diff --git a/Makefile b/Makefile
index c9b6ecac..528437d9 100644
--- a/Makefile
+++ b/Makefile
@@ -1,5 +1,5 @@
-GPU=0
-OPENCV=0
+GPU=1
+OPENCV=1
 DEBUG=0
 
 ARCH= --gpu-architecture=compute_20 --gpu-code=compute_20
@@ -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
+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
 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 coco_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 yolo_kernels.o
 endif
 
 OBJS = $(addprefix $(OBJDIR), $(OBJ))
diff --git a/src/blas.c b/src/blas.c
index d7948bb1..978f1ed7 100644
--- a/src/blas.c
+++ b/src/blas.c
@@ -115,13 +115,30 @@ void copy_cpu(int N, float *X, int INCX, float *Y, int INCY)
     for(i = 0; i < N; ++i) Y[i*INCY] = X[i*INCX];
 }
 
-void smooth_l1_cpu(int n, float *pred, float *truth, float *delta)
+void smooth_l1_cpu(int n, float *pred, float *truth, float *delta, float *error)
 {
     int i;
     for(i = 0; i < n; ++i){
         float diff = truth[i] - pred[i];
-        if(fabs(diff) > 1) delta[i] = diff;
-        else delta[i] = (diff > 0) ? 1 : -1;
+        float abs_val = fabs(diff);
+        if(abs_val < 1) {
+            error[i] = diff * diff;
+            delta[i] = diff;
+        }
+        else {
+            error[i] = 2*abs_val - 1;
+            delta[i] = (diff < 0) ? -1 : 1;
+        }
+    }
+}
+
+void l2_cpu(int n, float *pred, float *truth, float *delta, float *error)
+{
+    int i;
+    for(i = 0; i < n; ++i){
+        float diff = truth[i] - pred[i];
+        error[i] = diff * diff;
+        delta[i] = diff;
     }
 }
 
diff --git a/src/blas.h b/src/blas.h
index f5189e5e..030ef668 100644
--- a/src/blas.h
+++ b/src/blas.h
@@ -17,7 +17,6 @@ void fill_cpu(int N, float ALPHA, float * X, int INCX);
 float dot_cpu(int N, float *X, int INCX, float *Y, int INCY);
 void test_gpu_blas();
 void shortcut_cpu(int batch, int w1, int h1, int c1, float *add, int w2, int h2, int c2, float *out);
-void smooth_l1_cpu(int n, float *pred, float *truth, float *delta);
 
 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);
@@ -29,6 +28,9 @@ void mean_delta_cpu(float *delta, float *variance, int batch, int filters, int s
 void  variance_delta_cpu(float *x, float *delta, float *mean, float *variance, int batch, int filters, int spatial, float *variance_delta);
 void normalize_delta_cpu(float *x, float *mean, float *variance, float *mean_delta, float *variance_delta, int batch, int filters, int spatial, float *delta);
 
+void smooth_l1_cpu(int n, float *pred, float *truth, float *delta, float *error);
+void l2_cpu(int n, float *pred, float *truth, float *delta, float *error);
+
 #ifdef GPU
 void axpy_ongpu(int N, float ALPHA, float * X, int INCX, float * Y, int INCY);
 void axpy_ongpu_offset(int N, float ALPHA, float * X, int OFFX, int INCX, float * Y, int OFFY, int INCY);
@@ -53,9 +55,11 @@ void fast_variance_delta_gpu(float *x, float *delta, float *mean, float *varianc
 void fast_variance_gpu(float *x, float *mean, int batch, int filters, int spatial, float *variance);
 void fast_mean_gpu(float *x, int batch, int filters, int spatial, float *mean);
 void shortcut_gpu(int batch, int w1, int h1, int c1, float *add, int w2, int h2, int c2, float *out);
-void smooth_l1_gpu(int n, float *pred, float *truth, float *delta);
 void scale_bias_gpu(float *output, float *biases, int batch, int n, int size);
 void backward_scale_gpu(float *x_norm, float *delta, int batch, int n, int size, float *scale_updates);
 void scale_bias_gpu(float *output, float *biases, int batch, int n, int size);
+
+void smooth_l1_gpu(int n, float *pred, float *truth, float *delta, float *error);
+void l2_gpu(int n, float *pred, float *truth, float *delta, float *error);
 #endif
 #endif
diff --git a/src/blas_kernels.cu b/src/blas_kernels.cu
index 61db29f2..be0e553b 100644
--- a/src/blas_kernels.cu
+++ b/src/blas_kernels.cu
@@ -410,18 +410,41 @@ extern "C" void shortcut_gpu(int batch, int w1, int h1, int c1, float *add, int
     check_error(cudaPeekAtLastError());
 }
 
-__global__ void smooth_l1_kernel(int n, float *pred, float *truth, float *delta)
+__global__ void smooth_l1_kernel(int n, float *pred, float *truth, float *delta, float *error)
 {
     int i = (blockIdx.x + blockIdx.y*gridDim.x) * blockDim.x + threadIdx.x;
     if(i < n){
         float diff = truth[i] - pred[i];
-        if(abs(diff) > 1) delta[i] = diff;
-        else delta[i] = (diff > 0) ? 1 : -1;
+        float abs_val = abs(diff);
+        if(abs_val < 1) {
+            error[i] = diff * diff;
+            delta[i] = diff;
+        }
+        else {
+            error[i] = 2*abs_val - 1;
+            delta[i] = (diff < 0) ? -1 : 1;
+        }
     }
 }
 
-extern "C" void smooth_l1_gpu(int n, float *pred, float *truth, float *delta)
+extern "C" void smooth_l1_gpu(int n, float *pred, float *truth, float *delta, float *error)
 {
-    smooth_l1_kernel<<<cuda_gridsize(n), BLOCK>>>(n, pred, truth, delta);
+    smooth_l1_kernel<<<cuda_gridsize(n), BLOCK>>>(n, pred, truth, delta, error);
+    check_error(cudaPeekAtLastError());
+}
+
+__global__ void l2_kernel(int n, float *pred, float *truth, float *delta, float *error)
+{
+    int i = (blockIdx.x + blockIdx.y*gridDim.x) * blockDim.x + threadIdx.x;
+    if(i < n){
+        float diff = truth[i] - pred[i];
+        error[i] = diff * diff; //I know this is technically wrong, deal with it.
+        delta[i] = diff;
+    }
+}
+
+extern "C" void l2_gpu(int n, float *pred, float *truth, float *delta, float *error)
+{
+    l2_kernel<<<cuda_gridsize(n), BLOCK>>>(n, pred, truth, delta, error);
     check_error(cudaPeekAtLastError());
 }
diff --git a/src/cifar.c b/src/cifar.c
new file mode 100644
index 00000000..f8878770
--- /dev/null
+++ b/src/cifar.c
@@ -0,0 +1,95 @@
+#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_cifar(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();
+    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*.9 + loss*.1;
+        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(char *filename, char *weightfile)
+{
+    network net = parse_network_cfg(filename);
+    if(weightfile){
+        load_weights(&net, weightfile);
+    }
+    srand(time(0));
+
+    clock_t time;
+    float avg_acc = 0;
+    float avg_top5 = 0;
+    data test = load_cifar10_data("data/cifar/cifar-10-batches-bin/test_batch.bin");
+
+    time=clock();
+
+    float *acc = network_accuracies(net, test, 2);
+    avg_acc += acc[0];
+    avg_top5 += acc[1];
+    printf("top1: %f, %lf seconds, %d images\n", avg_acc, sec(clock()-time), test.X.rows);
+    free_data(test);
+}
+
+void run_cifar(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_cifar(cfg, weights);
+    else if(0==strcmp(argv[2], "test")) test_cifar(cfg, weights);
+}
+
+
diff --git a/src/classifier.c b/src/classifier.c
index 9924c371..fdbe5344 100644
--- a/src/classifier.c
+++ b/src/classifier.c
@@ -70,6 +70,11 @@ void train_classifier(char *datacfg, char *cfgfile, char *weightfile)
     load_args args = {0};
     args.w = net.w;
     args.h = net.h;
+
+    args.min = net.w;
+    args.max = net.max_crop;
+    args.size = net.w;
+
     args.paths = paths;
     args.classes = classes;
     args.n = imgs;
@@ -88,6 +93,16 @@ void train_classifier(char *datacfg, char *cfgfile, char *weightfile)
         load_thread = load_data_in_thread(args);
         printf("Loaded: %lf seconds\n", sec(clock()-time));
         time=clock();
+
+/*
+        int u;
+        for(u = 0; u < net.batch; ++u){
+            image im = float_to_image(net.w, net.h, 3, train.X.vals[u]);
+            show_image(im, "loaded");
+            cvWaitKey(0);
+        }
+        */
+
         float loss = train_network(net, train);
         if(avg_loss == -1) avg_loss = loss;
         avg_loss = avg_loss*.9 + loss*.1;
@@ -99,7 +114,7 @@ void train_classifier(char *datacfg, char *cfgfile, char *weightfile)
             sprintf(buff, "%s/%s_%d.weights",backup_directory,base, epoch);
             save_weights(net, buff);
         }
-        if(*net.seen%1000 == 0){
+        if(*net.seen%100 == 0){
             char buff[256];
             sprintf(buff, "%s/%s.backup",backup_directory,base);
             save_weights(net, buff);
@@ -152,13 +167,14 @@ void validate_classifier(char *datacfg, char *filename, char *weightfile)
     load_args args = {0};
     args.w = net.w;
     args.h = net.h;
+
     args.paths = paths;
     args.classes = classes;
     args.n = num;
     args.m = 0;
     args.labels = labels;
     args.d = &buffer;
-    args.type = CLASSIFICATION_DATA;
+    args.type = OLD_CLASSIFICATION_DATA;
 
     pthread_t load_thread = load_data_in_thread(args);
     for(i = 1; i <= splits; ++i){
@@ -221,19 +237,22 @@ void validate_classifier_10(char *datacfg, char *filename, char *weightfile)
                 break;
             }
         }
-        image im = load_image_color(paths[i], 256, 256);
+        int w = net.w;
+        int h = net.h;
+        image im = load_image_color(paths[i], w, h);
+        int shift = 32;
         image images[10];
-        images[0] = crop_image(im, -16, -16, 256, 256);
-        images[1] = crop_image(im, 16, -16, 256, 256);
-        images[2] = crop_image(im, 0, 0, 256, 256);
-        images[3] = crop_image(im, -16, 16, 256, 256);
-        images[4] = crop_image(im, 16, 16, 256, 256);
+        images[0] = crop_image(im, -shift, -shift, w, h);
+        images[1] = crop_image(im, shift, -shift, w, h);
+        images[2] = crop_image(im, 0, 0, w, h);
+        images[3] = crop_image(im, -shift, shift, w, h);
+        images[4] = crop_image(im, shift, shift, w, h);
         flip_image(im);
-        images[5] = crop_image(im, -16, -16, 256, 256);
-        images[6] = crop_image(im, 16, -16, 256, 256);
-        images[7] = crop_image(im, 0, 0, 256, 256);
-        images[8] = crop_image(im, -16, 16, 256, 256);
-        images[9] = crop_image(im, 16, 16, 256, 256);
+        images[5] = crop_image(im, -shift, -shift, w, h);
+        images[6] = crop_image(im, shift, -shift, w, h);
+        images[7] = crop_image(im, 0, 0, w, h);
+        images[8] = crop_image(im, -shift, shift, w, h);
+        images[9] = crop_image(im, shift, shift, w, h);
         float *pred = calloc(classes, sizeof(float));
         for(j = 0; j < 10; ++j){
             float *p = network_predict(net, images[j].data);
@@ -252,6 +271,122 @@ void validate_classifier_10(char *datacfg, char *filename, char *weightfile)
     }
 }
 
