    }
+    matrix pred2 = network_predict_data(net, test);
+    scale_matrix(pred, .5);
+    scale_matrix(pred2, .5);
+    matrix_add_matrix(pred2, pred);
+
+    matrix_to_csv(pred);
+    fprintf(stderr, "Accuracy: %f\n", matrix_topk_accuracy(test.y, pred, 1));
+    free_data(test);
+}
+
+void test_cifar_csvtrain(char *filename, char *weightfile)
+{
+    network net = parse_network_cfg(filename);
+    if(weightfile){
+        load_weights(&net, weightfile);
+    }
+    srand(time(0));
+
+    data test = load_all_cifar10();
+
+    matrix pred = network_predict_data(net, test);
+
+    int i;
+    for(i = 0; i < test.X.rows; ++i){
+        image im = float_to_image(32, 32, 3, test.X.vals[i]);
+        flip_image(im);
+    }
+    matrix pred2 = network_predict_data(net, test);
+    scale_matrix(pred, .5);
+    scale_matrix(pred2, .5);
+    matrix_add_matrix(pred2, pred);
+
+    matrix_to_csv(pred);
+    fprintf(stderr, "Accuracy: %f\n", matrix_topk_accuracy(test.y, pred, 1));
+    free_data(test);
+}
+
+void eval_cifar_csv()
+{
+    data test = load_cifar10_data("data/cifar/cifar-10-batches-bin/test_batch.bin");
+
+    matrix pred = csv_to_matrix("results/combined.csv");
+    fprintf(stderr, "%d %d\n", pred.rows, pred.cols);
+
+    fprintf(stderr, "Accuracy: %f\n", matrix_topk_accuracy(test.y, pred, 1));
+    free_data(test);
+    free_matrix(pred);
+}
+
+
 void run_cifar(int argc, char **argv)
 {
     if(argc < 4){
@@ -89,7 +245,12 @@ void run_cifar(int argc, char **argv)
     char *cfg = argv[3];
     char *weights = (argc > 4) ? argv[4] : 0;
     if(0==strcmp(argv[2], "train")) train_cifar(cfg, weights);
+    else if(0==strcmp(argv[2], "distill")) train_cifar_distill(cfg, weights);
     else if(0==strcmp(argv[2], "test")) test_cifar(cfg, weights);
+    else if(0==strcmp(argv[2], "multi")) test_cifar_multi(cfg, weights);
+    else if(0==strcmp(argv[2], "csv")) test_cifar_csv(cfg, weights);
+    else if(0==strcmp(argv[2], "csvtrain")) test_cifar_csvtrain(cfg, weights);
+    else if(0==strcmp(argv[2], "eval")) eval_cifar_csv();
 }
 
 

src/classifier.c

@@ -3,6 +3,7 @@
 #include "parser.h"
 #include "option_list.h"
 #include "blas.h"
+#include <sys/time.h>
 
 #ifdef OPENCV
 #include "opencv2/highgui/highgui_c.h"
@@ -239,8 +240,8 @@ void validate_classifier_10(char *datacfg, char *filename, char *weightfile)
         }
         int w = net.w;
         int h = net.h;
-        image im = load_image_color(paths[i], w, h);
         int shift = 32;
+        image im = load_image_color(paths[i], w+shift, h+shift);
         image images[10];
         images[0] = crop_image(im, -shift, -shift, w, h);
         images[1] = crop_image(im, shift, -shift, w, h);
@@ -299,6 +300,7 @@ void validate_classifier_full(char *datacfg, char *filename, char *weightfile)
     float avg_topk = 0;
     int *indexes = calloc(topk, sizeof(int));
 
+    int size = net.w;
     for(i = 0; i < m; ++i){
         int class = -1;
         char *path = paths[i];
@@ -309,13 +311,15 @@ void validate_classifier_full(char *datacfg, char *filename, char *weightfile)
             }
         }
         image im = load_image_color(paths[i], 0, 0);
-        resize_network(&net, im.w, im.h);
+        image resized = resize_min(im, size);
+        resize_network(&net, resized.w, resized.h);
         //show_image(im, "orig");
         //show_image(crop, "cropped");
         //cvWaitKey(0);
-        float *pred = network_predict(net, im.data);
+        float *pred = network_predict(net, resized.data);
 
         free_image(im);
+        free_image(resized);
         top_k(pred, classes, topk, indexes);
 
         if(indexes[0] == class) avg_acc += 1;
@@ -406,7 +410,7 @@ void validate_classifier_multi(char *datacfg, char *filename, char *weightfile)
 
     char **labels = get_labels(label_list);
     list *plist = get_paths(valid_list);
-    int scales[] = {224, 256, 384, 480, 512};
+    int scales[] = {192, 224, 288, 320, 352};
     int nscales = sizeof(scales)/sizeof(scales[0]);
 
     char **paths = (char **)list_to_array(plist);
@@ -429,16 +433,8 @@ void validate_classifier_multi(char *datacfg, char *filename, char *weightfile)
         float *pred = calloc(classes, sizeof(float));
         image im = load_image_color(paths[i], 0, 0);
         for(j = 0; j < nscales; ++j){
-            int w, h;
-            if(im.w < im.h){
-                w = scales[j];
-                h = (im.h*w)/im.w;
-            } else {
-                h = scales[j];
-                w = (im.w * h) / im.h;
-            }
-            resize_network(&net, w, h);
-            image r = resize_image(im, w, h);
+            image r = resize_min(im, scales[j]);
+            resize_network(&net, r.w, r.h);
             float *p = network_predict(net, r.data);
             axpy_cpu(classes, 1, p, 1, pred, 1);
             flip_image(r);
@@ -577,6 +573,73 @@ void test_classifier(char *datacfg, char *cfgfile, char *weightfile, int target_
 }
 
