diff --git a/Makefile b/Makefile
index de515d32..c9b6ecac 100644
--- a/Makefile
+++ b/Makefile
@@ -1,5 +1,5 @@
-GPU=1
-OPENCV=1
+GPU=0
+OPENCV=0
 DEBUG=0
 
 ARCH= --gpu-architecture=compute_20 --gpu-code=compute_20
diff --git a/cfg/rnn.cfg b/cfg/rnn.cfg
index a67e1fa1..68c032d2 100644
--- a/cfg/rnn.cfg
+++ b/cfg/rnn.cfg
@@ -1,29 +1,32 @@
 [net]
 subdivisions=1
 inputs=256
-batch = 128
+batch = 1
 momentum=0.9
 decay=0.001
-max_batches = 50000
-time_steps=900
+max_batches = 2000
+time_steps=1
 learning_rate=0.1
+policy=steps
+steps=1000,1500
+scales=.1,.1
 
 [rnn]
 batch_normalize=1
-output = 256
-hidden=512
+output = 1024
+hidden=1024
 activation=leaky
 
 [rnn]
 batch_normalize=1
-output = 256
-hidden=512
+output = 1024
+hidden=1024
 activation=leaky
 
 [rnn]
 batch_normalize=1
-output = 256
-hidden=512
+output = 1024
+hidden=1024
 activation=leaky
 
 [connected]
diff --git a/cfg/rnn.train.cfg b/cfg/rnn.train.cfg
new file mode 100644
index 00000000..9139757f
--- /dev/null
+++ b/cfg/rnn.train.cfg
@@ -0,0 +1,40 @@
+[net]
+subdivisions=1
+inputs=256
+batch = 128
+momentum=0.9
+decay=0.001
+max_batches = 2000
+time_steps=576
+learning_rate=0.1
+policy=steps
+steps=1000,1500
+scales=.1,.1
+
+[rnn]
+batch_normalize=1
+output = 1024
+hidden=1024
+activation=leaky
+
+[rnn]
+batch_normalize=1
+output = 1024
+hidden=1024
+activation=leaky
+
+[rnn]
+batch_normalize=1
+output = 1024
+hidden=1024
+activation=leaky
+
+[connected]
+output=256
+activation=leaky
+
+[softmax]
+
+[cost]
+type=sse
+
diff --git a/src/rnn.c b/src/rnn.c
index aee53ffb..3984d93d 100644
--- a/src/rnn.c
+++ b/src/rnn.c
@@ -12,22 +12,31 @@ typedef struct {
     float *y;
 } float_pair;
 
-float_pair get_rnn_data(char *text, int len, int batch, int steps)
+float_pair get_rnn_data(unsigned char *text, int characters, int len, int batch, int steps)
 {
-    float *x = calloc(batch * steps * 256, sizeof(float));
-    float *y = calloc(batch * steps * 256, sizeof(float));
+    float *x = calloc(batch * steps * characters, sizeof(float));
+    float *y = calloc(batch * steps * characters, sizeof(float));
     int i,j;
     for(i = 0; i < batch; ++i){
         int index = rand() %(len - steps - 1);
+        /*
         int done = 1;
         while(!done){
             index = rand() %(len - steps - 1);
             while(index < len-steps-1 && text[index++] != '\n');
             if (index < len-steps-1) done = 1;
-        }
+            }
+         */
         for(j = 0; j < steps; ++j){
-            x[(j*batch + i)*256 + text[index + j]] = 1;
-            y[(j*batch + i)*256 + text[index + j + 1]] = 1;
+            x[(j*batch + i)*characters + text[index + j]] = 1;
+            y[(j*batch + i)*characters + text[index + j + 1]] = 1;
+
+            if(text[index+j] > 255 || text[index+j] <= 0 || text[index+j+1] > 255 || text[index+j+1] <= 0){
+                text[index+j+2] = 0;
+                printf("%d %d %d %d %d\n", index, j, len, (int)text[index+j], (int)text[index+j+1]);
+                printf("%s", text+index);
+                error("Bad char");
+            }
         }
     }
     float_pair p;
@@ -38,7 +47,7 @@ float_pair get_rnn_data(char *text, int len, int batch, int steps)
 
 void train_char_rnn(char *cfgfile, char *weightfile, char *filename)
 {
-    FILE *fp = fopen(filename, "r");
+    FILE *fp = fopen(filename, "rb");
     //FILE *fp = fopen("data/ab.txt", "r");
     //FILE *fp = fopen("data/grrm/asoiaf.txt", "r");
 