+void validate_classifier_full(char *datacfg, char *filename, char *weightfile)
+{
+    int i, j;
+    network net = parse_network_cfg(filename);
+    set_batch_network(&net, 1);
+    if(weightfile){
+        load_weights(&net, weightfile);
+    }
+    srand(time(0));
+
+    list *options = read_data_cfg(datacfg);
+
+    char *label_list = option_find_str(options, "labels", "data/labels.list");
+    char *valid_list = option_find_str(options, "valid", "data/train.list");
+    int classes = option_find_int(options, "classes", 2);
+    int topk = option_find_int(options, "top", 1);
+
+    char **labels = get_labels(label_list);
+    list *plist = get_paths(valid_list);
+
+    char **paths = (char **)list_to_array(plist);
+    int m = plist->size;
+    free_list(plist);
+
+    float avg_acc = 0;
+    float avg_topk = 0;
+    int *indexes = calloc(topk, sizeof(int));
+
+    for(i = 0; i < m; ++i){
+        int class = -1;
+        char *path = paths[i];
+        for(j = 0; j < classes; ++j){
+            if(strstr(path, labels[j])){
+                class = j;
+                break;
+            }
+        }
+        image im = load_image_color(paths[i], 0, 0);
+        resize_network(&net, im.w, im.h);
+        //show_image(im, "orig");
+        //show_image(crop, "cropped");
+        //cvWaitKey(0);
+        float *pred = network_predict(net, im.data);
+
+        free_image(im);
+        top_k(pred, classes, topk, indexes);
+
+        if(indexes[0] == class) avg_acc += 1;
+        for(j = 0; j < topk; ++j){
+            if(indexes[j] == class) avg_topk += 1;
+        }
+
+        printf("%d: top 1: %f, top %d: %f\n", i, avg_acc/(i+1), topk, avg_topk/(i+1));
+    }
+}
+
+
+void validate_classifier_single(char *datacfg, char *filename, char *weightfile)
+{
+    int i, j;
+    network net = parse_network_cfg(filename);
+    set_batch_network(&net, 1);
+    if(weightfile){
+        load_weights(&net, weightfile);
+    }
+    srand(time(0));
+
+    list *options = read_data_cfg(datacfg);
+
+    char *label_list = option_find_str(options, "labels", "data/labels.list");
+    char *valid_list = option_find_str(options, "valid", "data/train.list");
+    int classes = option_find_int(options, "classes", 2);
+    int topk = option_find_int(options, "top", 1);
+
+    char **labels = get_labels(label_list);
+    list *plist = get_paths(valid_list);
+
+    char **paths = (char **)list_to_array(plist);
+    int m = plist->size;
+    free_list(plist);
+
+    float avg_acc = 0;
+    float avg_topk = 0;
+    int *indexes = calloc(topk, sizeof(int));
+
+    for(i = 0; i < m; ++i){
+        int class = -1;
+        char *path = paths[i];
+        for(j = 0; j < classes; ++j){
+            if(strstr(path, labels[j])){
+                class = j;
+                break;
+            }
+        }
+        image im = load_image_color(paths[i], 0, 0);
+        image resized = resize_min(im, net.w);
+        image crop = crop_image(resized, (resized.w - net.w)/2, (resized.h - net.h)/2, net.w, net.h);
+        //show_image(im, "orig");
+        //show_image(crop, "cropped");
+        //cvWaitKey(0);
+        float *pred = network_predict(net, crop.data);
+
+        free_image(im);
+        free_image(resized);
+        free_image(crop);
+        top_k(pred, classes, topk, indexes);
+
+        if(indexes[0] == class) avg_acc += 1;
+        for(j = 0; j < topk; ++j){
+            if(indexes[j] == class) avg_topk += 1;
+        }
+
+        printf("%d: top 1: %f, top %d: %f\n", i, avg_acc/(i+1), topk, avg_topk/(i+1));
+    }
+}
+
 void validate_classifier_multi(char *datacfg, char *filename, char *weightfile)
 {
     int i, j;
@@ -271,7 +406,7 @@ 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, 640};
+    int scales[] = {224, 256, 384, 480, 512};
     int nscales = sizeof(scales)/sizeof(scales[0]);
 
     char **paths = (char **)list_to_array(plist);
@@ -402,7 +537,7 @@ void test_classifier(char *datacfg, char *cfgfile, char *weightfile, int target_
     args.m = 0;
     args.labels = 0;
     args.d = &buffer;
-    args.type = CLASSIFICATION_DATA;
+    args.type = OLD_CLASSIFICATION_DATA;
 
     pthread_t load_thread = load_data_in_thread(args);
     for(curr = net.batch; curr < m; curr += net.batch){
@@ -420,7 +555,7 @@ void test_classifier(char *datacfg, char *cfgfile, char *weightfile, int target_
 
         time=clock();
         matrix pred = network_predict_data(net, val);
-        
+
         int i, j;
         if (target_layer >= 0){
             //layer l = net.layers[target_layer];
@@ -461,6 +596,8 @@ void run_classifier(int argc, char **argv)
     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);
     else if(0==strcmp(argv[2], "validmulti")) validate_classifier_multi(data, cfg, weights);
+    else if(0==strcmp(argv[2], "validsingle")) validate_classifier_single(data, cfg, weights);
+    else if(0==strcmp(argv[2], "validfull")) validate_classifier_full(data, cfg, weights);
 }
 
 
diff --git a/src/coco.c b/src/coco.c
index 41c2d80c..947bef21 100644
--- a/src/coco.c
+++ b/src/coco.c
@@ -389,10 +389,10 @@ void test_coco(char *cfgfile, char *weightfile, char *filename, float thresh)
 void demo_coco(char *cfgfile, char *weightfile, float thresh, int cam_index, char *filename);
 static void demo(char *cfgfile, char *weightfile, float thresh, int cam_index, char* filename)
 {
-    #if defined(OPENCV) && defined(GPU)
+    #if defined(OPENCV)
         demo_coco(cfgfile, weightfile, thresh, cam_index, filename);
     #else
-        fprintf(stderr, "Need to compile with GPU and OpenCV for demo.\n");
+        fprintf(stderr, "Need to compile with OpenCV for demo.\n");
     #endif
 }
 
diff --git a/src/coco_demo.c b/src/coco_demo.c
new file mode 100644
index 00000000..4ba8eef2
--- /dev/null
+++ b/src/coco_demo.c
@@ -0,0 +1,152 @@
+#include "network.h"
+#include "detection_layer.h"
+#include "cost_layer.h"
+#include "utils.h"
+#include "parser.h"
+#include "box.h"
+#include "image.h"
+#include <sys/time.h>
+
+#define FRAMES 1
+
+#ifdef OPENCV
+#include "opencv2/highgui/highgui.hpp"
+#include "opencv2/imgproc/imgproc.hpp"
+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);
+
+extern char *coco_classes[];
+extern image coco_labels[];
+
+static float **probs;
+static box *boxes;
+static network net;
+static image in   ;
+static image in_s ;
+static image det  ;
+static image det_s;
+static image disp ;
+static CvCapture * cap;
+static float fps = 0;
+static float demo_thresh = 0;
+
+static float *predictions[FRAMES];
+static int demo_index = 0;
+static image images[FRAMES];
+static float *avg;
+
+void *fetch_in_thread_coco(void *ptr)
+{
+    in = get_image_from_stream(cap);
+    in_s = resize_image(in, net.w, net.h);
+    return 0;
+}
+
+void *detect_in_thread_coco(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);
+
+    free_image(det_s);
+    convert_coco_detections(avg, 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);
+    return 0;
+}
+
+void demo_coco(char *cfgfile, char *weightfile, float thresh, int cam_index, const char *filename)
+{
+    demo_thresh = thresh;
+    printf("YOLO demo\n");
+    net = parse_network_cfg(cfgfile);
+    if(weightfile){
+        load_weights(&net, weightfile);
+    }
+    set_batch_network(&net, 1);
+
+    srand(2222222);
+
+    if(filename){
+        cap = cvCaptureFromFile(filename);
+    }else{
+        cap = cvCaptureFromCAM(cam_index);
+    }
+
+    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 *));
+
+    pthread_t fetch_thread;
+    pthread_t detect_thread;
+
+    fetch_in_thread_coco(0);
+    det = in;
+    det_s = in_s;
+
+    fetch_in_thread_coco(0);
+    detect_in_thread_coco(0);
+    disp = det;
+    det = in;
+    det_s = in_s;
+
+    for(j = 0; j < FRAMES/2; ++j){
+        fetch_in_thread_coco(0);
+        detect_in_thread_coco(0);
+        disp = det;
+        det = in;
+        det_s = in_s;
+    }
+
+    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");
+        show_image(disp, "YOLO");
+        save_image(disp, "YOLO");
+        free_image(disp);
+        cvWaitKey(10);
+        pthread_join(fetch_thread, 0);
+        pthread_join(detect_thread, 0);
+
+        disp  = det;
+        det   = in;
+        det_s = in_s;
+
+        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;
+    }
+}
+#else
+void demo_coco(char *cfgfile, char *weightfile, float thresh, int cam_index){
+    fprintf(stderr, "YOLO-COCO demo needs OpenCV for webcam images.\n");
+}
+#endif
+
diff --git a/src/convolutional_kernels.cu b/src/convolutional_kernels.cu
index 4fdc1a1e..4f474d68 100644
--- a/src/convolutional_kernels.cu
+++ b/src/convolutional_kernels.cu
@@ -121,11 +121,11 @@ void backward_bias_gpu(float *bias_updates, float *delta, int batch, int n, int
     check_error(cudaPeekAtLastError());
 }
 