 
+void demo_classifier(char *datacfg, char *cfgfile, char *weightfile, int cam_index, const char *filename)
+{
+#ifdef OPENCV
+    printf("Classifier Demo\n");
+    network net = parse_network_cfg(cfgfile);
+    if(weightfile){
+        load_weights(&net, weightfile);
+    }
+    set_batch_network(&net, 1);
+    list *options = read_data_cfg(datacfg);
+
+    srand(2222222);
+    CvCapture * cap;
+
+    if(filename){
+        cap = cvCaptureFromFile(filename);
+    }else{
+        cap = cvCaptureFromCAM(cam_index);
+    }
+
+    int top = option_find_int(options, "top", 1);
+
+    char *name_list = option_find_str(options, "names", 0);
+    char **names = get_labels(name_list);
+
+    int *indexes = calloc(top, sizeof(int));
+
+    if(!cap) error("Couldn't connect to webcam.\n");
+    cvNamedWindow("Classifier", CV_WINDOW_NORMAL); 
+    cvResizeWindow("Classifier", 512, 512);
+    float fps = 0;
+    int i;
+
+    while(1){
+        struct timeval tval_before, tval_after, tval_result;
+        gettimeofday(&tval_before, NULL);
+
+        image in = get_image_from_stream(cap);
+        image in_s = resize_image(in, net.w, net.h);
+        show_image(in, "Classifier");
+
+        float *predictions = network_predict(net, in_s.data);
+        top_predictions(net, top, indexes);
+
+        printf("\033[2J");
+        printf("\033[1;1H");
+        printf("\nFPS:%.0f\n",fps);
+
+        for(i = 0; i < top; ++i){
+            int index = indexes[i];
+            printf("%.1f%%: %s\n", predictions[index]*100, names[index]);
+        }
+
+        free_image(in_s);
+        free_image(in);
+
+        cvWaitKey(10);
+
+        gettimeofday(&tval_after, NULL);
+        timersub(&tval_after, &tval_before, &tval_result);
+        float curr = 1000000.f/((long int)tval_result.tv_usec);
+        fps = .9*fps + .1*curr;
+    }
+#endif
+}
+
+
 void run_classifier(int argc, char **argv)
 {
     if(argc < 4){
@@ -584,6 +647,7 @@ void run_classifier(int argc, char **argv)
         return;
     }
 
+    int cam_index = find_int_arg(argc, argv, "-c", 0);
     char *data = argv[3];
     char *cfg = argv[4];
     char *weights = (argc > 5) ? argv[5] : 0;
@@ -592,6 +656,7 @@ void run_classifier(int argc, char **argv)
     int layer = layer_s ? atoi(layer_s) : -1;
     if(0==strcmp(argv[2], "predict")) predict_classifier(data, cfg, weights, filename);
     else if(0==strcmp(argv[2], "train")) train_classifier(data, cfg, weights);
+    else if(0==strcmp(argv[2], "demo")) demo_classifier(data, cfg, weights, cam_index, filename);
     else if(0==strcmp(argv[2], "test")) test_classifier(data, cfg, weights, layer);
     else if(0==strcmp(argv[2], "valid")) validate_classifier(data, cfg, weights);
     else if(0==strcmp(argv[2], "valid10")) validate_classifier_10(data, cfg, weights);

src/coco_demo.c

@@ -71,7 +71,7 @@ void *detect_in_thread_coco(void *ptr)
 void demo_coco(char *cfgfile, char *weightfile, float thresh, int cam_index, const char *filename)
 {
     demo_thresh = thresh;
-    printf("YOLO demo\n");
+    printf("COCO demo\n");
     net = parse_network_cfg(cfgfile);
     if(weightfile){
         load_weights(&net, weightfile);
@@ -87,8 +87,8 @@ void demo_coco(char *cfgfile, char *weightfile, float thresh, int cam_index, con
     }
 
     if(!cap) error("Couldn't connect to webcam.\n");
-    cvNamedWindow("YOLO", CV_WINDOW_NORMAL); 
-    cvResizeWindow("YOLO", 512, 512);
+    cvNamedWindow("COCO", CV_WINDOW_NORMAL); 
+    cvResizeWindow("COCO", 512, 512);
 
     detection_layer l = net.layers[net.n-1];
     int j;
@@ -127,8 +127,8 @@ void demo_coco(char *cfgfile, char *weightfile, float thresh, int cam_index, con
         gettimeofday(&tval_before, NULL);
         if(pthread_create(&fetch_thread, 0, fetch_in_thread_coco, 0)) error("Thread creation failed");
         if(pthread_create(&detect_thread, 0, detect_in_thread_coco, 0)) error("Thread creation failed");
-        show_image(disp, "YOLO");
-        save_image(disp, "YOLO");
+        show_image(disp, "COCO");
+        //save_image(disp, "COCO");
         free_image(disp);
         cvWaitKey(10);
         pthread_join(fetch_thread, 0);

src/convolutional_kernels.cu

@@ -115,6 +115,46 @@ __global__ void backward_bias_kernel(float *bias_updates, float *delta, int batc
     }
 }
 