@@ -46,7 +55,7 @@ void train_char_rnn(char *cfgfile, char *weightfile, char *filename)
     size_t size = ftell(fp);
     fseek(fp, 0, SEEK_SET); 
 
-    char *text = calloc(size, sizeof(char));
+    unsigned char *text = calloc(size+1, sizeof(char));
     fread(text, 1, size, fp);
     fclose(fp);
 
@@ -60,6 +69,7 @@ void train_char_rnn(char *cfgfile, char *weightfile, char *filename)
     if(weightfile){
         load_weights(&net, weightfile);
     }
+    int inputs = get_network_input_size(net);
     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;
@@ -69,7 +79,7 @@ void train_char_rnn(char *cfgfile, char *weightfile, char *filename)
     while(get_current_batch(net) < net.max_batches){
         i += 1;
         time=clock();
-        float_pair p = get_rnn_data(text, size, batch/steps, steps);
+        float_pair p = get_rnn_data(text, inputs, size, batch/steps, steps);
 
         float loss = train_network_datum(net, p.x, p.y) / (batch);
         free(p.x);
@@ -104,12 +114,13 @@ void test_char_rnn(char *cfgfile, char *weightfile, int num, char *seed, float t
     if(weightfile){
         load_weights(&net, weightfile);
     }
-    
+    int inputs = get_network_input_size(net);
+
     int i, j;
     for(i = 0; i < net.n; ++i) net.layers[i].temperature = temp;
-    char c;
+    unsigned char c;
     int len = strlen(seed);
-    float *input = calloc(256, sizeof(float));
+    float *input = calloc(inputs, sizeof(float));
     for(i = 0; i < len-1; ++i){
         c = seed[i];
         input[(int)c] = 1;
@@ -125,7 +136,7 @@ void test_char_rnn(char *cfgfile, char *weightfile, int num, char *seed, float t
         input[(int)c] = 1;
         float *out = network_predict(net, input);
         input[(int)c] = 0;
-        for(j = 0; j < 256; ++j){
+        for(j = 0; j < inputs; ++j){
             sum += out[j];
             if(sum > r) break;
         }
@@ -134,20 +145,8 @@ void test_char_rnn(char *cfgfile, char *weightfile, int num, char *seed, float t
     printf("\n");
 }
 
-void valid_char_rnn(char *cfgfile, char *weightfile, char *filename)
+void valid_char_rnn(char *cfgfile, char *weightfile)
 {
-    FILE *fp = fopen(filename, "r");
-    //FILE *fp = fopen("data/ab.txt", "r");
-    //FILE *fp = fopen("data/grrm/asoiaf.txt", "r");
-
-    fseek(fp, 0, SEEK_END); 
-    size_t size = ftell(fp);
-    fseek(fp, 0, SEEK_SET); 
-
-    char *text = calloc(size, sizeof(char));
-    fread(text, 1, size, fp);
-    fclose(fp);
-
     char *base = basecfg(cfgfile);
     fprintf(stderr, "%s\n", base);
 
@@ -155,19 +154,25 @@ void valid_char_rnn(char *cfgfile, char *weightfile, char *filename)
     if(weightfile){
         load_weights(&net, weightfile);
     }
-    
-    int i;
-    char c;
-    float *input = calloc(256, sizeof(float));
+    int inputs = get_network_input_size(net);
+
+    int count = 0;
+    int c;
+    float *input = calloc(inputs, sizeof(float));
     float sum = 0;
-    for(i = 0; i < size-1; ++i){
-        c = text[i];
-        input[(int)c] = 1;
+    c = getc(stdin);
+    float log2 = log(2);
+    while(c != EOF){
+        int next = getc(stdin);
+        if(next == EOF) break;
+        ++count;
+        input[c] = 1;
         float *out = network_predict(net, input);
-        input[(int)c] = 0;
-        sum += log(out[(int)text[i+1]]);
+        input[c] = 0;
+        sum += log(out[next])/log2;
+        c = next;
     }
-    printf("Log Probability: %f\n", sum);
+    printf("Perplexity: %f\n", pow(2, -sum/count));
 }
 