-void swap_binary(convolutional_layer l)
+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)
@@ -139,7 +139,7 @@ void forward_convolutional_layer_gpu(convolutional_layer l, network_state state)
     fill_ongpu(l.outputs*l.batch, 0, l.output_gpu, 1);
     if(l.binary){
         binarize_filters_gpu(l.filters_gpu, l.n, l.c*l.size*l.size, l.binary_filters_gpu);
-        swap_binary(l);
+        swap_binary(&l);
     }
 
     for(i = 0; i < l.batch; ++i){
@@ -172,7 +172,7 @@ 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.binary) swap_binary(l);
+    if(l.binary) swap_binary(&l);
 }
 
 void backward_convolutional_layer_gpu(convolutional_layer l, network_state state)
@@ -206,7 +206,7 @@ void backward_convolutional_layer_gpu(convolutional_layer l, network_state state
         gemm_ongpu(0,1,m,n,k,1,a + i*m*k,k,b,k,1,c,n);
 
         if(state.delta){
-            if(l.binary) swap_binary(l);
+            if(l.binary) swap_binary(&l);
             float * a = l.filters_gpu;
             float * b = l.delta_gpu;
             float * c = l.col_image_gpu;
@@ -214,7 +214,7 @@ void backward_convolutional_layer_gpu(convolutional_layer l, network_state state
             gemm_ongpu(1,0,n,k,m,1,a,n,b + i*k*m,k,0,c,k);
 
             col2im_ongpu(l.col_image_gpu, l.c,  l.h,  l.w,  l.size,  l.stride, l.pad, state.delta + i*l.c*l.h*l.w);
-            if(l.binary) swap_binary(l);
+            if(l.binary) swap_binary(&l);
         }
     }
 }
diff --git a/src/cost_layer.c b/src/cost_layer.c
index 39ae8096..fdba7771 100644
--- a/src/cost_layer.c
+++ b/src/cost_layer.c
@@ -41,9 +41,11 @@ cost_layer make_cost_layer(int batch, int inputs, COST_TYPE cost_type, float sca
     l.outputs = inputs;
     l.cost_type = cost_type;
     l.delta = calloc(inputs*batch, sizeof(float));
-    l.output = calloc(1, sizeof(float));
+    l.output = calloc(inputs*batch, sizeof(float));
+    l.cost = calloc(1, sizeof(float));
     #ifdef GPU
-    l.delta_gpu = cuda_make_array(l.delta, inputs*batch);
+    l.delta_gpu = cuda_make_array(l.output, inputs*batch);
+    l.output_gpu = cuda_make_array(l.delta, inputs*batch);
     #endif
     return l;
 }
@@ -53,9 +55,12 @@ void resize_cost_layer(cost_layer *l, int inputs)
     l->inputs = inputs;
     l->outputs = inputs;
     l->delta = realloc(l->delta, inputs*l->batch*sizeof(float));
+    l->output = realloc(l->output, inputs*l->batch*sizeof(float));
 #ifdef GPU
     cuda_free(l->delta_gpu);
+    cuda_free(l->output_gpu);
     l->delta_gpu = cuda_make_array(l->delta, inputs*l->batch);
+    l->output_gpu = cuda_make_array(l->output, inputs*l->batch);
 #endif
 }
 
@@ -69,13 +74,11 @@ void forward_cost_layer(cost_layer l, network_state state)
         }
     }
     if(l.cost_type == SMOOTH){
-        smooth_l1_cpu(l.batch*l.inputs, state.input, state.truth, l.delta);
+        smooth_l1_cpu(l.batch*l.inputs, state.input, state.truth, l.delta, l.output);
     } else {
-        copy_cpu(l.batch*l.inputs, state.truth, 1, l.delta, 1);
-        axpy_cpu(l.batch*l.inputs, -1, state.input, 1, l.delta, 1);
+        l2_cpu(l.batch*l.inputs, state.input, state.truth, l.delta, l.output);
     }
-    *(l.output) = dot_cpu(l.batch*l.inputs, l.delta, 1, l.delta, 1);
-    //printf("cost: %f\n", *l.output);
+    l.cost[0] = sum_array(l.output, l.batch*l.inputs);
 }
 
 void backward_cost_layer(const cost_layer l, network_state state)
@@ -103,14 +106,13 @@ void forward_cost_layer_gpu(cost_layer l, network_state state)
     }
 
     if(l.cost_type == SMOOTH){
-        smooth_l1_gpu(l.batch*l.inputs, state.input, state.truth, l.delta_gpu);
+        smooth_l1_gpu(l.batch*l.inputs, state.input, state.truth, l.delta_gpu, l.output_gpu);
     } else {
-        copy_ongpu(l.batch*l.inputs, state.truth, 1, l.delta_gpu, 1);
-        axpy_ongpu(l.batch*l.inputs, -1, state.input, 1, l.delta_gpu, 1);
+        l2_gpu(l.batch*l.inputs, state.input, state.truth, l.delta_gpu, l.output_gpu);
     }
 
-    cuda_pull_array(l.delta_gpu, l.delta, l.batch*l.inputs);
-    *(l.output) = dot_cpu(l.batch*l.inputs, l.delta, 1, l.delta, 1);
+    cuda_pull_array(l.output_gpu, l.output, l.batch*l.inputs);
+    l.cost[0] = sum_array(l.output, l.batch*l.inputs);
 }
 