+__global__ void dot_kernel(float *output, float scale, int batch, int n, int size, float *delta)
+{
+    int index = (blockIdx.x + blockIdx.y*gridDim.x) * blockDim.x + threadIdx.x;
+    int f1 = index / n;
+    int f2 = index % n;
+    if (f2 <= f1) return;
+    
+    float sum = 0;
+    float norm1 = 0;
+    float norm2 = 0;
+    int b, i;
+    for(b = 0; b <  batch; ++b){
+        for(i = 0; i < size; ++i){
+            int i1 = b * size * n + f1 * size + i;
+            int i2 = b * size * n + f2 * size + i;
+            sum += output[i1] * output[i2];
+            norm1 += output[i1] * output[i1];
+            norm2 += output[i2] * output[i2];
+        }
+    }
+    norm1 = sqrt(norm1);
+    norm2 = sqrt(norm2);
+    float norm = norm1 * norm2;
+    sum = sum / norm;
+    for(b = 0; b <  batch; ++b){
+        for(i = 0; i < size; ++i){
+            int i1 = b * size * n + f1 * size + i;
+            int i2 = b * size * n + f2 * size + i;
+            delta[i1] += - scale * sum * output[i2] / norm;
+            delta[i2] += - scale * sum * output[i1] / norm;
+        }
+    }
+}
+
+void dot_error_gpu(layer l)
+{
+    dot_kernel<<<cuda_gridsize(l.n*l.n), BLOCK>>>(l.output_gpu, l.dot, l.batch, l.n, l.out_w * l.out_h, l.delta_gpu);
+    check_error(cudaPeekAtLastError());
+}
+
 void backward_bias_gpu(float *bias_updates, float *delta, int batch, int n, int size)
 {
     backward_bias_kernel<<<n, BLOCK>>>(bias_updates, delta, batch, n, size);
@@ -123,9 +163,9 @@ void backward_bias_gpu(float *bias_updates, float *delta, int batch, int n, int
 
 void swap_binary(convolutional_layer *l)
 {
-        float *swap = l->filters_gpu;
-        l->filters_gpu = l->binary_filters_gpu;
-        l->binary_filters_gpu = swap;
+    float *swap = l->filters_gpu;
+    l->filters_gpu = l->binary_filters_gpu;
+    l->binary_filters_gpu = swap;
 }
 
 void forward_convolutional_layer_gpu(convolutional_layer l, network_state state)
@@ -150,8 +190,8 @@ void forward_convolutional_layer_gpu(convolutional_layer l, network_state state)
         gemm_ongpu(0,0,m,n,k,1.,a,k,b,n,1.,c+i*m*n,n);
     }
 
-    if(l.batch_normalize){
-        if(state.train){
+    if (l.batch_normalize) {
+        if (state.train) {
             fast_mean_gpu(l.output_gpu, l.batch, l.n, l.out_h*l.out_w, l.mean_gpu);
             fast_variance_gpu(l.output_gpu, l.mean_gpu, l.batch, l.n, l.out_h*l.out_w, l.variance_gpu);
 
@@ -172,6 +212,7 @@ void forward_convolutional_layer_gpu(convolutional_layer l, network_state state)
     add_bias_gpu(l.output_gpu, l.biases_gpu, l.batch, l.n, n);
 
     activate_array_ongpu(l.output_gpu, m*n*l.batch, l.activation);
+    if(l.dot > 0) dot_error_gpu(l);
     if(l.binary) swap_binary(&l);
 }
 

src/darknet.c

@@ -24,6 +24,7 @@ extern void run_char_rnn(int argc, char **argv);
 extern void run_vid_rnn(int argc, char **argv);
 extern void run_tag(int argc, char **argv);
 extern void run_cifar(int argc, char **argv);
+extern void run_go(int argc, char **argv);
 
 void change_rate(char *filename, float scale, float add)
 {
@@ -171,13 +172,13 @@ void denormalize_net(char *cfgfile, char *weightfile, char *outfile)
 {
     gpu_index = -1;
     network net = parse_network_cfg(cfgfile);
-    if(weightfile){
+    if (weightfile) {
         load_weights(&net, weightfile);
     }
     int i;
-    for(i = 0; i < net.n; ++i){
+    for (i = 0; i < net.n; ++i) {
         layer l = net.layers[i];
-        if(l.type == CONVOLUTIONAL){
+        if (l.type == CONVOLUTIONAL && l.batch_normalize) {
             denormalize_convolutional_layer(l);
             net.layers[i].batch_normalize=0;
         }
@@ -228,6 +229,8 @@ int main(int argc, char **argv)
         run_yolo(argc, argv);
     } else if (0 == strcmp(argv[1], "cifar")){
         run_cifar(argc, argv);
+    } else if (0 == strcmp(argv[1], "go")){
+        run_go(argc, argv);
     } else if (0 == strcmp(argv[1], "rnn")){
         run_char_rnn(argc, argv);
     } else if (0 == strcmp(argv[1], "vid")){

src/data.c

@@ -95,6 +95,11 @@ matrix load_image_cropped_paths(char **paths, int n, int min, int max, int size)
         image crop = random_crop_image(im, min, max, size);
         int flip = rand_r(&data_seed)%2;
         if (flip) flip_image(crop);
+        /*
+        show_image(im, "orig");
+        show_image(crop, "crop");
+        cvWaitKey(0);
+        */
         free_image(im);
         X.vals[i] = crop.data;
         X.cols = crop.h*crop.w*crop.c;
@@ -863,6 +868,17 @@ void get_next_batch(data d, int n, int offset, float *X, float *y)
     }
 }
 
