Added groups= and groupd_id= params to the [route] layer

include/darknet.h

@@ -206,6 +206,7 @@ struct layer {
     int n;
     int max_boxes;
     int groups;
+    int group_id;
     int size;
     int side;
     int stride;

src/parser.c

@@ -776,7 +776,10 @@ route_layer parse_route(list *options, size_params params)
     }
     int batch = params.batch;
 
-    route_layer layer = make_route_layer(batch, n, layers, sizes);
+    int groups = option_find_int_quiet(options, "groups", 1);
+    int group_id = option_find_int_quiet(options, "group_id", 0);
+
+    route_layer layer = make_route_layer(batch, n, layers, sizes, groups, group_id);
 
     convolutional_layer first = params.net.layers[layers[0]];
     layer.out_w = first.out_w;
@@ -791,6 +794,7 @@ route_layer parse_route(list *options, size_params params)
             layer.out_h = layer.out_w = layer.out_c = 0;
         }
     }
+    layer.out_c = layer.out_c / layer.groups;
 
     return layer;
 }

src/route_layer.c

@@ -3,7 +3,7 @@
 #include "blas.h"
 #include <stdio.h>
 
-route_layer make_route_layer(int batch, int n, int *input_layers, int *input_sizes)
+route_layer make_route_layer(int batch, int n, int *input_layers, int *input_sizes, int groups, int group_id)
 {
     fprintf(stderr,"route ");
     route_layer l = { (LAYER_TYPE)0 };
@@ -12,6 +12,8 @@ route_layer make_route_layer(int batch, int n, int *input_layers, int *input_siz
     l.n = n;
     l.input_layers = input_layers;
     l.input_sizes = input_sizes;
+    l.groups = groups;
+    l.group_id = group_id;
     int i;
     int outputs = 0;
     for(i = 0; i < n; ++i){
@@ -19,6 +21,7 @@ route_layer make_route_layer(int batch, int n, int *input_layers, int *input_siz
         outputs += input_sizes[i];
     }
     fprintf(stderr, "\n");
+    outputs = outputs / groups;
     l.outputs = outputs;
     l.inputs = outputs;
     l.delta = (float*)calloc(outputs * batch, sizeof(float));
@@ -57,6 +60,8 @@ void resize_route_layer(route_layer *l, network *net)
             l->out_h = l->out_w = l->out_c = 0;
         }
     }
+    l->out_c = l->out_c / l->groups;
+    l->outputs = l->outputs / l->groups;
     l->inputs = l->outputs;
     l->delta = (float*)realloc(l->delta, l->outputs * l->batch * sizeof(float));
     l->output = (float*)realloc(l->output, l->outputs * l->batch * sizeof(float));
@@ -78,10 +83,13 @@ void forward_route_layer(const route_layer l, network_state state)
         int index = l.input_layers[i];
         float *input = state.net.layers[index].output;
         int input_size = l.input_sizes[i];
+        int part_input_size = input_size / l.groups;
         for(j = 0; j < l.batch; ++j){
-            copy_cpu(input_size, input + j*input_size, 1, l.output + offset + j*l.outputs, 1);
+            //copy_cpu(input_size, input + j*input_size, 1, l.output + offset + j*l.outputs, 1);
+            copy_cpu(part_input_size, input + j*input_size + part_input_size*l.group_id, 1, l.output + offset + j*l.outputs, 1);
         }
-        offset += input_size;
+        //offset += input_size;
+        offset += part_input_size;
     }
 }
 
@@ -93,10 +101,13 @@ void backward_route_layer(const route_layer l, network_state state)
         int index = l.input_layers[i];
         float *delta = state.net.layers[index].delta;
         int input_size = l.input_sizes[i];
+        int part_input_size = input_size / l.groups;
         for(j = 0; j < l.batch; ++j){
-            axpy_cpu(input_size, 1, l.delta + offset + j*l.outputs, 1, delta + j*input_size, 1);
+            //axpy_cpu(input_size, 1, l.delta + offset + j*l.outputs, 1, delta + j*input_size, 1);
+            axpy_cpu(part_input_size, 1, l.delta + offset + j*l.outputs, 1, delta + j*input_size + part_input_size*l.group_id, 1);
         }
-        offset += input_size;
+        //offset += input_size;
+        offset += part_input_size;
     }
 }
 
@@ -109,11 +120,14 @@ void forward_route_layer_gpu(const route_layer l, network_state state)
         int index = l.input_layers[i];
         float *input = state.net.layers[index].output_gpu;
         int input_size = l.input_sizes[i];
+        int part_input_size = input_size / l.groups;
         for(j = 0; j < l.batch; ++j){
             //copy_ongpu(input_size, input + j*input_size, 1, l.output_gpu + offset + j*l.outputs, 1);
-            simple_copy_ongpu(input_size, input + j*input_size, l.output_gpu + offset + j*l.outputs);
+            //simple_copy_ongpu(input_size, input + j*input_size, l.output_gpu + offset + j*l.outputs);
+            simple_copy_ongpu(part_input_size, input + j*input_size + part_input_size*l.group_id, l.output_gpu + offset + j*l.outputs);
         }
-        offset += input_size;
+        //offset += input_size;
+        offset += part_input_size;
     }
 }
 
@@ -125,10 +139,13 @@ void backward_route_layer_gpu(const route_layer l, network_state state)
         int index = l.input_layers[i];
         float *delta = state.net.layers[index].delta_gpu;
         int input_size = l.input_sizes[i];
+        int part_input_size = input_size / l.groups;
         for(j = 0; j < l.batch; ++j){
-            axpy_ongpu(input_size, 1, l.delta_gpu + offset + j*l.outputs, 1, delta + j*input_size, 1);
+            //axpy_ongpu(input_size, 1, l.delta_gpu + offset + j*l.outputs, 1, delta + j*input_size, 1);
+            axpy_ongpu(part_input_size, 1, l.delta_gpu + offset + j*l.outputs, 1, delta + j*input_size + part_input_size*l.group_id, 1);
         }
-        offset += input_size;
+        //offset += input_size;
+        offset += part_input_size;
     }
 }
 #endif

src/route_layer.h

@@ -8,7 +8,7 @@ typedef layer route_layer;
 #ifdef __cplusplus
 extern "C" {
 #endif
-route_layer make_route_layer(int batch, int n, int *input_layers, int *input_size);
+route_layer make_route_layer(int batch, int n, int *input_layers, int *input_size, int groups, int group_id);
 void forward_route_layer(const route_layer l, network_state state);
 void backward_route_layer(const route_layer l, network_state state);
 void resize_route_layer(route_layer *l, network *net);