diff --git a/cfg/gru.cfg b/cfg/gru.cfg
new file mode 100644
index 00000000..f9a06999
--- /dev/null
+++ b/cfg/gru.cfg
@@ -0,0 +1,34 @@
+[net]
+subdivisions=1
+inputs=256
+batch = 1
+momentum=0.9
+decay=0.001
+time_steps=1
+learning_rate=0.5
+
+policy=poly
+power=4
+max_batches=2000
+
+[gru]
+batch_normalize=1
+output = 1024
+
+[gru]
+batch_normalize=1
+output = 1024
+
+[gru]
+batch_normalize=1
+output = 1024
+
+[connected]
+output=256
+activation=linear
+
+[softmax]
+
+[cost]
+type=sse
+
diff --git a/src/blas.c b/src/blas.c
index 35a4c40a..00f0c3a3 100644
--- a/src/blas.c
+++ b/src/blas.c
@@ -2,6 +2,14 @@
 #include "math.h"
 #include <assert.h>
 
+void weighted_sum_cpu(float *a, float *b, float *s, int n, float *c)
+{
+    int i;
+    for(i = 0; i < n; ++i){
+        c[i] = s[i]*a[i] + (1-s[i])*(b ? b[i] : 0);
+    }
+}
+
 void shortcut_cpu(int batch, int w1, int h1, int c1, float *add, int w2, int h2, int c2, float *out)
 {
     int stride = w1/w2;
diff --git a/src/blas.h b/src/blas.h
index 47d930c3..b4cfcf2e 100644
--- a/src/blas.h
+++ b/src/blas.h
@@ -31,6 +31,7 @@ void normalize_delta_cpu(float *x, float *mean, float *variance, float *mean_del
 
 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);
+void weighted_sum_cpu(float *a, float *b, float *s, int num, float *c);
 
 #ifdef GPU
 void axpy_ongpu(int N, float ALPHA, float * X, int INCX, float * Y, int INCY);
diff --git a/src/convolutional_layer.c b/src/convolutional_layer.c
index 5575aacf..c377802f 100644
--- a/src/convolutional_layer.c
+++ b/src/convolutional_layer.c
@@ -413,6 +413,7 @@ void forward_convolutional_layer(convolutional_layer l, network_state state)
        }
      */
 
+/*
     if(l.binary){
         int m = l.n;
         int k = l.size*l.size*l.c;
@@ -434,6 +435,7 @@ void forward_convolutional_layer(convolutional_layer l, network_state state)
         activate_array(l.output, m*n*l.batch, l.activation);
         return;
     }
+    */
 
     int m = l.n;
     int k = l.size*l.size*l.c;
diff --git a/src/detection_layer.c b/src/detection_layer.c
index 90b672b1..1adda06b 100644
--- a/src/detection_layer.c
+++ b/src/detection_layer.c
@@ -175,6 +175,38 @@ void forward_detection_layer(const detection_layer l, network_state state)
                         LOGISTIC, l.delta + index + locations*l.classes);
             }
         }
+
+        if(1){
+            float *costs = calloc(l.batch*locations*l.n, sizeof(float));
+            for (b = 0; b < l.batch; ++b) {
+                int index = b*l.inputs;
+                for (i = 0; i < locations; ++i) {
+                    int truth_index = (b*locations + i)*(1+l.coords+l.classes);
+                    for (j = 0; j < l.n; ++j) {
+                        int p_index = index + locations*l.classes + i*l.n + j;
+                        costs[b*locations*l.n + i*l.n + j] = l.delta[p_index]*l.delta[p_index];
+                    }
+                }
+            }
+            int indexes[100];
+            top_k(costs, l.batch*locations*l.n, 100, indexes);
+            float cutoff = costs[indexes[99]];
+            for (b = 0; b < l.batch; ++b) {
+                int index = b*l.inputs;
+                for (i = 0; i < locations; ++i) {
+                    int truth_index = (b*locations + i)*(1+l.coords+l.classes);
+                    for (j = 0; j < l.n; ++j) {
+                        int p_index = index + locations*l.classes + i*l.n + j;
+                        if (l.delta[p_index]*l.delta[p_index] < cutoff) l.delta[p_index] = 0;
+                    }
+                }
+            }
+            free(costs);
+        }
+
+
+
+
         printf("Detection Avg IOU: %f, Pos Cat: %f, All Cat: %f, Pos Obj: %f, Any Obj: %f, count: %d\n", avg_iou/count, avg_cat/count, avg_allcat/(count*l.classes), avg_obj/count, avg_anyobj/(l.batch*locations*l.n), count);
     }
 }
diff --git a/src/gru_layer.c b/src/gru_layer.c
index 1c41cbfb..4c720ce3 100644
--- a/src/gru_layer.c
+++ b/src/gru_layer.c
@@ -76,6 +76,14 @@ layer make_gru_layer(int batch, int inputs, int outputs, int steps, int batch_no
     l.outputs = outputs;
     l.output = calloc(outputs*batch*steps, sizeof(float));
     l.delta = calloc(outputs*batch*steps, sizeof(float));
+    l.state = calloc(outputs*batch, sizeof(float));
+    l.prev_state = calloc(outputs*batch, sizeof(float));
+    l.forgot_state = calloc(outputs*batch, sizeof(float));
+    l.forgot_delta = calloc(outputs*batch, sizeof(float));
+
+    l.r_cpu = calloc(outputs*batch, sizeof(float));
+    l.z_cpu = calloc(outputs*batch, sizeof(float));
+    l.h_cpu = calloc(outputs*batch, sizeof(float));
 