+void smooth_data(data d)
+{
+    int i, j;
+    int scale = 1. / d.y.cols;
+    int eps = .1;
+    for(i = 0; i < d.y.rows; ++i){
+        for(j = 0; j < d.y.cols; ++j){
+            d.y.vals[i][j] = eps * scale + (1-eps) * d.y.vals[i][j];
+        }
+    }
+}
 
 data load_all_cifar10()
 {
@@ -894,9 +910,55 @@ data load_all_cifar10()
     //normalize_data_rows(d);
     //translate_data_rows(d, -128);
     scale_data_rows(d, 1./255);
+   // smooth_data(d);
     return d;
 }
 
+data load_go(char *filename)
+{
+    FILE *fp = fopen(filename, "rb");
+    matrix X = make_matrix(128, 361);
+    matrix y = make_matrix(128, 361);
+    int row, col;
+
+    if(!fp) file_error(filename);
+    char *label;
+    int count = 0;
+    while((label = fgetl(fp))){
+        int i;
+        if(count == X.rows){
+            X = resize_matrix(X, count*2);
+            y = resize_matrix(y, count*2);
+        }
+        sscanf(label, "%d %d", &row, &col);
+        char *board = fgetl(fp);
+
+        int index = row*19 + col;
+        y.vals[count][index] = 1;
+
+        for(i = 0; i < 19*19; ++i){
+            float val = 0;
+            if(board[i] == '1') val = 1;
+            else if(board[i] == '2') val = -1;
+            X.vals[count][i] = val;
+        }
+        ++count;
+    }
+    X = resize_matrix(X, count);
+    y = resize_matrix(y, count);
+
+    data d;
+    d.shallow = 0;
+    d.X = X;
+    d.y = y;
+
+
+    fclose(fp);
+
+    return d;
+}
+
+
 void randomize_data(data d)
 {
     int i;
@@ -936,6 +998,29 @@ void normalize_data_rows(data d)
     }
 }
 
+data get_random_data(data d, int num)
+{
+    data r = {0};
+    r.shallow = 1;
+
+    r.X.rows = num;
+    r.y.rows = num;
+
+    r.X.cols = d.X.cols;
+    r.y.cols = d.y.cols;
+
+    r.X.vals = calloc(num, sizeof(float *));
+    r.y.vals = calloc(num, sizeof(float *));
+
+    int i;
+    for(i = 0; i < num; ++i){
+        int index = rand()%d.X.rows;
+        r.X.vals[i] = d.X.vals[index];
+        r.y.vals[i] = d.y.vals[index];
+    }
+    return r;
+}
+
 data *split_data(data d, int part, int total)
 {
     data *split = calloc(2, sizeof(data));

src/data.h

@@ -70,6 +70,7 @@ data load_data(char **paths, int n, int m, char **labels, int k, int w, int h);
 data load_data_detection(int n, char **paths, int m, int classes, int w, int h, int num_boxes, int background);
 data load_data_tag(char **paths, int n, int m, int k, int min, int max, int size);
 data load_data_augment(char **paths, int n, int m, char **labels, int k, int min, int max, int size);
+data load_go(char *filename);
 
 box_label *read_boxes(char *filename, int *n);
 data load_cifar10_data(char *filename);
@@ -80,6 +81,7 @@ data load_data_writing(char **paths, int n, int m, int w, int h, int out_w, int
 list *get_paths(char *filename);
 char **get_labels(char *filename);
 void get_random_batch(data d, int n, float *X, float *y);
+data get_random_data(data d, int num);
 void get_next_batch(data d, int n, int offset, float *X, float *y);
 data load_categorical_data_csv(char *filename, int target, int k);
 void normalize_data_rows(data d);