 void backward_cost_layer_gpu(const cost_layer l, network_state state)
diff --git a/src/crnn_layer.c b/src/crnn_layer.c
new file mode 100644
index 00000000..ed65665f
--- /dev/null
+++ b/src/crnn_layer.c
@@ -0,0 +1,277 @@
+#include "crnn_layer.h"
+#include "convolutional_layer.h"
+#include "utils.h"
+#include "cuda.h"
+#include "blas.h"
+#include "gemm.h"
+
+#include <math.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+
+static void increment_layer(layer *l, int steps)
+{
+    int num = l->outputs*l->batch*steps;
+    l->output += num;
+    l->delta += num;
+    l->x += num;
+    l->x_norm += num;
+
+#ifdef GPU
+    l->output_gpu += num;
+    l->delta_gpu += num;
+    l->x_gpu += num;
+    l->x_norm_gpu += num;
+#endif
+}
+
+layer make_crnn_layer(int batch, int h, int w, int c, int hidden_filters, int output_filters, int steps, ACTIVATION activation, int batch_normalize)
+{
+    fprintf(stderr, "CRNN Layer: %d x %d x %d image, %d filters\n", h,w,c,output_filters);
+    batch = batch / steps;
+    layer l = {0};
+    l.batch = batch;
+    l.type = CRNN;
+    l.steps = steps;
+    l.h = h;
+    l.w = w;
+    l.c = c;
+    l.out_h = h;
+    l.out_w = w;
+    l.out_c = output_filters;
+    l.inputs = h*w*c;
+    l.hidden = h * w * hidden_filters;
+    l.outputs = l.out_h * l.out_w * l.out_c;
+
+    l.state = calloc(l.hidden*batch*(steps+1), sizeof(float));
+
+    l.input_layer = malloc(sizeof(layer));
+    fprintf(stderr, "\t\t");
+    *(l.input_layer) = make_convolutional_layer(batch*steps, h, w, c, hidden_filters, 3, 1, 1,  activation, batch_normalize, 0);
+    l.input_layer->batch = batch;
+
+    l.self_layer = malloc(sizeof(layer));
+    fprintf(stderr, "\t\t");
+    *(l.self_layer) = make_convolutional_layer(batch*steps, h, w, hidden_filters, hidden_filters, 3, 1, 1,  activation, batch_normalize, 0);
+    l.self_layer->batch = batch;
+
+    l.output_layer = malloc(sizeof(layer));
+    fprintf(stderr, "\t\t");
+    *(l.output_layer) = make_convolutional_layer(batch*steps, h, w, hidden_filters, output_filters, 3, 1, 1,  activation, batch_normalize, 0);
+    l.output_layer->batch = batch;
+
+    l.output = l.output_layer->output;
+    l.delta = l.output_layer->delta;
+
+#ifdef GPU
+    l.state_gpu = cuda_make_array(l.state, l.hidden*batch*(steps+1));
+    l.output_gpu = l.output_layer->output_gpu;
+    l.delta_gpu = l.output_layer->delta_gpu;
+#endif
+
+    return l;
+}
+
+void update_crnn_layer(layer l, int batch, float learning_rate, float momentum, float decay)
+{
+    update_convolutional_layer(*(l.input_layer), batch, learning_rate, momentum, decay);
+    update_convolutional_layer(*(l.self_layer), batch, learning_rate, momentum, decay);
+    update_convolutional_layer(*(l.output_layer), batch, learning_rate, momentum, decay);
+}
+
+void forward_crnn_layer(layer l, network_state state)
+{
+    network_state s = {0};
+    s.train = state.train;
+    int i;
+    layer input_layer = *(l.input_layer);
+    layer self_layer = *(l.self_layer);
+    layer output_layer = *(l.output_layer);
+
+    fill_cpu(l.outputs * l.batch * l.steps, 0, output_layer.delta, 1);
+    fill_cpu(l.hidden * l.batch * l.steps, 0, self_layer.delta, 1);
+    fill_cpu(l.hidden * l.batch * l.steps, 0, input_layer.delta, 1);
+    if(state.train) fill_cpu(l.hidden * l.batch, 0, l.state, 1);
+
+    for (i = 0; i < l.steps; ++i) {
+        s.input = state.input;
+        forward_convolutional_layer(input_layer, s);
+
+        s.input = l.state;
+        forward_convolutional_layer(self_layer, s);
+
+        float *old_state = l.state;
+        if(state.train) l.state += l.hidden*l.batch;
+        if(l.shortcut){
+            copy_cpu(l.hidden * l.batch, old_state, 1, l.state, 1);
+        }else{
+            fill_cpu(l.hidden * l.batch, 0, l.state, 1);
+        }
+        axpy_cpu(l.hidden * l.batch, 1, input_layer.output, 1, l.state, 1);
+        axpy_cpu(l.hidden * l.batch, 1, self_layer.output, 1, l.state, 1);
+
+        s.input = l.state;
+        forward_convolutional_layer(output_layer, s);
+
+        state.input += l.inputs*l.batch;
+        increment_layer(&input_layer, 1);
+        increment_layer(&self_layer, 1);
+        increment_layer(&output_layer, 1);
+    }
+}
+
+void backward_crnn_layer(layer l, network_state state)
+{
+    network_state s = {0};
+    s.train = state.train;
+    int i;
+    layer input_layer = *(l.input_layer);
+    layer self_layer = *(l.self_layer);
+    layer output_layer = *(l.output_layer);
+
+    increment_layer(&input_layer, l.steps-1);
+    increment_layer(&self_layer, l.steps-1);
+    increment_layer(&output_layer, l.steps-1);
+
+    l.state += l.hidden*l.batch*l.steps;
+    for (i = l.steps-1; i >= 0; --i) {
+        copy_cpu(l.hidden * l.batch, input_layer.output, 1, l.state, 1);
+        axpy_cpu(l.hidden * l.batch, 1, self_layer.output, 1, l.state, 1);
+
+        s.input = l.state;
+        s.delta = self_layer.delta;
+        backward_convolutional_layer(output_layer, s);
+
+        l.state -= l.hidden*l.batch;
+        /*
+           if(i > 0){
+           copy_cpu(l.hidden * l.batch, input_layer.output - l.hidden*l.batch, 1, l.state, 1);
+           axpy_cpu(l.hidden * l.batch, 1, self_layer.output - l.hidden*l.batch, 1, l.state, 1);
+           }else{
+           fill_cpu(l.hidden * l.batch, 0, l.state, 1);
+           }
+         */
+
+        s.input = l.state;
+        s.delta = self_layer.delta - l.hidden*l.batch;
+        if (i == 0) s.delta = 0;
+        backward_convolutional_layer(self_layer, s);
+
+        copy_cpu(l.hidden*l.batch, self_layer.delta, 1, input_layer.delta, 1);
+        if (i > 0 && l.shortcut) axpy_cpu(l.hidden*l.batch, 1, self_layer.delta, 1, self_layer.delta - l.hidden*l.batch, 1);
+        s.input = state.input + i*l.inputs*l.batch;
+        if(state.delta) s.delta = state.delta + i*l.inputs*l.batch;
+        else s.delta = 0;
+        backward_convolutional_layer(input_layer, s);
+
+        increment_layer(&input_layer, -1);
+        increment_layer(&self_layer, -1);
+        increment_layer(&output_layer, -1);
+    }
+}
+
+#ifdef GPU
+
+void pull_crnn_layer(layer l)
+{
+    pull_convolutional_layer(*(l.input_layer));
+    pull_convolutional_layer(*(l.self_layer));
+    pull_convolutional_layer(*(l.output_layer));
+}
+
+void push_crnn_layer(layer l)
+{
+    push_convolutional_layer(*(l.input_layer));
+    push_convolutional_layer(*(l.self_layer));
+    push_convolutional_layer(*(l.output_layer));
+}
+
+void update_crnn_layer_gpu(layer l, int batch, float learning_rate, float momentum, float decay)
+{
+    update_convolutional_layer_gpu(*(l.input_layer), batch, learning_rate, momentum, decay);
+    update_convolutional_layer_gpu(*(l.self_layer), batch, learning_rate, momentum, decay);
+    update_convolutional_layer_gpu(*(l.output_layer), batch, learning_rate, momentum, decay);
+}
+
+void forward_crnn_layer_gpu(layer l, network_state state)
+{
+    network_state s = {0};
+    s.train = state.train;
+    int i;
+    layer input_layer = *(l.input_layer);
+    layer self_layer = *(l.self_layer);
+    layer output_layer = *(l.output_layer);
+
+    fill_ongpu(l.outputs * l.batch * l.steps, 0, output_layer.delta_gpu, 1);
+    fill_ongpu(l.hidden * l.batch * l.steps, 0, self_layer.delta_gpu, 1);
+    fill_ongpu(l.hidden * l.batch * l.steps, 0, input_layer.delta_gpu, 1);
+    if(state.train) fill_ongpu(l.hidden * l.batch, 0, l.state_gpu, 1);
+
+    for (i = 0; i < l.steps; ++i) {
+        s.input = state.input;
+        forward_convolutional_layer_gpu(input_layer, s);
+
+        s.input = l.state_gpu;
+        forward_convolutional_layer_gpu(self_layer, s);
+
+        float *old_state = l.state_gpu;
+        if(state.train) l.state_gpu += l.hidden*l.batch;
+        if(l.shortcut){
+            copy_ongpu(l.hidden * l.batch, old_state, 1, l.state_gpu, 1);
+        }else{
+            fill_ongpu(l.hidden * l.batch, 0, l.state_gpu, 1);
+        }
+        axpy_ongpu(l.hidden * l.batch, 1, input_layer.output_gpu, 1, l.state_gpu, 1);
+        axpy_ongpu(l.hidden * l.batch, 1, self_layer.output_gpu, 1, l.state_gpu, 1);
+
+        s.input = l.state_gpu;
+        forward_convolutional_layer_gpu(output_layer, s);
+
+        state.input += l.inputs*l.batch;
+        increment_layer(&input_layer, 1);
+        increment_layer(&self_layer, 1);
+        increment_layer(&output_layer, 1);
+    }
+}
+
+void backward_crnn_layer_gpu(layer l, network_state state)
+{
+    network_state s = {0};
+    s.train = state.train;
+    int i;
+    layer input_layer = *(l.input_layer);
+    layer self_layer = *(l.self_layer);
+    layer output_layer = *(l.output_layer);
+    increment_layer(&input_layer,  l.steps - 1);
+    increment_layer(&self_layer,   l.steps - 1);
+    increment_layer(&output_layer, l.steps - 1);
+    l.state_gpu += l.hidden*l.batch*l.steps;
+    for (i = l.steps-1; i >= 0; --i) {
+        copy_ongpu(l.hidden * l.batch, input_layer.output_gpu, 1, l.state_gpu, 1);
+        axpy_ongpu(l.hidden * l.batch, 1, self_layer.output_gpu, 1, l.state_gpu, 1);
+
+        s.input = l.state_gpu;
+        s.delta = self_layer.delta_gpu;
+        backward_convolutional_layer_gpu(output_layer, s);
+
+        l.state_gpu -= l.hidden*l.batch;
+
+        s.input = l.state_gpu;
+        s.delta = self_layer.delta_gpu - l.hidden*l.batch;
+        if (i == 0) s.delta = 0;
+        backward_convolutional_layer_gpu(self_layer, s);
+
+        copy_ongpu(l.hidden*l.batch, self_layer.delta_gpu, 1, input_layer.delta_gpu, 1);
+        if (i > 0 && l.shortcut) axpy_ongpu(l.hidden*l.batch, 1, self_layer.delta_gpu, 1, self_layer.delta_gpu - l.hidden*l.batch, 1);
+        s.input = state.input + i*l.inputs*l.batch;
+        if(state.delta) s.delta = state.delta + i*l.inputs*l.batch;
+        else s.delta = 0;
+        backward_convolutional_layer_gpu(input_layer, s);
+
+        increment_layer(&input_layer,  -1);
+        increment_layer(&self_layer,   -1);
+        increment_layer(&output_layer, -1);
+    }
+}
+#endif
diff --git a/src/crnn_layer.h b/src/crnn_layer.h
new file mode 100644
index 00000000..0da942ee
--- /dev/null
+++ b/src/crnn_layer.h
@@ -0,0 +1,24 @@
+
+#ifndef CRNN_LAYER_H
+#define CRNN_LAYER_H
+
+#include "activations.h"
+#include "layer.h"
+#include "network.h"
+
+layer make_crnn_layer(int batch, int h, int w, int c, int hidden_filters, int output_filters, int steps, ACTIVATION activation, int batch_normalize);
+
+void forward_crnn_layer(layer l, network_state state);
+void backward_crnn_layer(layer l, network_state state);
+void update_crnn_layer(layer l, int batch, float learning_rate, float momentum, float decay);
+
+#ifdef GPU
+void forward_crnn_layer_gpu(layer l, network_state state);
+void backward_crnn_layer_gpu(layer l, network_state state);
+void update_crnn_layer_gpu(layer l, int batch, float learning_rate, float momentum, float decay);
+void push_crnn_layer(layer l);
+void pull_crnn_layer(layer l);
+#endif
+
+#endif
+
diff --git a/src/darknet.c b/src/darknet.c
index c4006cee..5722729a 100644
--- a/src/darknet.c
+++ b/src/darknet.c
@@ -21,6 +21,9 @@ extern void run_dice(int argc, char **argv);
 extern void run_compare(int argc, char **argv);
 extern void run_classifier(int argc, char **argv);
 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);
 
 void change_rate(char *filename, float scale, float add)
 {
@@ -223,12 +226,18 @@ int main(int argc, char **argv)
         average(argc, argv);
     } else if (0 == strcmp(argv[1], "yolo")){
         run_yolo(argc, argv);
+    } else if (0 == strcmp(argv[1], "cifar")){
+        run_cifar(argc, argv);
     } else if (0 == strcmp(argv[1], "rnn")){
         run_char_rnn(argc, argv);
+    } else if (0 == strcmp(argv[1], "vid")){
+        run_vid_rnn(argc, argv);
     } else if (0 == strcmp(argv[1], "coco")){
         run_coco(argc, argv);
     } else if (0 == strcmp(argv[1], "classifier")){
         run_classifier(argc, argv);
+    } else if (0 == strcmp(argv[1], "tag")){
+        run_tag(argc, argv);
     } else if (0 == strcmp(argv[1], "compare")){
         run_compare(argc, argv);
     } else if (0 == strcmp(argv[1], "dice")){
diff --git a/src/data.c b/src/data.c
index 88c89917..c429a73a 100644
--- a/src/data.c
+++ b/src/data.c
@@ -82,6 +82,27 @@ matrix load_image_paths(char **paths, int n, int w, int h)
     return X;
 }
 
+matrix load_image_cropped_paths(char **paths, int n, int min, int max, int size)
+{
+    int i;
+    matrix X;
+    X.rows = n;
+    X.vals = calloc(X.rows, sizeof(float*));
+    X.cols = 0;
+
+    for(i = 0; i < n; ++i){
+        image im = load_image_color(paths[i], 0, 0);
+        image crop = random_crop_image(im, min, max, size);
+        int flip = rand_r(&data_seed)%2;
+        if (flip) flip_image(crop);
+        free_image(im);
+        X.vals[i] = crop.data;
+        X.cols = crop.h*crop.w*crop.c;
+    }
+    return X;
+}
+
+
 box_label *read_boxes(char *filename, int *n)
 {
     box_label *boxes = calloc(1, sizeof(box_label));
@@ -386,6 +407,33 @@ matrix load_labels_paths(char **paths, int n, char **labels, int k)
     return y;
 }
 
+matrix load_tags_paths(char **paths, int n, int k)
+{
+    matrix y = make_matrix(n, k);
+    int i;
+    int count = 0;
+    for(i = 0; i < n; ++i){
+        char *label = find_replace(paths[i], "imgs", "labels");
+        label = find_replace(label, "_iconl.jpeg", ".txt");
+        FILE *file = fopen(label, "r");
+        if(!file){
+            label = find_replace(label, "labels", "labels2");
+            file = fopen(label, "r");
+            if(!file) continue;
+        }
+        ++count;
+        int tag;
+        while(fscanf(file, "%d", &tag) == 1){
+            if(tag < k){
+                y.vals[i][tag] = 1;
+            }
+        }
+        fclose(file);
+    }
+    printf("%d/%d\n", count, n);
+    return y;
+}
+
 char **get_labels(char *filename)
 {
     list *plist = get_paths(filename);
@@ -641,8 +689,10 @@ void *load_thread(void *ptr)
 
