padding_mode for Pooling, fix #261

src/layer/arm/pooling_arm.cpp

@@ -58,7 +58,7 @@ int Pooling_arm::forward(const Mat& bottom_blob, Mat& top_blob) const
         w = bottom_blob_bordered.w;
         h = bottom_blob_bordered.h;
     }
-    else if (pad_w == -233 && pad_h == -233)
+    else if (pad_mode == 2) // tensorflow padding=SAME
     {
         int wpad = kernel_w + (w - 1) / stride_w * stride_w - w;
         int hpad = kernel_h + (h - 1) / stride_h * stride_h - h;
@@ -76,12 +76,12 @@ int Pooling_arm::forward(const Mat& bottom_blob, Mat& top_blob) const
     int outw = (w - kernel_w) / stride_w + 1;
     int outh = (h - kernel_h) / stride_h + 1;
 
-    int wtail = (w - kernel_w) % stride_w;
-    int htail = (h - kernel_h) % stride_h;
-    if ((pad_w == -233 && pad_h == -233) || (pad_w == -2333 && pad_h == -2333))
+    int wtail = 0;
+    int htail = 0;
+    if (pad_mode == 0) // full padding
     {
-        wtail = 0;
-        htail = 0;
+        wtail = (w - kernel_w) % stride_w;
+        htail = (h - kernel_h) % stride_h;
     }
     if (wtail != 0 || htail != 0)
     {

src/layer/pooling.cpp

@@ -35,6 +35,7 @@ int Pooling::load_param(const ParamDict& pd)
     pad_w = pd.get(3, 0);
     pad_h = pd.get(13, pad_w);
     global_pooling = pd.get(4, 0);
+    pad_mode = pd.get(5, 0);
 
     return 0;
 }
@@ -105,7 +106,7 @@ int Pooling::forward(const Mat& bottom_blob, Mat& top_blob) const
         w = bottom_blob_bordered.w;
         h = bottom_blob_bordered.h;
     }
-    else if (pad_w == -233 && pad_h == -233)
+    else if (pad_mode == 2) // tensorflow padding=SAME
     {
         int wpad = kernel_w + (w - 1) / stride_w * stride_w - w;
         int hpad = kernel_h + (h - 1) / stride_h * stride_h - h;
@@ -123,12 +124,12 @@ int Pooling::forward(const Mat& bottom_blob, Mat& top_blob) const
     int outw = (w - kernel_w) / stride_w + 1;
     int outh = (h - kernel_h) / stride_h + 1;
 
-    int wtail = (w - kernel_w) % stride_w;
-    int htail = (h - kernel_h) % stride_h;
-    if ((pad_w == -233 && pad_h == -233) || (pad_w == -2333 && pad_h == -2333))
+    int wtail = 0;
+    int htail = 0;
+    if (pad_mode == 0) // full padding
     {
-        wtail = 0;
-        htail = 0;
+        wtail = (w - kernel_w) % stride_w;
+        htail = (h - kernel_h) % stride_h;
     }
     if (wtail != 0 || htail != 0)
     {

src/layer/pooling.h

@@ -40,6 +40,7 @@ public:
     int pad_w;
     int pad_h;
     int global_pooling;
+    int pad_mode;// 0=full 1=valid 2=SAME
 };
 
 } // namespace ncnn

tools/mxnet/mxnet2ncnn.cpp

@@ -1092,9 +1092,14 @@ int main(int argc, char** argv)
                 pool = 1;
             }
 
+            int pad_mode = 1;
             if (pooling_convention == "valid")
             {
-                // TODO valid and full mode
+                pad_mode = 1;
+            }
+            else if (pooling_convention == "full")
+            {
+                pad_mode = 0;
             }
 
             fprintf(pp, " 0=%d", pool);
@@ -1105,6 +1110,7 @@ int main(int argc, char** argv)
             if (!pad.empty())
                 fprintf(pp, " 3=%d", pad[0]);
             fprintf(pp, " 4=%d", global_pool);
+            fprintf(pp, " 5=%d", pad_mode);
         }
         else if (n.op == "SliceChannel")
         {

tools/tensorflow/tensorflow2ncnn.cpp

@@ -530,6 +530,7 @@ int main(int argc, char** argv)
             int pad = 0;
 
             int global_pooling = 0;
+            int pad_mode = 1;
 
             tensorflow::AttrValue value_ksize;
             if (find_attr_value(node, "ksize", value_ksize))
@@ -552,11 +553,11 @@ int main(int argc, char** argv)
             {
                 if (value_padding.s() == "VALID")
                 {
-                    pad = 0;
+                    pad_mode = 1;
                 }
                 else if (value_padding.s() == "SAME")
                 {
-                    pad = -233;
+                    pad_mode = 2;
                 }
             }
 
@@ -567,6 +568,7 @@ int main(int argc, char** argv)
             fprintf(pp, " 12=%d", stride_h);
             fprintf(pp, " 3=%d", pad);
             fprintf(pp, " 4=%d", global_pooling);
+            fprintf(pp, " 5=%d", pad_mode);
         }
         else if (node.op() == "Concat" || node.op() == "ConcatV2")
         {
@@ -1075,6 +1077,7 @@ int main(int argc, char** argv)
             int pad = 0;
 
             int global_pooling = 0;
+            int pad_mode = 1;
 
             tensorflow::AttrValue value_ksize;
             if (find_attr_value(node, "ksize", value_ksize))
@@ -1097,11 +1100,11 @@ int main(int argc, char** argv)
             {
                 if (value_padding.s() == "VALID")
                 {
-                    pad = -2333;
+                    pad_mode = 1;
                 }
                 else if (value_padding.s() == "SAME")
                 {
-                    pad = -233;
+                    pad_mode = 2;
                 }
             }
 
@@ -1112,6 +1115,7 @@ int main(int argc, char** argv)
             fprintf(pp, " 12=%d", stride_h);
             fprintf(pp, " 3=%d", pad);
             fprintf(pp, " 4=%d", global_pooling);
+            fprintf(pp, " 5=%d", pad_mode);
         }
         else if (node.op() == "Min" || node.op() == "Minimum")
         {