src/go.c

@@ -0,0 +1,249 @@
+#include "network.h"
+#include "utils.h"
+#include "parser.h"
+#include "option_list.h"
+#include "blas.h"
+
+#ifdef OPENCV
+#include "opencv2/highgui/highgui_c.h"
+#endif
+
+void train_go(char *cfgfile, char *weightfile)
+{
+    data_seed = time(0);
+    srand(time(0));
+    float avg_loss = -1;
+    char *base = basecfg(cfgfile);
+    printf("%s\n", base);
+    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);
+
+    char *backup_directory = "/home/pjreddie/backup/";
+
+    data train = load_go("/home/pjreddie/backup/go.train");
+    int N = train.X.rows;
+    int epoch = (*net.seen)/N;
+    while(get_current_batch(net) < net.max_batches || net.max_batches == 0){
+        clock_t time=clock();
+
+        data batch = get_random_data(train, net.batch);
+        int i;
+        for(i = 0; i < batch.X.rows; ++i){
+            int flip = rand()%2;
+            int rotate = rand()%4;
+            image in = float_to_image(19, 19, 1, batch.X.vals[i]);
+            image out = float_to_image(19, 19, 1, batch.y.vals[i]);
+            //show_image_normalized(in, "in");
+            //show_image_normalized(out, "out");
+            if(flip){
+                flip_image(in);
+                flip_image(out);
+            }
+            rotate_image_cw(in, rotate);
+            rotate_image_cw(out, rotate);
+            //show_image_normalized(in, "in2");
+            //show_image_normalized(out, "out2");
+            //cvWaitKey(0);
+        }
+        float loss = train_network(net, batch);
+        free_data(batch);
+        if(avg_loss == -1) avg_loss = loss;
+        avg_loss = avg_loss*.95 + loss*.05;
+        printf("%d, %.3f: %f, %f avg, %f rate, %lf seconds, %d images\n", get_current_batch(net), (float)(*net.seen)/N, loss, avg_loss, get_current_rate(net), sec(clock()-time), *net.seen);
+        if(*net.seen/N > epoch){
+            epoch = *net.seen/N;
+            char buff[256];
+            sprintf(buff, "%s/%s_%d.weights",backup_directory,base, epoch);
+            save_weights(net, buff);
+        }
+        if(get_current_batch(net)%100 == 0){
+            char buff[256];
+            sprintf(buff, "%s/%s.backup",backup_directory,base);
+            save_weights(net, buff);
+        }
+    }
+    char buff[256];
+    sprintf(buff, "%s/%s.weights", backup_directory, base);
+    save_weights(net, buff);
+
+    free_network(net);
+    free(base);
+    free_data(train);
+}
+
+void propagate_liberty(float *board, int *lib, int *visited, int row, int col, int num, int side)
+{
+    if (!num) return;
+    if (row < 0 || row > 18 || col < 0 || col > 18) return;
+    int index = row*19 + col;
+    if (board[index] != side) return;
+    if (visited[index]) return;
+    visited[index] = 1;
+    lib[index] += num;
+    propagate_liberty(board, lib, visited, row+1, col, num, side);
+    propagate_liberty(board, lib, visited, row-1, col, num, side);
+    propagate_liberty(board, lib, visited, row, col+1, num, side);
+    propagate_liberty(board, lib, visited, row, col-1, num, side);
+}
+
+int *calculate_liberties(float *board)
+{
+    int *lib = calloc(19*19, sizeof(int));
+    int visited[361];
+    int i, j;
+    for(j = 0; j < 19; ++j){
+        for(i = 0; i < 19; ++i){
+            memset(visited, 0, 19*19*sizeof(int));
+            int index = j*19 + i;
+            if(board[index]){
+                printf("%d %d\n", j, i);
+                int side = board[index];
+                int num = 0;
+                if (i > 0  && board[j*19 + i -  1] == 0) ++num;
+                if (i < 18 && board[j*19 + i +  1] == 0) ++num;
+                if (j > 0  && board[j*19 + i - 19] == 0) ++num;
+                if (j < 18 && board[j*19 + i + 19] == 0) ++num;
+                propagate_liberty(board, lib, visited, j, i, num, side);
+            }
+        }
+    }
+    return lib;
+}
+
+void update_board(float *board)
+{
+    int i;
+    int *l = calculate_liberties(board);
+    for(i = 0; i < 19*19; ++i){
+        if (board[i] && !l[i]) board[i] = 0;
+    }
+    free(l);
+}
+
+void print_board(float *board)
+{
+    int i,j;
+    printf("\n\n");
+    printf("   ");
+    for(i = 0; i < 19; ++i){
+        printf("%c ", 'A' + i + 1*(i > 7));
+    }
+    printf("\n");
+    for(j = 0; j < 19; ++j){
+        printf("%2d ", 19-j);
+        for(i = 0; i < 19; ++i){
+            int index = j*19 + i;
+            if(board[index] > 0) printf("\u25C9 ");
+            else if(board[index] < 0) printf("\u25EF ");
+            else printf("  ");
+        }
+        printf("\n");
+    }
+}
+
+void flip_board(float *board)
+{
+    int i;
+    for(i = 0; i < 19*19; ++i){
+        board[i] = -board[i];
+    }
+}
+
+void test_go(char *filename, char *weightfile)
+{
+    network net = parse_network_cfg(filename);
+    if(weightfile){
+        load_weights(&net, weightfile);
+    }
+    srand(time(0));
+    set_batch_network(&net, 1);
+    float *board = calloc(19*19, sizeof(float));
+    float *move = calloc(19*19, sizeof(float));
+    image bim = float_to_image(19, 19, 1, board);
+    while(1){
+        float *output = network_predict(net, board);
+        copy_cpu(19*19, output, 1, move, 1);
+        int i;
+        for(i = 1; i < 8; ++i){
+            rotate_image_cw(bim, i);
+            if(i >= 4) flip_image(bim);
+
+            float *output = network_predict(net, board);
+            image oim = float_to_image(19, 19, 1, output);
+
+            if(i >= 4) flip_image(oim);
+            rotate_image_cw(oim, -i);
+
+            int index = max_index(output, 19*19);
+            int row = index / 19;
+            int col = index % 19;
+            printf("Suggested: %c %d, %.2f%%\n", col + 'A' + 1*(col > 7), 19 - row, output[index]*100);
+
+            axpy_cpu(19*19, 1, output, 1, move, 1);
+
+            if(i >= 4) flip_image(bim);
+            rotate_image_cw(bim, -i);
+        }
+        scal_cpu(19*19, 1./8., move, 1);
+        for(i = 0; i < 19*19; ++i){
+            if(board[i]) move[i] = 0;
+        }
+
+        int indexes[3];
+        int row, col;
+        top_k(move, 19*19, 3, indexes);
+        print_board(board);
+        for(i = 0; i < 3; ++i){
+            int index = indexes[i];
+            row = index / 19;
+            col = index % 19;
+            printf("Suggested: %c %d, %.2f%%\n", col + 'A' + 1*(col > 7), 19 - row, move[index]*100);
+        }
+            int index = indexes[0];
+            row = index / 19;
+            col = index % 19;
+
+        printf("\u25C9 Enter move: ");
+        char c;
+        char *line = fgetl(stdin);
+        int num = sscanf(line, "%c %d", &c, &row);
+        if (c < 'A' || c > 'T'){
+            if (c == 'p'){
+                board[row*19 + col] = 1;
+            }else{
+                char g;
+                num = sscanf(line, "%c %c %d", &g, &c, &row);
+                row = 19 - row;
+                col = c - 'A';
+                if (col > 7) col -= 1;
+                board[row*19 + col] = 0;
+            }
+        } else {
+            row = 19 - row;
+            col = c - 'A';
+            if (col > 7) col -= 1;
+            if(num == 2) board[row*19 + col] = 1;
+        }
+        update_board(board);
+        flip_board(board);
+    }
+
+}
+
+void run_go(int argc, char **argv)
+{
+    if(argc < 4){
+        fprintf(stderr, "usage: %s %s [train/test/valid] [cfg] [weights (optional)]\n", argv[0], argv[1]);
+        return;
+    }
+
+    char *cfg = argv[3];
+    char *weights = (argc > 4) ? argv[4] : 0;
+    if(0==strcmp(argv[2], "train")) train_go(cfg, weights);
+    else if(0==strcmp(argv[2], "test")) test_go(cfg, weights);
+}
+
+