 #ifdef GPU
     l.forgot_state_gpu = cuda_make_array(l.output, batch*outputs);
@@ -101,6 +109,78 @@ void update_gru_layer(layer l, int batch, float learning_rate, float momentum, f
 
 void forward_gru_layer(layer l, network_state state)
 {
+    network_state s = {0};
+    s.train = state.train;
+    int i;
+    layer input_z_layer = *(l.input_z_layer);
+    layer input_r_layer = *(l.input_r_layer);
+    layer input_h_layer = *(l.input_h_layer);
+
+    layer state_z_layer = *(l.state_z_layer);
+    layer state_r_layer = *(l.state_r_layer);
+    layer state_h_layer = *(l.state_h_layer);
+
+    fill_cpu(l.outputs * l.batch * l.steps, 0, input_z_layer.delta, 1);
+    fill_cpu(l.outputs * l.batch * l.steps, 0, input_r_layer.delta, 1);
+    fill_cpu(l.outputs * l.batch * l.steps, 0, input_h_layer.delta, 1);
+
+    fill_cpu(l.outputs * l.batch * l.steps, 0, state_z_layer.delta, 1);
+    fill_cpu(l.outputs * l.batch * l.steps, 0, state_r_layer.delta, 1);
+    fill_cpu(l.outputs * l.batch * l.steps, 0, state_h_layer.delta, 1);
+    if(state.train) {
+        fill_cpu(l.outputs * l.batch * l.steps, 0, l.delta, 1);
+        copy_cpu(l.outputs*l.batch, l.state, 1, l.prev_state, 1);
+    }
+
+    for (i = 0; i < l.steps; ++i) {
+        s.input = l.state;
+        forward_connected_layer(state_z_layer, s);
+        forward_connected_layer(state_r_layer, s);
+
+        s.input = state.input;
+        forward_connected_layer(input_z_layer, s);
+        forward_connected_layer(input_r_layer, s);
+        forward_connected_layer(input_h_layer, s);
+
+
+        copy_cpu(l.outputs*l.batch, input_z_layer.output, 1, l.z_cpu, 1);
+        axpy_cpu(l.outputs*l.batch, 1, state_z_layer.output, 1, l.z_cpu, 1);
+
+        copy_cpu(l.outputs*l.batch, input_r_layer.output, 1, l.r_cpu, 1);
+        axpy_cpu(l.outputs*l.batch, 1, state_r_layer.output, 1, l.r_cpu, 1);
+
+        activate_array(l.z_cpu, l.outputs*l.batch, LOGISTIC);
+        activate_array(l.r_cpu, l.outputs*l.batch, LOGISTIC);
+
+        copy_cpu(l.outputs*l.batch, l.state, 1, l.forgot_state, 1);
+        mul_cpu(l.outputs*l.batch, l.r_cpu, 1, l.forgot_state, 1);
+
+        s.input = l.forgot_state;
+        forward_connected_layer(state_h_layer, s);
+
+        copy_cpu(l.outputs*l.batch, input_h_layer.output, 1, l.h_cpu, 1);
+        axpy_cpu(l.outputs*l.batch, 1, state_h_layer.output, 1, l.h_cpu, 1);
+
+        #ifdef USET
+        activate_array(l.h_cpu, l.outputs*l.batch, TANH);
+        #else
+        activate_array(l.h_cpu, l.outputs*l.batch, LOGISTIC);
+        #endif
+
+        weighted_sum_cpu(l.state, l.h_cpu, l.z_cpu, l.outputs*l.batch, l.output);
+
+        copy_cpu(l.outputs*l.batch, l.output, 1, l.state, 1);
+
+        state.input += l.inputs*l.batch;
+        l.output += l.outputs*l.batch;
+        increment_layer(&input_z_layer, 1);
+        increment_layer(&input_r_layer, 1);
+        increment_layer(&input_h_layer, 1);
+
+        increment_layer(&state_z_layer, 1);
+        increment_layer(&state_r_layer, 1);
+        increment_layer(&state_h_layer, 1);
+    }
 }
 
 void backward_gru_layer(layer l, network_state state)
diff --git a/src/layer.h b/src/layer.h
index d53fe387..d2250a6d 100644
--- a/src/layer.h
+++ b/src/layer.h
@@ -28,6 +28,7 @@ typedef enum {
     CRNN,
     BATCHNORM,
     NETWORK,
+    XNOR,
     BLANK
 } LAYER_TYPE;
 
@@ -102,6 +103,9 @@ struct layer{
     char  *cfilters;
     float *filter_updates;
     float *state;
+    float *prev_state;
+    float *forgot_state;
+    float *forgot_delta;
     float *state_delta;
 
     float *concat;
@@ -159,6 +163,10 @@ struct layer{
     struct layer *input_h_layer;
     struct layer *state_h_layer;
 
+    float *z_cpu;
+    float *r_cpu;
+    float *h_cpu;
+
     size_t workspace_size;
 
     #ifdef GPU
diff --git a/src/tag.c b/src/tag.c
index a00a161a..f97621c3 100644
--- a/src/tag.c
+++ b/src/tag.c
@@ -6,7 +6,7 @@
 #include "opencv2/highgui/highgui_c.h"
 #endif
 
-void train_tag(char *cfgfile, char *weightfile)
+void train_tag(char *cfgfile, char *weightfile, int clear)
 {
     data_seed = time(0);
     srand(time(0));
@@ -18,6 +18,7 @@ void train_tag(char *cfgfile, char *weightfile)
     if(weightfile){
         load_weights(&net, weightfile);
     }
+    if(clear) *net.seen = 0;
     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");
@@ -138,10 +139,11 @@ void run_tag(int argc, char **argv)
         return;
     }
 
+    int clear = find_arg(argc, argv, "-clear");
     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);
+    if(0==strcmp(argv[2], "train")) train_tag(cfg, weights, clear);
     else if(0==strcmp(argv[2], "test")) test_tag(cfg, weights, filename);
 }