 
@@ -179,13 +184,13 @@ void run_char_rnn(int argc, char **argv)
     }
     char *filename = find_char_arg(argc, argv, "-file", "data/shakespeare.txt");
     char *seed = find_char_arg(argc, argv, "-seed", "\n");
-    int len = find_int_arg(argc, argv, "-len", 100);
-    float temp = find_float_arg(argc, argv, "-temp", 1);
+    int len = find_int_arg(argc, argv, "-len", 1000);
+    float temp = find_float_arg(argc, argv, "-temp", .7);
     int rseed = find_int_arg(argc, argv, "-srand", time(0));
 
     char *cfg = argv[3];
     char *weights = (argc > 4) ? argv[4] : 0;
     if(0==strcmp(argv[2], "train")) train_char_rnn(cfg, weights, filename);
-    else if(0==strcmp(argv[2], "valid")) valid_char_rnn(cfg, weights, filename);
+    else if(0==strcmp(argv[2], "valid")) valid_char_rnn(cfg, weights);
     else if(0==strcmp(argv[2], "test")) test_char_rnn(cfg, weights, len, seed, temp, rseed);
 }
diff --git a/src/rnn_layer.c b/src/rnn_layer.c
index e58e0a4b..a6b025da 100644
--- a/src/rnn_layer.c
+++ b/src/rnn_layer.c
@@ -10,6 +10,19 @@
 #include <stdlib.h>
 #include <string.h>
 
+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;
+
+    l->output_gpu += num;
+    l->delta_gpu += num;
+    l->x_gpu += num;
+    l->x_norm_gpu += num;
+}
 
 layer make_rnn_layer(int batch, int inputs, int hidden, int outputs, int steps, ACTIVATION activation, int batch_normalize, int log)
 {
@@ -22,7 +35,7 @@ layer make_rnn_layer(int batch, int inputs, int hidden, int outputs, int steps,
     l.hidden = hidden;
     l.inputs = inputs;
 
-    l.state = calloc(batch*hidden, sizeof(float));
+    l.state = calloc(batch*hidden*(steps+1), sizeof(float));
 
     l.input_layer = malloc(sizeof(layer));
     fprintf(stderr, "\t\t");
@@ -43,11 +56,11 @@ layer make_rnn_layer(int batch, int inputs, int hidden, int outputs, int steps,
     l.output = l.output_layer->output;
     l.delta = l.output_layer->delta;
 
-    #ifdef GPU
-    l.state_gpu = cuda_make_array(l.state, batch*hidden);
+#ifdef GPU
+    l.state_gpu = cuda_make_array(l.state, batch*hidden*(steps+1));
     l.output_gpu = l.output_layer->output_gpu;
     l.delta_gpu = l.output_layer->delta_gpu;
-    #endif
+#endif
 
     return l;
 }
@@ -80,16 +93,23 @@ void forward_rnn_layer(layer l, network_state state)
         s.input = l.state;
         forward_connected_layer(self_layer, s);
 
-        copy_cpu(l.hidden * l.batch, input_layer.output, 1, l.state, 1);
+        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_connected_layer(output_layer, s);
 
         state.input += l.inputs*l.batch;
-        input_layer.output += l.hidden*l.batch;
-        self_layer.output += l.hidden*l.batch;
-        output_layer.output += l.outputs*l.batch;
+        increment_layer(&input_layer, 1);
+        increment_layer(&self_layer, 1);
+        increment_layer(&output_layer, 1);
     }
 }
 
@@ -101,14 +121,12 @@ void backward_rnn_layer(layer l, network_state state)
     layer input_layer = *(l.input_layer);
     layer self_layer = *(l.self_layer);
     layer output_layer = *(l.output_layer);
-    input_layer.output += l.hidden*l.batch*(l.steps-1);
-    input_layer.delta  += l.hidden*l.batch*(l.steps-1);
 
-    self_layer.output += l.hidden*l.batch*(l.steps-1);
-    self_layer.delta  += l.hidden*l.batch*(l.steps-1);
+    increment_layer(&input_layer, l.steps-1);
+    increment_layer(&self_layer, l.steps-1);
+    increment_layer(&output_layer, l.steps-1);
 
-    output_layer.output += l.outputs*l.batch*(l.steps-1);
-    output_layer.delta  += l.outputs*l.batch*(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);
@@ -116,13 +134,16 @@ void backward_rnn_layer(layer l, network_state state)
         s.input = l.state;
         s.delta = self_layer.delta;
         backward_connected_layer(output_layer, s);
-        
-        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);
-        }
+
+        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;
@@ -130,19 +151,15 @@ void backward_rnn_layer(layer l, network_state state)
         backward_connected_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_connected_layer(input_layer, s);
 
-        input_layer.output  -= l.hidden*l.batch;
-        input_layer.delta   -= l.hidden*l.batch;
-
-        self_layer.output   -= l.hidden*l.batch;
-        self_layer.delta    -= l.hidden*l.batch;
-
-        output_layer.output -= l.outputs*l.batch;
-        output_layer.delta  -= l.outputs*l.batch;
+        increment_layer(&input_layer, -1);
+        increment_layer(&self_layer, -1);
+        increment_layer(&output_layer, -1);
     }
 }
 