src/image.h

@@ -36,6 +36,7 @@ image resize_min(image im, int min);
 void translate_image(image m, float s);
 void normalize_image(image p);
 image rotate_image(image m, float rad);
+void rotate_image_cw(image im, int times);
 void embed_image(image source, image dest, int dx, int dy);
 void saturate_image(image im, float sat);
 void exposure_image(image im, float sat);
@@ -52,6 +53,7 @@ image collapse_images_horz(image *ims, int n);
 image collapse_images_vert(image *ims, int n);
 
 void show_image(image p, const char *name);
+void show_image_normalized(image im, const char *name);
 void save_image(image p, const char *name);
 void show_images(image *ims, int n, char *window);
 void show_image_layers(image p, char *name);

src/layer.h

@@ -56,6 +56,7 @@ struct layer{
     int binary;
     int steps;
     int hidden;
+    float dot;
     float angle;
     float jitter;
     float saturation;

src/matrix.c

@@ -33,6 +33,35 @@ float matrix_topk_accuracy(matrix truth, matrix guess, int k)
     return (float)correct/truth.rows;
 }
 
+void scale_matrix(matrix m, float scale)
+{
+    int i,j;
+    for(i = 0; i < m.rows; ++i){
+        for(j = 0; j < m.cols; ++j){
+            m.vals[i][j] *= scale;
+        }
+    }
+}
+
+matrix resize_matrix(matrix m, int size)
+{
+    int i;
+    if (m.rows == size) return m;
+    if (m.rows < size) {
+        m.vals = realloc(m.vals, size*sizeof(float*));
+        for (i = m.rows; i < size; ++i) {
+            m.vals[i] = calloc(m.cols, sizeof(float));
+        }
+    } else if (m.rows > size) {
+        for (i = size; i < m.rows; ++i) {
+            free(m.vals[i]);
+        }
+        m.vals = realloc(m.vals, size*sizeof(float*));
+    }
+    m.rows = size;
+    return m;
+}
+
 void matrix_add_matrix(matrix from, matrix to)
 {
     assert(from.rows == to.rows && from.cols == to.cols);
@@ -114,6 +143,19 @@ matrix csv_to_matrix(char *filename)
     return m;
 }
 
+void matrix_to_csv(matrix m)
+{
+    int i, j;
+
+    for(i = 0; i < m.rows; ++i){
+        for(j = 0; j < m.cols; ++j){
+            if(j > 0) printf(",");
+            printf("%.17g", m.vals[i][j]);
+        }
+        printf("\n");
+    }
+}
+
 void print_matrix(matrix m)
 {
     int i, j;

src/matrix.h

@@ -10,9 +10,12 @@ void free_matrix(matrix m);
 void print_matrix(matrix m);
 
 matrix csv_to_matrix(char *filename);
+void matrix_to_csv(matrix m);
 matrix hold_out_matrix(matrix *m, int n);
 float matrix_topk_accuracy(matrix truth, matrix guess, int k);
 void matrix_add_matrix(matrix from, matrix to);
+void scale_matrix(matrix m, float scale);
+matrix resize_matrix(matrix m, int size);
 
 float *pop_column(matrix *m, int c);
 

src/parser.c

@@ -160,6 +160,7 @@ convolutional_layer parse_convolutional(list *options, size_params params)
 
