Made the converter add zero padding layers when needed by Eltwise to replicate

the behavior of dlib's add_prev layers.

tools/convert_dlib_nets_to_caffe/main.cpp

@@ -21,6 +21,8 @@ struct layer
     string type; // comp, loss, or input
     int idx;
 
+    matrix<long,4,1> output_tensor_shape; // (N,K,NR,NC)
+
     string detail_name; // The name of the tag inside the layer tag. e.g. fc, con, max_pool, input_rgb_image.
     std::map<string,double> attributes;
     matrix<double> params;
@@ -49,30 +51,32 @@ struct layer
 // ----------------------------------------------------------------------------------------
 
 std::vector<layer> parse_dlib_xml(
+    const matrix<long,4,1>& input_tensor_shape, 
     const string& xml_filename
 );
 
 // ----------------------------------------------------------------------------------------
 
 template <typename iterator>
-string find_layer_caffe_name (
+const layer& find_layer (
     iterator i,
     long tag_id
 )
 /*!
     requires
-        - i is an iterator pointing to a layer in the list of layers produced by parse_dlib_xml().
+        - i is a reverse iterator pointing to a layer in the list of layers produced by parse_dlib_xml().
         - i is not an input layer.
     ensures
         - if (tag_id == -1) then
-            - returns the caffe string name for the previous layer to layer i.
+            - returns the previous layer (i.e. closer to the input) to layer i.
         - else
-            - returns the caffe string name for the previous layer to layer i with the given tag_id.
+            - returns the previous layer (i.e. closer to the input) to layer i with the
+              given tag_id.
 !*/
 {
     if (tag_id == -1)
     {
-        return (i-1)->caffe_layer_name();
+        return *(i-1);
     }
     else
     {
@@ -81,7 +85,7 @@ string find_layer_caffe_name (
             i--;
             // if we hit the end of the network before we found what we were looking for
             if (i->tag_id == tag_id)
-                return i->caffe_layer_name();
+                return *i;
             if (i->type == "input")
                 throw dlib::error("Network definition is bad, a layer wanted to skip back to a non-existing layer.");
         }
@@ -89,7 +93,19 @@ string find_layer_caffe_name (
 }
 
 template <typename iterator>
-string find_input_layer_caffe_name (iterator i) { return find_layer_caffe_name(i, i->skip_id); }
+const layer& find_input_layer (iterator i) { return find_layer(i, i->skip_id); }
+
+template <typename iterator>
+string find_layer_caffe_name (
+    iterator i,
+    long tag_id
+)
+{
+    return find_layer(i,tag_id).caffe_layer_name();
+}
+
+template <typename iterator>
+string find_input_layer_caffe_name (iterator i) { return find_input_layer(i).caffe_layer_name(); }
 
 // ----------------------------------------------------------------------------------------
 
@@ -116,7 +132,8 @@ void convert_dlib_xml_to_caffe_python_code(
     cout << "Writing model to " << out_filename << endl;
     ofstream fout(out_filename);
     fout.precision(9);
-    const auto layers = parse_dlib_xml(xml_filename);
+    const auto layers = parse_dlib_xml({N,K,NR,NC}, xml_filename);
+
 