     //printf("Loading data: %d\n", rand_r(&data_seed));
     load_args a = *(struct load_args*)ptr;
-    if (a.type == CLASSIFICATION_DATA){
+    if (a.type == OLD_CLASSIFICATION_DATA){
         *a.d = load_data(a.paths, a.n, a.m, a.labels, a.classes, a.w, a.h);
+    } else if (a.type == CLASSIFICATION_DATA){
+        *a.d = load_data_augment(a.paths, a.n, a.m, a.labels, a.classes, a.min, a.max, a.size);
     } else if (a.type == DETECTION_DATA){
         *a.d = load_data_detection(a.n, a.paths, a.m, a.classes, a.w, a.h, a.num_boxes, a.background);
     } else if (a.type == WRITING_DATA){
@@ -656,6 +706,9 @@ void *load_thread(void *ptr)
     } else if (a.type == IMAGE_DATA){
         *(a.im) = load_image_color(a.path, 0, 0);
         *(a.resized) = resize_image(*(a.im), a.w, a.h);
+    } else if (a.type == TAG_DATA){
+        *a.d = load_data_tag(a.paths, a.n, a.m, a.classes, a.min, a.max, a.size);
+        //*a.d = load_data(a.paths, a.n, a.m, a.labels, a.classes, a.w, a.h);
     }
     free(ptr);
     return 0;
@@ -696,6 +749,30 @@ data load_data(char **paths, int n, int m, char **labels, int k, int w, int h)
     return d;
 }
 
+data load_data_augment(char **paths, int n, int m, char **labels, int k, int min, int max, int size)
+{
+    if(m) paths = get_random_paths(paths, n, m);
+    data d;
+    d.shallow = 0;
+    d.X = load_image_cropped_paths(paths, n, min, max, size);
+    d.y = load_labels_paths(paths, n, labels, k);
+    if(m) free(paths);
+    return d;
+}
+
+data load_data_tag(char **paths, int n, int m, int k, int min, int max, int size)
+{
+    if(m) paths = get_random_paths(paths, n, m);
+    data d = {0};
+    d.w = size;
+    d.h = size;
+    d.shallow = 0;
+    d.X = load_image_cropped_paths(paths, n, min, max, size);
+    d.y = load_tags_paths(paths, n, k);
+    if(m) free(paths);
+    return d;
+}
+
 matrix concat_matrix(matrix m1, matrix m2)
 {
     int i, count = 0;
@@ -759,8 +836,8 @@ data load_cifar10_data(char *filename)
             X.vals[i][j] = (double)bytes[j+1];
         }
     }
-    translate_data_rows(d, -128);
-    scale_data_rows(d, 1./128);
+    //translate_data_rows(d, -128);
+    scale_data_rows(d, 1./255);
     //normalize_data_rows(d);
     fclose(fp);
     return d;
@@ -800,7 +877,7 @@ data load_all_cifar10()
 
     for(b = 0; b < 5; ++b){
         char buff[256];
-        sprintf(buff, "data/cifar10/data_batch_%d.bin", b+1);
+        sprintf(buff, "data/cifar/cifar-10-batches-bin/data_batch_%d.bin", b+1);
         FILE *fp = fopen(buff, "rb");
         if(!fp) file_error(buff);
         for(i = 0; i < 10000; ++i){
@@ -815,8 +892,8 @@ data load_all_cifar10()
         fclose(fp);
     }
     //normalize_data_rows(d);
-    translate_data_rows(d, -128);
-    scale_data_rows(d, 1./128);
+    //translate_data_rows(d, -128);
+    scale_data_rows(d, 1./255);
     return d;
 }
 
diff --git a/src/data.h b/src/data.h
index 0ebdfc3f..a3036a81 100644
--- a/src/data.h
+++ b/src/data.h
@@ -27,7 +27,7 @@ typedef struct{
 } data;
 
 typedef enum {
-    CLASSIFICATION_DATA, DETECTION_DATA, CAPTCHA_DATA, REGION_DATA, IMAGE_DATA, COMPARE_DATA, WRITING_DATA, SWAG_DATA
+    CLASSIFICATION_DATA, DETECTION_DATA, CAPTCHA_DATA, REGION_DATA, IMAGE_DATA, COMPARE_DATA, WRITING_DATA, SWAG_DATA, TAG_DATA, OLD_CLASSIFICATION_DATA
 } data_type;
 
 typedef struct load_args{
@@ -43,6 +43,7 @@ typedef struct load_args{
     int nh;
     int nw;
     int num_boxes;
+    int min, max, size;
     int classes;
     int background;
     float jitter;
@@ -67,6 +68,8 @@ data load_data_captcha(char **paths, int n, int m, int k, int w, int h);
 data load_data_captcha_encode(char **paths, int n, int m, int w, int h);
 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);
 
 box_label *read_boxes(char *filename, int *n);
 data load_cifar10_data(char *filename);
diff --git a/src/image.c b/src/image.c
index 60ccfb8c..e2cf97fc 100644
--- a/src/image.c
+++ b/src/image.c
@@ -4,11 +4,6 @@
 #include <stdio.h>
 #include <math.h>
 
-#ifdef OPENCV
-#include "opencv2/highgui/highgui_c.h"
-#include "opencv2/imgproc/imgproc_c.h"
-#endif
-
 #define STB_IMAGE_IMPLEMENTATION
 #include "stb_image.h"
 #define STB_IMAGE_WRITE_IMPLEMENTATION
@@ -329,6 +324,16 @@ void save_image(image im, const char *name)
     if(!success) fprintf(stderr, "Failed to write image %s\n", buff);
 }
 
+#ifdef OPENCV
+image get_image_from_stream(CvCapture *cap)
+{
+    IplImage* src = cvQueryFrame(cap);
+    image im = ipl_to_image(src);
+    rgbgr_image(im);
+    return im;
+}
+#endif
+
 #ifdef OPENCV
 void save_image_jpg(image p, char *name)
 {
@@ -459,6 +464,39 @@ image crop_image(image im, int dx, int dy, int w, int h)
     return cropped;
 }
 
+image resize_min(image im, int min)
+{
+    int w = im.w;
+    int h = im.h;
+    if(w < h){
+        h = (h * min) / w;
+        w = min;
+    } else {
+        w = (w * min) / h;
+        h = min;
+    }
+    image resized = resize_image(im, w, h);
+    return resized;
+}
+
+image random_crop_image(image im, int low, int high, int size)
+{
+    int r = rand_int(low, high);
+    image resized = resize_min(im, r);
+    int dx = rand_int(0, resized.w - size);
+    int dy = rand_int(0, resized.h - size);
+    image crop = crop_image(resized, dx, dy, size, size);
+
+    /*
+       show_image(im, "orig");
+       show_image(crop, "cropped");
+       cvWaitKey(0);
+     */
+
+    free_image(resized);
+    return crop;
+}
+
 float three_way_max(float a, float b, float c)
 {
     return (a > b) ? ( (a > c) ? a : c) : ( (b > c) ? b : c) ;
@@ -724,7 +762,7 @@ void test_resize(char *filename)
     image exp5 = copy_image(im);
     exposure_image(exp5, .5);
 
-    #ifdef GPU
+#ifdef GPU
     image r = resize_image(im, im.w, im.h);
     image black = make_image(im.w*2 + 3, im.h*2 + 3, 9);
     image black2 = make_image(im.w, im.h, 3);
@@ -741,7 +779,7 @@ void test_resize(char *filename)
     cuda_pull_array(black2_gpu, black2.data, black2.w*black2.h*black2.c);
     show_image_layers(black, "Black");
     show_image(black2, "Recreate");
-    #endif
+#endif
 
     show_image(im, "Original");
     show_image(gray, "Gray");
@@ -788,8 +826,12 @@ image load_image_cv(char *filename, int channels)
 
     if( (src = cvLoadImage(filename, flag)) == 0 )
     {
-        printf("Cannot load image \"%s\"\n", filename);
-        exit(0);
+        fprintf(stderr, "Cannot load image \"%s\"\n", filename);
+        char buff[256];
+        sprintf(buff, "echo %s >> bad.list", filename);
+        system(buff);
+        return make_image(10,10,3);
+        //exit(0);
     }
     image out = ipl_to_image(src);
     cvReleaseImage(&src);
diff --git a/src/image.h b/src/image.h
index 4846bc19..b4a7a233 100644
--- a/src/image.h
+++ b/src/image.h
@@ -8,6 +8,11 @@
 #include <math.h>
 #include "box.h"
 
+#ifdef OPENCV
+#include "opencv2/highgui/highgui_c.h"
+#include "opencv2/imgproc/imgproc_c.h"
+#endif
+
 typedef struct {
     int h;
     int w;
@@ -25,8 +30,9 @@ void draw_detections(image im, int num, float thresh, box *boxes, float **probs,
 image image_distance(image a, image b);
 void scale_image(image m, float s);
 image crop_image(image im, int dx, int dy, int w, int h);
+image random_crop_image(image im, int low, int high, int size);
 image resize_image(image im, int w, int h);
-image resize_image2(image im, int w, int h);
+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);
@@ -53,6 +59,8 @@ void show_image_collapsed(image p, char *name);
 
 #ifdef OPENCV
 void save_image_jpg(image p, char *name);
+image get_image_from_stream(CvCapture *cap);
+image ipl_to_image(IplImage* src);
 #endif
 
 void print_image(image m);
diff --git a/src/imagenet.c b/src/imagenet.c
index 4c4d2bd6..16255263 100644
--- a/src/imagenet.c
+++ b/src/imagenet.c
@@ -39,7 +39,7 @@ void train_imagenet(char *cfgfile, char *weightfile)
     args.m = N;
     args.labels = labels;
     args.d = &buffer;
-    args.type = CLASSIFICATION_DATA;
+    args.type = OLD_CLASSIFICATION_DATA;
 
     load_thread = load_data_in_thread(args);
     int epoch = (*net.seen)/N;
@@ -115,7 +115,7 @@ void validate_imagenet(char *filename, char *weightfile)
     args.m = 0;
     args.labels = labels;
     args.d = &buffer;
-    args.type = CLASSIFICATION_DATA;
+    args.type = OLD_CLASSIFICATION_DATA;
 
     pthread_t load_thread = load_data_in_thread(args);
     for(i = 1; i <= splits; ++i){
diff --git a/src/layer.h b/src/layer.h
index 91042a21..93083708 100644
--- a/src/layer.h
+++ b/src/layer.h
@@ -22,7 +22,8 @@ typedef enum {
     LOCAL,
     SHORTCUT,
     ACTIVE,
-    RNN
+    RNN,
+    CRNN
 } LAYER_TYPE;
 