     convolutional_layer layer = make_convolutional_layer(batch,h,w,c,n,size,stride,pad,activation, batch_normalize, binary);
     layer.flipped = option_find_int_quiet(options, "flipped", 0);
+    layer.dot = option_find_float_quiet(options, "dot", 0);
 
     char *weights = option_find_str(options, "weights", 0);
     char *biases = option_find_str(options, "biases", 0);
@@ -850,7 +851,15 @@ void load_convolutional_weights(layer l, FILE *fp)
         fread(l.scales, sizeof(float), l.n, fp);
         fread(l.rolling_mean, sizeof(float), l.n, fp);
         fread(l.rolling_variance, sizeof(float), l.n, fp);
+        /*
+        int i;
+        for(i = 0; i < l.n; ++i){
+            if(l.rolling_mean[i] > 1 || l.rolling_mean[i] < -1 || l.rolling_variance[i] > 1 || l.rolling_variance[i] < -1)
+            printf("%f %f\n", l.rolling_mean[i], l.rolling_variance[i]);
+        }
+        */
     }
+    fflush(stdout);
     fread(l.filters, sizeof(float), num, fp);
     if (l.flipped) {
         transpose_matrix(l.filters, l.c*l.size*l.size, l.n);

src/tag.c

@@ -99,6 +99,7 @@ void test_tag(char *cfgfile, char *weightfile, char *filename)
     int indexes[10];
     char buff[256];
     char *input = buff;
+    int size = net.w;
     while(1){
         if(filename){
             strncpy(input, filename, 256);
@@ -109,11 +110,12 @@ void test_tag(char *cfgfile, char *weightfile, char *filename)
             if(!input) return;
             strtok(input, "\n");
         }
-        image im = load_image_color(input, net.w, net.h);
-        //resize_network(&net, im.w, im.h);
-        printf("%d %d\n", im.w, im.h);
+        image im = load_image_color(input, 0, 0);
+        image r = resize_min(im, size);
+        resize_network(&net, r.w, r.h);
+        printf("%d %d\n", r.w, r.h);
 
-        float *X = im.data;
+        float *X = r.data;
         time=clock();
         float *predictions = network_predict(net, X);
         top_predictions(net, 10, indexes);
@@ -122,6 +124,7 @@ void test_tag(char *cfgfile, char *weightfile, char *filename)
             int index = indexes[i];
             printf("%.1f%%: %s\n", predictions[index]*100, names[index]);
         }
+        free_image(r);
         free_image(im);
         if (filename) break;
     }

src/yolo.c

@@ -395,13 +395,7 @@ void demo_swag(char *cfgfile, char *weightfile, float thresh){}
 #endif
  */
 
-void demo_yolo(char *cfgfile, char *weightfile, float thresh, int cam_index);
-#ifndef GPU
-void demo_yolo(char *cfgfile, char *weightfile, float thresh, int cam_index)
-{
-    fprintf(stderr, "Darknet must be compiled with CUDA for YOLO demo.\n");
-}
-#endif
+void demo_yolo(char *cfgfile, char *weightfile, float thresh, int cam_index, char *filename);
 
 void run_yolo(int argc, char **argv)
 {
@@ -426,5 +420,5 @@ void run_yolo(int argc, char **argv)
     else if(0==strcmp(argv[2], "train")) train_yolo(cfg, weights);
     else if(0==strcmp(argv[2], "valid")) validate_yolo(cfg, weights);
     else if(0==strcmp(argv[2], "recall")) validate_yolo_recall(cfg, weights);
-    else if(0==strcmp(argv[2], "demo")) demo_yolo(cfg, weights, thresh, cam_index);
+    else if(0==strcmp(argv[2], "demo")) demo_yolo(cfg, weights, thresh, cam_index, filename);
 }

src/yolo_demo.c

@@ -1,8 +1,3 @@
-#include "cuda_runtime.h"
-#include "curand.h"
-#include "cublas_v2.h"
-
-extern "C" {
 #include "network.h"
 #include "detection_layer.h"
 #include "cost_layer.h"
@@ -11,16 +6,16 @@ extern "C" {
 #include "box.h"
 #include "image.h"
 #include <sys/time.h>
-}
 
 #ifdef OPENCV
 #include "opencv2/highgui/highgui.hpp"
 #include "opencv2/imgproc/imgproc.hpp"
-extern "C" image ipl_to_image(IplImage* src);
-extern "C" void convert_coco_detections(float *predictions, int classes, int num, int square, int side, int w, int h, float thresh, float **probs, box *boxes, int only_objectness);
+image ipl_to_image(IplImage* src);
+void convert_yolo_detections(float *predictions, int classes, int num, int square, int side, int w, int h, float thresh, float **probs, box *boxes, int only_objectness);
+void draw_yolo(image im, int num, float thresh, box *boxes, float **probs);
 
-extern "C" char *coco_classes[];
-extern "C" image coco_labels[];
+extern char *voc_names[];
+extern image voc_labels[];
 
 static float **probs;
 static box *boxes;
@@ -30,55 +25,36 @@ static image in_s ;
 static image det  ;
 static image det_s;
 static image disp ;
-static cv::VideoCapture cap;
+static CvCapture * cap;
 static float fps = 0;
 static float demo_thresh = 0;
 
