Merge pull request #4302 from dccho/master

enable random resize training for efficient net

src/convolutional_layer.c

@@ -750,6 +750,7 @@ void resize_convolutional_layer(convolutional_layer *l, int w, int h)
         //l->binary_input = realloc(l->inputs*l->batch, sizeof(float));
     }
 
+    if (l->activation == SWISH || l->activation == MISH) l->activation_input = (float*)realloc(l->activation_input, total_batch*l->outputs * sizeof(float));
 #ifdef GPU
     if (old_w < w || old_h < h) {
         if (l->train) {
@@ -772,6 +773,11 @@ void resize_convolutional_layer(convolutional_layer *l, int w, int h)
             cuda_free(l->binary_input_gpu);
             l->binary_input_gpu = cuda_make_array(0, l->inputs*l->batch);
         }
+
+        if (l->activation == SWISH || l->activation == MISH) {
+            cuda_free(l->activation_input_gpu);
+            l->activation_input_gpu = cuda_make_array(l->activation_input, total_batch*l->outputs);
+        }
     }
 #ifdef CUDNN
     cudnn_convolutional_setup(l, cudnn_fastest);

src/dropout_layer.c

@@ -27,11 +27,12 @@ dropout_layer make_dropout_layer(int batch, int inputs, float probability)
 
 void resize_dropout_layer(dropout_layer *l, int inputs)
 {
+    l->inputs = l->outputs = inputs;
     l->rand = (float*)realloc(l->rand, l->inputs * l->batch * sizeof(float));
     #ifdef GPU
     cuda_free(l->rand_gpu);
 
-    l->rand_gpu = cuda_make_array(l->rand, inputs*l->batch);
+    l->rand_gpu = cuda_make_array(l->rand, l->inputs*l->batch);
     #endif
 }
 

src/network.c

@@ -535,8 +535,16 @@ int resize_network(network *net, int w, int h)
             resize_route_layer(&l, net);
         }else if (l.type == SHORTCUT) {
             resize_shortcut_layer(&l, w, h);
-        //}else if (l.type == SCALE_CHANNELS) {
+        }else if (l.type == SCALE_CHANNELS) {
-        //    resize_scale_channels_layer(&l, w, h);
+            resize_scale_channels_layer(&l, net);
+        }else if (l.type == DROPOUT) {
+            resize_dropout_layer(&l, inputs);
+            l.output = net->layers[i - 1].output;
+            l.delta = net->layers[i - 1].delta;
+#ifdef GPU
+            l.output_gpu = net->layers[i-1].output_gpu;
+            l.delta_gpu = net->layers[i-1].delta_gpu;
+#endif  
         }else if (l.type == UPSAMPLE) {
             resize_upsample_layer(&l, w, h);
         }else if(l.type == REORG){
@@ -556,9 +564,12 @@ int resize_network(network *net, int w, int h)
         if(l.workspace_size > workspace_size) workspace_size = l.workspace_size;
         inputs = l.outputs;
         net->layers[i] = l;
+        if(l.type != DROPOUT)
+        {
             w = l.out_w;
             h = l.out_h;
-        if(l.type == AVGPOOL) break;
+        }
+        //if(l.type == AVGPOOL) break;
     }
 #ifdef GPU
     const int size = get_network_input_size(*net) * net->batch;

src/scale_channels_layer.c

@@ -39,10 +39,11 @@ layer make_scale_channels_layer(int batch, int index, int w, int h, int c, int w
     return l;
 }
 
-void resize_scale_channels_layer(layer *l, int w, int h)
+void resize_scale_channels_layer(layer *l, network *net)
 {
-    l->out_w = w;
+    layer first = net->layers[l->index];
-    l->out_h = h;
+    l->out_w = first.out_w;
+    l->out_h = first.out_h;
     l->outputs = l->out_w*l->out_h*l->out_c;
     l->inputs = l->outputs;
     l->delta = (float*)realloc(l->delta, l->outputs * l->batch * sizeof(float));

src/scale_channels_layer.h

@@ -10,7 +10,7 @@ extern "C" {
 layer make_scale_channels_layer(int batch, int index, int w, int h, int c, int w2, int h2, int c2);
 void forward_scale_channels_layer(const layer l, network_state state);
 void backward_scale_channels_layer(const layer l, network_state state);
-void resize_scale_channels_layer(layer *l, int w, int h);
+void resize_scale_channels_layer(layer *l, network *net);
 
 #ifdef GPU
 void forward_scale_channels_layer_gpu(const layer l, network_state state);