 typedef enum{
diff --git a/src/network.c b/src/network.c
index 32c3ba14..e6fb51e5 100644
--- a/src/network.c
+++ b/src/network.c
@@ -9,6 +9,7 @@
 #include "crop_layer.h"
 #include "connected_layer.h"
 #include "rnn_layer.h"
+#include "crnn_layer.h"
 #include "local_layer.h"
 #include "convolutional_layer.h"
 #include "activation_layer.h"
@@ -85,6 +86,8 @@ char *get_layer_string(LAYER_TYPE a)
             return "connected";
         case RNN:
             return "rnn";
+        case CRNN:
+            return "crnn";
         case MAXPOOL:
             return "maxpool";
         case AVGPOOL:
@@ -149,6 +152,8 @@ void forward_network(network net, network_state state)
             forward_connected_layer(l, state);
         } else if(l.type == RNN){
             forward_rnn_layer(l, state);
+        } else if(l.type == CRNN){
+            forward_crnn_layer(l, state);
         } else if(l.type == CROP){
             forward_crop_layer(l, state);
         } else if(l.type == COST){
@@ -185,6 +190,8 @@ void update_network(network net)
             update_connected_layer(l, update_batch, rate, net.momentum, net.decay);
         } else if(l.type == RNN){
             update_rnn_layer(l, update_batch, rate, net.momentum, net.decay);
+        } else if(l.type == CRNN){
+            update_crnn_layer(l, update_batch, rate, net.momentum, net.decay);
         } else if(l.type == LOCAL){
             update_local_layer(l, update_batch, rate, net.momentum, net.decay);
         }
@@ -205,7 +212,7 @@ float get_network_cost(network net)
     int count = 0;
     for(i = 0; i < net.n; ++i){
         if(net.layers[i].type == COST){
-            sum += net.layers[i].output[0];
+            sum += net.layers[i].cost[0];
             ++count;
         }
         if(net.layers[i].type == DETECTION){
@@ -261,6 +268,8 @@ void backward_network(network net, network_state state)
             backward_connected_layer(l, state);
         } else if(l.type == RNN){
             backward_rnn_layer(l, state);
+        } else if(l.type == CRNN){
+            backward_crnn_layer(l, state);
         } else if(l.type == LOCAL){
             backward_local_layer(l, state);
         } else if(l.type == COST){
diff --git a/src/network.h b/src/network.h
index 3d7c5746..f4f8b5cd 100644
--- a/src/network.h
+++ b/src/network.h
@@ -36,6 +36,7 @@ typedef struct network{
 
     int inputs;
     int h, w, c;
+    int max_crop;
 
     #ifdef GPU
     float **input_gpu;
diff --git a/src/network_kernels.cu b/src/network_kernels.cu
index ea128194..730634ef 100644
--- a/src/network_kernels.cu
+++ b/src/network_kernels.cu
@@ -16,6 +16,7 @@ extern "C" {
 #include "crop_layer.h"
 #include "connected_layer.h"
 #include "rnn_layer.h"
+#include "crnn_layer.h"
 #include "detection_layer.h"
 #include "convolutional_layer.h"
 #include "activation_layer.h"
@@ -59,6 +60,8 @@ void forward_network_gpu(network net, network_state state)
             forward_connected_layer_gpu(l, state);
         } else if(l.type == RNN){
             forward_rnn_layer_gpu(l, state);
+        } else if(l.type == CRNN){
+            forward_crnn_layer_gpu(l, state);
         } else if(l.type == CROP){
             forward_crop_layer_gpu(l, state);
         } else if(l.type == COST){
@@ -122,6 +125,8 @@ void backward_network_gpu(network net, network_state state)
             backward_connected_layer_gpu(l, state);
         } else if(l.type == RNN){
             backward_rnn_layer_gpu(l, state);
+        } else if(l.type == CRNN){
+            backward_crnn_layer_gpu(l, state);
         } else if(l.type == COST){
             backward_cost_layer_gpu(l, state);
         } else if(l.type == ROUTE){
@@ -147,6 +152,8 @@ void update_network_gpu(network net)
             update_connected_layer_gpu(l, update_batch, rate, net.momentum, net.decay);
         } else if(l.type == RNN){
             update_rnn_layer_gpu(l, update_batch, rate, net.momentum, net.decay);
+        } else if(l.type == CRNN){
+            update_crnn_layer_gpu(l, update_batch, rate, net.momentum, net.decay);
         } else if(l.type == LOCAL){
             update_local_layer_gpu(l, update_batch, rate, net.momentum, net.decay);
         }
diff --git a/src/nightmare.c b/src/nightmare.c
index 2b1c76cd..ec7166cc 100644
--- a/src/nightmare.c
+++ b/src/nightmare.c
@@ -8,6 +8,8 @@
 #include "opencv2/highgui/highgui_c.h"
 #endif
 
+// ./darknet nightmare cfg/extractor.recon.cfg ~/trained/yolo-coco.conv frame6.png -reconstruct -iters 500 -i 3 -lambda .1 -rate .01 -smooth 2
+
 float abs_mean(float *x, int n)
 {
     int i;
@@ -31,8 +33,8 @@ void calculate_loss(float *output, float *delta, int n, float thresh)
 
 void optimize_picture(network *net, image orig, int max_layer, float scale, float rate, float thresh, int norm)
 {
-    scale_image(orig, 2);
-    translate_image(orig, -1);
+    //scale_image(orig, 2);
+    //translate_image(orig, -1);
     net->n = max_layer + 1;
 
     int dx = rand()%16 - 8;
@@ -98,8 +100,8 @@ void optimize_picture(network *net, image orig, int max_layer, float scale, floa
        translate_image(orig, mean);
      */
 
-    translate_image(orig, 1);
-    scale_image(orig, .5);
+    //translate_image(orig, 1);
+    //scale_image(orig, .5);
     //normalize_image(orig);
 
     constrain_image(orig);
@@ -133,50 +135,47 @@ void smooth(image recon, image update, float lambda, int num)
     }
 }
 
-void reconstruct_picture(network net, float *features, image recon, image update, float rate, float momentum, float lambda, int smooth_size)
+void reconstruct_picture(network net, float *features, image recon, image update, float rate, float momentum, float lambda, int smooth_size, int iters)
 {
-    scale_image(recon, 2);
-    translate_image(recon, -1);
+    int iter = 0;
+    for (iter = 0; iter < iters; ++iter) {
+        image delta = make_image(recon.w, recon.h, recon.c);
 
-    image delta = make_image(recon.w, recon.h, recon.c);
-
-    network_state state = {0};
+        network_state state = {0};
 #ifdef GPU
-    state.input = cuda_make_array(recon.data, recon.w*recon.h*recon.c);
-    state.delta = cuda_make_array(delta.data, delta.w*delta.h*delta.c);
-    state.truth = cuda_make_array(features, get_network_output_size(net));
+        state.input = cuda_make_array(recon.data, recon.w*recon.h*recon.c);
+        state.delta = cuda_make_array(delta.data, delta.w*delta.h*delta.c);
+        state.truth = cuda_make_array(features, get_network_output_size(net));
 
-    forward_network_gpu(net, state);
-    backward_network_gpu(net, state);
+        forward_network_gpu(net, state);
+        backward_network_gpu(net, state);
 
-    cuda_pull_array(state.delta, delta.data, delta.w*delta.h*delta.c);
+        cuda_pull_array(state.delta, delta.data, delta.w*delta.h*delta.c);
 
-    cuda_free(state.input);
-    cuda_free(state.delta);
-    cuda_free(state.truth);
+        cuda_free(state.input);
+        cuda_free(state.delta);
+        cuda_free(state.truth);
 #else
-    state.input = recon.data;
-    state.delta = delta.data;
-    state.truth = features;
+        state.input = recon.data;
+        state.delta = delta.data;
+        state.truth = features;
 
-    forward_network(net, state);
-    backward_network(net, state);
+        forward_network(net, state);
+        backward_network(net, state);
 #endif
 
-    axpy_cpu(recon.w*recon.h*recon.c, 1, delta.data, 1, update.data, 1);
-    smooth(recon, update, lambda, smooth_size);
+        axpy_cpu(recon.w*recon.h*recon.c, 1, delta.data, 1, update.data, 1);
+        smooth(recon, update, lambda, smooth_size);
 
-    axpy_cpu(recon.w*recon.h*recon.c, rate, update.data, 1, recon.data, 1);
-    scal_cpu(recon.w*recon.h*recon.c, momentum, update.data, 1);
+        axpy_cpu(recon.w*recon.h*recon.c, rate, update.data, 1, recon.data, 1);
+        scal_cpu(recon.w*recon.h*recon.c, momentum, update.data, 1);
 
-    translate_image(recon, 1);
-    scale_image(recon, .5);
+        //float mag = mag_array(recon.data, recon.w*recon.h*recon.c);
+        //scal_cpu(recon.w*recon.h*recon.c, 600/mag, recon.data, 1);
 
-    float mag = mag_array(recon.data, recon.w*recon.h*recon.c);
-    scal_cpu(recon.w*recon.h*recon.c, 600/mag, recon.data, 1);
-
-    constrain_image(recon);
-    free_image(delta);
+        constrain_image(recon);
+        free_image(delta);
+    }
 }
 
 
@@ -226,7 +225,7 @@ void run_nightmare(int argc, char **argv)
         im = resized;
     }
 
-    float *features;
+    float *features = 0;
     image update;
     if (reconstruct){
         resize_network(&net, im.w, im.h);
@@ -241,13 +240,19 @@ void run_nightmare(int argc, char **argv)
         printf("%d features\n", out_im.w*out_im.h*out_im.c);
 
 
-        im = resize_image(im, im.w*2, im.h);
-        f_im = resize_image(f_im, f_im.w*2, f_im.h);
+        im = resize_image(im, im.w, im.h);
+        f_im = resize_image(f_im, f_im.w, f_im.h);
         features = f_im.data;
 
+        int i;
+        for(i = 0; i < 14*14*512; ++i){
+            features[i] += rand_uniform(-.19, .19);
+        }
+
         free_image(im);
         im = make_random_image(im.w, im.h, im.c);
         update = make_image(im.w, im.h, im.c);
+
     }
 
     int e;
@@ -259,11 +264,12 @@ void run_nightmare(int argc, char **argv)
             fprintf(stderr, "%d, ", n);
             fflush(stderr);
             if(reconstruct){
-                reconstruct_picture(net, features, im, update, rate, momentum, lambda, smooth_size);
+                reconstruct_picture(net, features, im, update, rate, momentum, lambda, smooth_size, 1);
+                //if ((n+1)%30 == 0) rate *= .5;
                 show_image(im, "reconstruction");
-                #ifdef OPENCV
+#ifdef OPENCV
                 cvWaitKey(10);
-                #endif
+#endif
             }else{
                 int layer = max_layer + rand()%range - range/2;
                 int octave = rand()%octaves;
diff --git a/src/parser.c b/src/parser.c
index 8051fd7f..97ce7a16 100644
--- a/src/parser.c
+++ b/src/parser.c
@@ -12,6 +12,7 @@
 #include "deconvolutional_layer.h"
 #include "connected_layer.h"
 #include "rnn_layer.h"
+#include "crnn_layer.h"
 #include "maxpool_layer.h"
 #include "softmax_layer.h"
 #include "dropout_layer.h"
@@ -36,6 +37,7 @@ int is_local(section *s);
 int is_deconvolutional(section *s);
 int is_connected(section *s);
 int is_rnn(section *s);
+int is_crnn(section *s);
 int is_maxpool(section *s);
 int is_avgpool(section *s);
 int is_dropout(section *s);
@@ -169,6 +171,21 @@ convolutional_layer parse_convolutional(list *options, size_params params)
     return layer;
 }
 