-static const int frames = 3;
-static float *predictions[frames];
-static int demo_index = 0;
-static image images[frames];
-static float *avg;
-
-void *fetch_in_thread_coco(void *ptr)
+void *fetch_in_thread(void *ptr)
 {
-    cv::Mat frame_m;
-    cap >> frame_m;
-    IplImage frame = frame_m;
-    in = ipl_to_image(&frame);
-    rgbgr_image(in);
+    in = get_image_from_stream(cap);
     in_s = resize_image(in, net.w, net.h);
     return 0;
 }
 
-void *detect_in_thread_coco(void *ptr)
+void *detect_in_thread(void *ptr)
 {
     float nms = .4;
 
     detection_layer l = net.layers[net.n-1];
     float *X = det_s.data;
-    float *prediction = network_predict(net, X);
-
-    memcpy(predictions[demo_index], prediction, l.outputs*sizeof(float));
-    mean_arrays(predictions, frames, l.outputs, avg);
-
+    float *predictions = network_predict(net, X);
     free_image(det_s);
-    convert_coco_detections(avg, l.classes, l.n, l.sqrt, l.side, 1, 1, demo_thresh, probs, boxes, 0);
+    convert_yolo_detections(predictions, l.classes, l.n, l.sqrt, l.side, 1, 1, demo_thresh, probs, boxes, 0);
     if (nms > 0) do_nms(boxes, probs, l.side*l.side*l.n, l.classes, nms);
     printf("\033[2J");
     printf("\033[1;1H");
     printf("\nFPS:%.0f\n",fps);
     printf("Objects:\n\n");
-
-    images[demo_index] = det;
-    det = images[(demo_index + frames/2 + 1)%frames];
-    demo_index = (demo_index + 1)%frames;
-
-    draw_detections(det, l.side*l.side*l.n, demo_thresh, boxes, probs, coco_classes, coco_labels, 80);
+    draw_detections(det, l.side*l.side*l.n, demo_thresh, boxes, probs, voc_names, voc_labels, 20);
     return 0;
 }
 
-extern "C" void demo_coco(char *cfgfile, char *weightfile, float thresh, int cam_index, const char *filename)
+void demo_yolo(char *cfgfile, char *weightfile, float thresh, int cam_index, char *filename)
 {
     demo_thresh = thresh;
     printf("YOLO demo\n");
@@ -91,20 +67,18 @@ extern "C" void demo_coco(char *cfgfile, char *weightfile, float thresh, int cam
     srand(2222222);
 
     if(filename){
-        cap.open(filename);
+        cap = cvCaptureFromFile(filename);
     }else{
-        cap.open(cam_index);
+        cap = cvCaptureFromCAM(cam_index);
     }
 
-    if(!cap.isOpened()) error("Couldn't connect to webcam.\n");
+    if(!cap) error("Couldn't connect to webcam.\n");
+    cvNamedWindow("YOLO", CV_WINDOW_NORMAL); 
+    cvResizeWindow("YOLO", 512, 512);
 
     detection_layer l = net.layers[net.n-1];
     int j;
 
-    avg = (float *) calloc(l.outputs, sizeof(float));
-    for(j = 0; j < frames; ++j) predictions[j] = (float *) calloc(l.outputs, sizeof(float));
-    for(j = 0; j < frames; ++j) images[j] = make_image(1,1,3);
-
     boxes = (box *)calloc(l.side*l.side*l.n, sizeof(box));
     probs = (float **)calloc(l.side*l.side*l.n, sizeof(float *));
     for(j = 0; j < l.side*l.side*l.n; ++j) probs[j] = (float *)calloc(l.classes, sizeof(float *));
@@ -112,12 +86,12 @@ extern "C" void demo_coco(char *cfgfile, char *weightfile, float thresh, int cam
     pthread_t fetch_thread;
     pthread_t detect_thread;
 
-    fetch_in_thread_coco(0);
+    fetch_in_thread(0);
     det = in;
     det_s = in_s;
 
-    fetch_in_thread_coco(0);
-    detect_in_thread_coco(0);
+    fetch_in_thread(0);
+    detect_in_thread(0);
     disp = det;
     det = in;
     det_s = in_s;
@@ -125,8 +99,8 @@ extern "C" void demo_coco(char *cfgfile, char *weightfile, float thresh, int cam
     while(1){
         struct timeval tval_before, tval_after, tval_result;
         gettimeofday(&tval_before, NULL);
-        if(pthread_create(&fetch_thread, 0, fetch_in_thread_coco, 0)) error("Thread creation failed");
-        if(pthread_create(&detect_thread, 0, detect_in_thread_coco, 0)) error("Thread creation failed");
+        if(pthread_create(&fetch_thread, 0, fetch_in_thread, 0)) error("Thread creation failed");
+        if(pthread_create(&detect_thread, 0, detect_in_thread, 0)) error("Thread creation failed");
         show_image(disp, "YOLO");
         free_image(disp);
         cvWaitKey(1);
@@ -144,8 +118,8 @@ extern "C" void demo_coco(char *cfgfile, char *weightfile, float thresh, int cam
     }
 }
 #else
-extern "C" void demo_coco(char *cfgfile, char *weightfile, float thresh, int cam_index){
-    fprintf(stderr, "YOLO-COCO demo needs OpenCV for webcam images.\n");
+void demo_yolo(char *cfgfile, char *weightfile, float thresh, int cam_index, char *filename){
+    fprintf(stderr, "YOLO demo needs OpenCV for webcam images.\n");
 }
 #endif