@@ -190,23 +207,23 @@ void forward_rnn_layer_gpu(layer l, network_state state)
         s.input = l.state_gpu;
         forward_connected_layer_gpu(self_layer, s);
 
-        copy_ongpu(l.hidden * l.batch, input_layer.output_gpu, 1, l.state_gpu, 1);
+        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_connected_layer_gpu(output_layer, s);
 
         state.input += l.inputs*l.batch;
-        input_layer.output_gpu += l.hidden*l.batch;
-        input_layer.x_gpu += l.hidden*l.batch;
-        input_layer.x_norm_gpu += l.hidden*l.batch;
-
-        self_layer.output_gpu += l.hidden*l.batch;
-        self_layer.x_gpu += l.hidden*l.batch;
-        self_layer.x_norm_gpu += l.hidden*l.batch;
-
-        output_layer.output_gpu += l.outputs*l.batch;
-        output_layer.x_gpu += l.outputs*l.batch;
-        output_layer.x_norm_gpu += l.outputs*l.batch;
+        increment_layer(&input_layer, 1);
+        increment_layer(&self_layer, 1);
+        increment_layer(&output_layer, 1);
     }
 }
 
@@ -218,20 +235,10 @@ void backward_rnn_layer_gpu(layer l, network_state state)
     layer input_layer = *(l.input_layer);
     layer self_layer = *(l.self_layer);
     layer output_layer = *(l.output_layer);
-    input_layer.output_gpu += l.hidden*l.batch*(l.steps-1);
-    input_layer.delta_gpu  += l.hidden*l.batch*(l.steps-1);
-    input_layer.x_gpu  += l.hidden*l.batch*(l.steps-1);
-    input_layer.x_norm_gpu  += l.hidden*l.batch*(l.steps-1);
-
-    self_layer.output_gpu += l.hidden*l.batch*(l.steps-1);
-    self_layer.delta_gpu  += l.hidden*l.batch*(l.steps-1);
-    self_layer.x_gpu  += l.hidden*l.batch*(l.steps-1);
-    self_layer.x_norm_gpu  += l.hidden*l.batch*(l.steps-1);
-
-    output_layer.output_gpu += l.outputs*l.batch*(l.steps-1);
-    output_layer.delta_gpu  += l.outputs*l.batch*(l.steps-1);
-    output_layer.x_gpu  += l.outputs*l.batch*(l.steps-1);
-    output_layer.x_norm_gpu  += l.outputs*l.batch*(l.steps-1);
+    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);
@@ -239,13 +246,8 @@ void backward_rnn_layer_gpu(layer l, network_state state)
         s.input = l.state_gpu;
         s.delta = self_layer.delta_gpu;
         backward_connected_layer_gpu(output_layer, s);
-        
-        if(i > 0){
-            copy_ongpu(l.hidden * l.batch, input_layer.output_gpu - l.hidden*l.batch, 1, l.state_gpu, 1);
-            axpy_ongpu(l.hidden * l.batch, 1, self_layer.output_gpu - l.hidden*l.batch, 1, l.state_gpu, 1);
-        }else{
-            fill_ongpu(l.hidden * l.batch, 0, l.state_gpu, 1);
-        }
+
+        l.state_gpu -= l.hidden*l.batch;
 
         s.input = l.state_gpu;
         s.delta = self_layer.delta_gpu - l.hidden*l.batch;
@@ -253,25 +255,15 @@ void backward_rnn_layer_gpu(layer l, network_state state)
         backward_connected_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_connected_layer_gpu(input_layer, s);
 
-        input_layer.output_gpu  -= l.hidden*l.batch;
-        input_layer.delta_gpu   -= l.hidden*l.batch;
-        input_layer.x_gpu   -= l.hidden*l.batch;
-        input_layer.x_norm_gpu   -= l.hidden*l.batch;
-
-        self_layer.output_gpu   -= l.hidden*l.batch;
-        self_layer.delta_gpu    -= l.hidden*l.batch;
-        self_layer.x_gpu    -= l.hidden*l.batch;
-        self_layer.x_norm_gpu    -= l.hidden*l.batch;
-
-        output_layer.output_gpu -= l.outputs*l.batch;
-        output_layer.delta_gpu  -= l.outputs*l.batch;
-        output_layer.x_gpu  -= l.outputs*l.batch;
-        output_layer.x_norm_gpu  -= l.outputs*l.batch;
+        increment_layer(&input_layer,  -1);
+        increment_layer(&self_layer,   -1);
+        increment_layer(&output_layer, -1);
     }
 }
 #endif