+layer parse_crnn(list *options, size_params params)
+{
+    int output_filters = option_find_int(options, "output_filters",1);
+    int hidden_filters = option_find_int(options, "hidden_filters",1);
+    char *activation_s = option_find_str(options, "activation", "logistic");
+    ACTIVATION activation = get_activation(activation_s);
+    int batch_normalize = option_find_int_quiet(options, "batch_normalize", 0);
+
+    layer l = make_crnn_layer(params.batch, params.w, params.h, params.c, hidden_filters, output_filters, params.time_steps, activation, batch_normalize);
+
+    l.shortcut = option_find_int_quiet(options, "shortcut", 0);
+
+    return l;
+}
+
 layer parse_rnn(list *options, size_params params)
 {
     int output = option_find_int(options, "output",1);
@@ -419,6 +436,7 @@ void parse_net_options(list *options, network *net)
     net->w = option_find_int_quiet(options, "width",0);
     net->c = option_find_int_quiet(options, "channels",0);
     net->inputs = option_find_int_quiet(options, "inputs", net->h * net->w * net->c);
+    net->max_crop = option_find_int_quiet(options, "max_crop",net->w*2);
 
     if(!net->inputs && !(net->h && net->w && net->c)) error("No input parameters supplied");
 
@@ -501,6 +519,8 @@ network parse_network_cfg(char *filename)
             l = parse_deconvolutional(options, params);
         }else if(is_rnn(s)){
             l = parse_rnn(options, params);
+        }else if(is_crnn(s)){
+            l = parse_crnn(options, params);
         }else if(is_connected(s)){
             l = parse_connected(options, params);
         }else if(is_crop(s)){
@@ -591,6 +611,10 @@ int is_network(section *s)
     return (strcmp(s->type, "[net]")==0
             || strcmp(s->type, "[network]")==0);
 }
+int is_crnn(section *s)
+{
+    return (strcmp(s->type, "[crnn]")==0);
+}
 int is_rnn(section *s)
 {
     return (strcmp(s->type, "[rnn]")==0);
@@ -705,6 +729,23 @@ void save_weights_double(network net, char *filename)
     fclose(fp);
 }
 
+void save_convolutional_weights(layer l, FILE *fp)
+{
+#ifdef GPU
+    if(gpu_index >= 0){
+        pull_convolutional_layer(l);
+    }
+#endif
+    int num = l.n*l.c*l.size*l.size;
+    fwrite(l.biases, sizeof(float), l.n, fp);
+    if (l.batch_normalize){
+        fwrite(l.scales, sizeof(float), l.n, fp);
+        fwrite(l.rolling_mean, sizeof(float), l.n, fp);
+        fwrite(l.rolling_variance, sizeof(float), l.n, fp);
+    }
+    fwrite(l.filters, sizeof(float), num, fp);
+}
+
 void save_connected_weights(layer l, FILE *fp)
 {
 #ifdef GPU
@@ -739,25 +780,17 @@ void save_weights_upto(network net, char *filename, int cutoff)
     for(i = 0; i < net.n && i < cutoff; ++i){
         layer l = net.layers[i];
         if(l.type == CONVOLUTIONAL){
-#ifdef GPU
-            if(gpu_index >= 0){
-                pull_convolutional_layer(l);
-            }
-#endif
-            int num = l.n*l.c*l.size*l.size;
-            fwrite(l.biases, sizeof(float), l.n, fp);
-            if (l.batch_normalize){
-                fwrite(l.scales, sizeof(float), l.n, fp);
-                fwrite(l.rolling_mean, sizeof(float), l.n, fp);
-                fwrite(l.rolling_variance, sizeof(float), l.n, fp);
-            }
-            fwrite(l.filters, sizeof(float), num, fp);
+            save_convolutional_weights(l, fp);
         } if(l.type == CONNECTED){
             save_connected_weights(l, fp);
         } if(l.type == RNN){
             save_connected_weights(*(l.input_layer), fp);
             save_connected_weights(*(l.self_layer), fp);
             save_connected_weights(*(l.output_layer), fp);
+        } if(l.type == CRNN){
+            save_convolutional_weights(*(l.input_layer), fp);
+            save_convolutional_weights(*(l.self_layer), fp);
+            save_convolutional_weights(*(l.output_layer), fp);
         } if(l.type == LOCAL){
 #ifdef GPU
             if(gpu_index >= 0){
@@ -809,6 +842,27 @@ void load_connected_weights(layer l, FILE *fp, int transpose)
 #endif
 }
 
+void load_convolutional_weights(layer l, FILE *fp)
+{
+    int num = l.n*l.c*l.size*l.size;
+    fread(l.biases, sizeof(float), l.n, fp);
+    if (l.batch_normalize && (!l.dontloadscales)){
+        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);
+    }
+    fread(l.filters, sizeof(float), num, fp);
+    if (l.flipped) {
+        transpose_matrix(l.filters, l.c*l.size*l.size, l.n);
+    }
+#ifdef GPU
+    if(gpu_index >= 0){
+        push_convolutional_layer(l);
+    }
+#endif
+}
+
+
 void load_weights_upto(network *net, char *filename, int cutoff)
 {
     fprintf(stderr, "Loading weights from %s...", filename);
@@ -830,22 +884,7 @@ void load_weights_upto(network *net, char *filename, int cutoff)
         layer l = net->layers[i];
         if (l.dontload) continue;
         if(l.type == CONVOLUTIONAL){
-            int num = l.n*l.c*l.size*l.size;
-            fread(l.biases, sizeof(float), l.n, fp);
-            if (l.batch_normalize && (!l.dontloadscales)){
-                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);
-            }
-            fread(l.filters, sizeof(float), num, fp);
-            if (l.flipped) {
-                transpose_matrix(l.filters, l.c*l.size*l.size, l.n);
-            }
-#ifdef GPU
-            if(gpu_index >= 0){
-                push_convolutional_layer(l);
-            }
-#endif
+            load_convolutional_weights(l, fp);
         }
         if(l.type == DECONVOLUTIONAL){
             int num = l.n*l.c*l.size*l.size;
@@ -860,6 +899,11 @@ void load_weights_upto(network *net, char *filename, int cutoff)
         if(l.type == CONNECTED){
             load_connected_weights(l, fp, transpose);
         }
+        if(l.type == CRNN){
+            load_convolutional_weights(*(l.input_layer), fp);
+            load_convolutional_weights(*(l.self_layer), fp);
+            load_convolutional_weights(*(l.output_layer), fp);
+        }
         if(l.type == RNN){
             load_connected_weights(*(l.input_layer), fp, transpose);
             load_connected_weights(*(l.self_layer), fp, transpose);
diff --git a/src/rnn.c b/src/rnn.c
index 3865209b..30fa4bd0 100644
--- a/src/rnn.c
+++ b/src/rnn.c
@@ -71,6 +71,7 @@ void train_char_rnn(char *cfgfile, char *weightfile, char *filename)
     fprintf(stderr, "Learning Rate: %g, Momentum: %g, Decay: %g\n", net.learning_rate, net.momentum, net.decay);
     int batch = net.batch;
     int steps = net.time_steps;
+    //*net.seen = 0;
     int i = (*net.seen)/net.batch;
 
     clock_t time;
diff --git a/src/rnn_layer.c b/src/rnn_layer.c
index 384169a4..35cf9923 100644
--- a/src/rnn_layer.c
+++ b/src/rnn_layer.c
@@ -10,7 +10,7 @@
 #include <stdlib.h>
 #include <string.h>
 
-void increment_layer(layer *l, int steps)
+static void increment_layer(layer *l, int steps)
 {
     int num = l->outputs*l->batch*steps;
     l->output += num;
diff --git a/src/rnn_vid.c b/src/rnn_vid.c
new file mode 100644
index 00000000..183ae779
--- /dev/null
+++ b/src/rnn_vid.c
@@ -0,0 +1,210 @@
+#include "network.h"
+#include "cost_layer.h"
+#include "utils.h"
+#include "parser.h"
+#include "blas.h"
+
+#ifdef OPENCV
+#include "opencv2/highgui/highgui_c.h"
+
+void reconstruct_picture(network net, float *features, image recon, image update, float rate, float momentum, float lambda, int smooth_size, int iters);
+
+
+typedef struct {
+    float *x;
+    float *y;
+} float_pair;
+
+float_pair get_rnn_vid_data(network net, char **files, int n, int batch, int steps)
+{
+    int b;
+    assert(net.batch == steps + 1);
+    image out_im = get_network_image(net);
+    int output_size = out_im.w*out_im.h*out_im.c;
+    printf("%d %d %d\n", out_im.w, out_im.h, out_im.c);
+    float *feats = calloc(net.batch*batch*output_size, sizeof(float));
+    for(b = 0; b < batch; ++b){
+        int input_size = net.w*net.h*net.c;
+        float *input = calloc(input_size*net.batch, sizeof(float));
+        char *filename = files[rand()%n];
+        CvCapture *cap = cvCaptureFromFile(filename);
+        int frames = cvGetCaptureProperty(cap, CV_CAP_PROP_FRAME_COUNT);
+        int index = rand() % (frames - steps - 2);
+        if (frames < (steps + 4)){
+            --b;
+            free(input);
+            continue;
+        }
+
+        printf("frames: %d, index: %d\n", frames, index);
+        cvSetCaptureProperty(cap, CV_CAP_PROP_POS_FRAMES, index);
+
+        int i;
+        for(i = 0; i < net.batch; ++i){
+            IplImage* src = cvQueryFrame(cap);
+            image im = ipl_to_image(src);
+            rgbgr_image(im);
+            image re = resize_image(im, net.w, net.h);
+            //show_image(re, "loaded");
+            //cvWaitKey(10);
+            memcpy(input + i*input_size, re.data, input_size*sizeof(float));
+            free_image(im);
+            free_image(re);
+        }
+        float *output = network_predict(net, input);
+
+        free(input);
+
+        for(i = 0; i < net.batch; ++i){
+            memcpy(feats + (b + i*batch)*output_size, output + i*output_size, output_size*sizeof(float));
+        }
+
+        cvReleaseCapture(&cap);
+    }
+
+    //printf("%d %d %d\n", out_im.w, out_im.h, out_im.c);
+    float_pair p = {0};
+    p.x = feats;
+    p.y = feats + output_size*batch; //+ out_im.w*out_im.h*out_im.c;
+
+    return p;
+}
+
+
+void train_vid_rnn(char *cfgfile, char *weightfile)
+{
+    char *train_videos = "data/vid/train.txt";
+    char *backup_directory = "/home/pjreddie/backup/";
+    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 = net.batch*net.subdivisions;
+    int i = *net.seen/imgs;
+
+    list *plist = get_paths(train_videos);
+    int N = plist->size;
+    char **paths = (char **)list_to_array(plist);
+    clock_t time;
+    int steps = net.time_steps;
+    int batch = net.batch / net.time_steps;
+
+    network extractor = parse_network_cfg("cfg/extractor.cfg");
+    load_weights(&extractor, "/home/pjreddie/trained/yolo-coco.conv");
+
+    while(get_current_batch(net) < net.max_batches){
+        i += 1;
+        time=clock();
+        float_pair p = get_rnn_vid_data(extractor, paths, N, batch, steps);
+
+        float loss = train_network_datum(net, p.x, p.y) / (net.batch);
+
+
+        free(p.x);
+        if (avg_loss < 0) avg_loss = loss;
+        avg_loss = avg_loss*.9 + loss*.1;
+
+        fprintf(stderr, "%d: %f, %f avg, %f rate, %lf seconds\n", i, loss, avg_loss, get_current_rate(net), sec(clock()-time));
+        if(i%100==0){
+            char buff[256];
+            sprintf(buff, "%s/%s_%d.weights", backup_directory, base, i);
+            save_weights(net, buff);
+        }
+        if(i%10==0){
+            char buff[256];
+            sprintf(buff, "%s/%s.backup", backup_directory, base);
+            save_weights(net, buff);
+        }
+    }
+    char buff[256];
+    sprintf(buff, "%s/%s_final.weights", backup_directory, base);
+    save_weights(net, buff);
+}
+
+
+image save_reconstruction(network net, image *init, float *feat, char *name, int i)
+{
+    image recon;
+    if (init) {
+        recon = copy_image(*init);
+    } else {
+        recon = make_random_image(net.w, net.h, 3);
+    }
+
+    image update = make_image(net.w, net.h, 3);
+    reconstruct_picture(net, feat, recon, update, .01, .9, .1, 2, 50);
+    char buff[256];
+    sprintf(buff, "%s%d", name, i);
+    save_image(recon, buff);
+    free_image(update);
+    return recon;
+}
+
+void generate_vid_rnn(char *cfgfile, char *weightfile)
+{
+    network extractor = parse_network_cfg("cfg/extractor.recon.cfg");
+    load_weights(&extractor, "/home/pjreddie/trained/yolo-coco.conv");
+
+    network net = parse_network_cfg(cfgfile);
+    if(weightfile){
+        load_weights(&net, weightfile);
+    }
+    set_batch_network(&extractor, 1);
+    set_batch_network(&net, 1);
+
+    int i;
+    CvCapture *cap = cvCaptureFromFile("/extra/vid/ILSVRC2015/Data/VID/snippets/val/ILSVRC2015_val_00007030.mp4");
+    float *feat;
+    float *next;
+    image last;
+    for(i = 0; i < 25; ++i){
+        image im = get_image_from_stream(cap);
+        image re = resize_image(im, extractor.w, extractor.h);
+        feat = network_predict(extractor, re.data);
+        if(i > 0){
+            printf("%f %f\n", mean_array(feat, 14*14*512), variance_array(feat, 14*14*512));
+            printf("%f %f\n", mean_array(next, 14*14*512), variance_array(next, 14*14*512));
+            printf("%f\n", mse_array(feat, 14*14*512));
+            axpy_cpu(14*14*512, -1, feat, 1, next, 1);
+            printf("%f\n", mse_array(next, 14*14*512));
+        }
+        next = network_predict(net, feat);
+
+        free_image(im);
+
+        free_image(save_reconstruction(extractor, 0, feat, "feat", i));
+        free_image(save_reconstruction(extractor, 0, next, "next", i));
+        if (i==24) last = copy_image(re);
+        free_image(re);
+    }
+    for(i = 0; i < 30; ++i){
+        next = network_predict(net, next);
+        image new = save_reconstruction(extractor, &last, next, "new", i);
+        free_image(last);
+        last = new;
+    }
+}
+
+void run_vid_rnn(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;
+    //char *filename = (argc > 5) ? argv[5]: 0;
+    if(0==strcmp(argv[2], "train")) train_vid_rnn(cfg, weights);
+    else if(0==strcmp(argv[2], "generate")) generate_vid_rnn(cfg, weights);
+}
+#else
+void run_vid_rnn(int argc, char **argv){}
+#endif
+
diff --git a/src/tag.c b/src/tag.c
new file mode 100644
index 00000000..8b63d31c
--- /dev/null
+++ b/src/tag.c
@@ -0,0 +1,144 @@
+#include "network.h"
+#include "utils.h"
+#include "parser.h"
+
+#ifdef OPENCV
+#include "opencv2/highgui/highgui_c.h"
+#endif
+
+void train_tag(char *cfgfile, char *weightfile)
+{
+    data_seed = time(0);
+    srand(time(0));
+    float avg_loss = -1;
+    char *base = basecfg(cfgfile);
+    char *backup_directory = "/home/pjreddie/backup/";
+    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);
+    int imgs = 1024;
+    list *plist = get_paths("/home/pjreddie/tag/train.list");
+    char **paths = (char **)list_to_array(plist);
+    printf("%d\n", plist->size);
+    int N = plist->size;
+    clock_t time;
+    pthread_t load_thread;
+    data train;
+    data buffer;
+
+    load_args args = {0};
+    args.w = net.w;
+    args.h = net.h;
+
+    args.min = net.w;
+    args.max = net.max_crop;
+    args.size = net.w;
+
+    args.paths = paths;
+    args.classes = net.outputs;
+    args.n = imgs;
+    args.m = N;
+    args.d = &buffer;
+    args.type = TAG_DATA;
+
+    fprintf(stderr, "%d classes\n", net.outputs);
+
+    load_thread = load_data_in_thread(args);
+    int epoch = (*net.seen)/N;
+    while(get_current_batch(net) < net.max_batches || net.max_batches == 0){
+        time=clock();
+        pthread_join(load_thread, 0);
+        train = buffer;
+
+        load_thread = load_data_in_thread(args);
+        printf("Loaded: %lf seconds\n", sec(clock()-time));
+        time=clock();
+        float loss = train_network(net, train);
+        if(avg_loss == -1) avg_loss = loss;
+        avg_loss = avg_loss*.9 + loss*.1;
+        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);
+        free_data(train);
+        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);
+
+    pthread_join(load_thread, 0);
+    free_data(buffer);
+    free_network(net);
+    free_ptrs((void**)paths, plist->size);
+    free_list(plist);
+    free(base);
+}
+
+void test_tag(char *cfgfile, char *weightfile, char *filename)
+{
+    network net = parse_network_cfg(cfgfile);
+    if(weightfile){
+        load_weights(&net, weightfile);
+    }
+    set_batch_network(&net, 1);
+    srand(2222222);
+    int i = 0;
+    char **names = get_labels("data/tags.txt");
+    clock_t time;
+    int indexes[10];
+    char buff[256];
+    char *input = buff;
+    while(1){
+        if(filename){
+            strncpy(input, filename, 256);
+        }else{
+            printf("Enter Image Path: ");
+            fflush(stdout);
+            input = fgets(input, 256, stdin);
+            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);
+
+        float *X = im.data;
+        time=clock();
+        float *predictions = network_predict(net, X);
+        top_predictions(net, 10, indexes);
+        printf("%s: Predicted in %f seconds.\n", input, sec(clock()-time));
+        for(i = 0; i < 10; ++i){
+            int index = indexes[i];
+            printf("%.1f%%: %s\n", predictions[index]*100, names[index]);
+        }
+        free_image(im);
+        if (filename) break;
+    }
+}
+
+
+void run_tag(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;
+    char *filename = (argc > 5) ? argv[5] : 0;
+    if(0==strcmp(argv[2], "train")) train_tag(cfg, weights);
+    else if(0==strcmp(argv[2], "test")) test_tag(cfg, weights, filename);
+}
+
diff --git a/src/utils.c b/src/utils.c
index ec87a265..398d18a8 100644
--- a/src/utils.c
+++ b/src/utils.c
@@ -2,6 +2,7 @@
 #include <stdlib.h>
 #include <string.h>
 #include <math.h>
+#include <assert.h>
 #include <unistd.h>
 #include <float.h>
 #include <limits.h>
@@ -137,15 +138,18 @@ void pm(int M, int N, float *A)
 char *find_replace(char *str, char *orig, char *rep)
 {
     static char buffer[4096];
+    static char buffer2[4096];
+    static char buffer3[4096];
     char *p;
 
     if(!(p = strstr(str, orig)))  // Is 'orig' even in 'str'?
         return str;
 
-    strncpy(buffer, str, p-str); // Copy characters from 'str' start to 'orig' st$
-    buffer[p-str] = '\0';
+    strncpy(buffer2, str, p-str); // Copy characters from 'str' start to 'orig' st$
+    buffer2[p-str] = '\0';
 
-    sprintf(buffer+(p-str), "%s%s", rep, p+strlen(orig));
+    sprintf(buffer3, "%s%s%s", buffer2, rep, p+strlen(orig));
+    sprintf(buffer, "%s", buffer3);
 
     return buffer;
 }
@@ -174,7 +178,8 @@ void top_k(float *a, int n, int k, int *index)
 void error(const char *s)
 {
     perror(s);
-    exit(0);
+    assert(0);
+    exit(-1);
 }
 
 void malloc_error()
@@ -450,6 +455,12 @@ int max_index(float *a, int n)
     return max_i;
 }
 
+int rand_int(int min, int max)
+{
+    int r = (rand()%(max - min + 1)) + min;
+    return r;
+}
+
 // From http://en.wikipedia.org/wiki/Box%E2%80%93Muller_transform
 #define TWO_PI 6.2831853071795864769252866
 float rand_normal()
diff --git a/src/utils.h b/src/utils.h
index 96bd6cfe..3af85d33 100644
--- a/src/utils.h
+++ b/src/utils.h
@@ -35,6 +35,7 @@ float constrain(float min, float max, float a);
 float mse_array(float *a, int n);
 float rand_normal();
 float rand_uniform(float min, float max);
+int rand_int(int min, int max);
 float sum_array(float *a, int n);
 float mean_array(float *a, int n);
 void mean_arrays(float **a, int n, int els, float *avg);