     fout << "#\n";
     fout << "# !!! This file was automatically generated by dlib's tools/convert_dlib_nets_to_caffe utility.     !!!\n";
@@ -300,12 +317,61 @@ void convert_dlib_xml_to_caffe_python_code(
             fout << ");\n";
         }
         else if (i->detail_name == "add_prev")
+        {
+            auto in_shape1 = find_input_layer(i).output_tensor_shape;
+            auto in_shape2 = find_layer(i,i->attribute("tag")).output_tensor_shape;
+            if (in_shape1 != in_shape2)
+            {
+                // if only the number of channels differs then we will use a dummy layer to
+                // pad with zeros.  But otherwise we will throw an error.
+                if (in_shape1(0) == in_shape2(0) && 
+                    in_shape1(2) == in_shape2(2) && 
+                    in_shape1(3) == in_shape2(3))
+                {
+                    fout << "    n." << i->caffe_layer_name() << "_zeropad = L.DummyData(num=" << in_shape1(0);
+                    fout << ", channels="<<std::abs(in_shape1(1)-in_shape2(1));
+                    fout << ", height="<<in_shape1(2);
+                    fout << ", width="<<in_shape1(3);
+                    fout << ");\n";
+
+                    string smaller_layer = find_input_layer_caffe_name(i);
+                    string bigger_layer = find_layer_caffe_name(i, i->attribute("tag"));
+                    if (in_shape1(1) > in_shape2(1))
+                        swap(smaller_layer, bigger_layer);
+
+                    fout << "    n." << i->caffe_layer_name() << "_concat = L.Concat(n." << smaller_layer;
+                    fout << ", n." << i->caffe_layer_name() << "_zeropad";
+                    fout << ");\n";
+
+                    fout << "    n." << i->caffe_layer_name() << " = L.Eltwise(n." << i->caffe_layer_name() << "_concat";
+                    fout << ", n." << bigger_layer;
+                    fout << ", operation=P.Eltwise.SUM";
+                    fout << ");\n";
+                }
+                else
+                {
+                    std::ostringstream sout;
+                    sout << "The dlib network contained an add_prev layer (layer idx " << i->idx << ") that adds two previous ";
+                    sout << "layers with different output tensor dimensions.  Caffe's equivalent layer, Eltwise, doesn't support ";
+                    sout << "adding layers together with different dimensions.  In the special case where the only difference is "; 
+                    sout << "in the number of channels, this converter program will add a dummy layer that outputs a tensor full of zeros ";
+                    sout << "and concat it appropriately so this will work.  However, this network you are converting has tensor dimensions ";
+                    sout << "different in values other than the number of channels.  In particular, here are the two tensor shapes (batch size, channels, rows, cols): ";
+                    std::ostringstream sout2;
+                    sout2 << wrap_string(sout.str()) << endl;
+                    sout2 << trans(in_shape1);
+                    sout2 << trans(in_shape2);
+                    throw dlib::error(sout2.str());
+                }
+            }
+            else
             {
                 fout << "    n." << i->caffe_layer_name() << " = L.Eltwise(n." << find_input_layer_caffe_name(i);
                 fout << ", n." << find_layer_caffe_name(i, i->attribute("tag"));
                 fout << ", operation=P.Eltwise.SUM";
                 fout << ");\n";
             }
+        }
         else
         {
             throw dlib::error("No known transformation from dlib's " + i->detail_name + " layer to caffe.");
@@ -549,7 +615,68 @@ public:
 
 // ----------------------------------------------------------------------------------------
 
+void compute_output_tensor_shapes(const matrix<long,4,1>& input_tensor_shape, std::vector<layer>& layers)
+{
+    DLIB_CASSERT(layers.back().type == "input");
+    layers.back().output_tensor_shape = input_tensor_shape;
+    for (auto i = ++layers.rbegin(); i != layers.rend(); ++i)
+    {
+        const auto input_shape = find_input_layer(i).output_tensor_shape;
+        if (i->type == "comp")
+        {
+            if (i->detail_name == "fc" || i->detail_name == "fc_no_bias")
+            {
+                long num_outputs = i->attribute("num_outputs");
+                i->output_tensor_shape = {input_shape(0), num_outputs, 1, 1};
+            }
+            else if (i->detail_name == "con")
+            {
+                long num_filters = i->attribute("num_filters");
+                long filter_nc = i->attribute("nc");
+                long filter_nr = i->attribute("nr");
+                long stride_x = i->attribute("stride_x");
+                long stride_y = i->attribute("stride_y");
+                long padding_x = i->attribute("padding_x");
+                long padding_y = i->attribute("padding_y");
+                long nr = 1+(input_shape(2) + 2*padding_y - filter_nr)/stride_y;
+                long nc = 1+(input_shape(3) + 2*padding_x - filter_nc)/stride_x;
+                i->output_tensor_shape = {input_shape(0), num_filters, nr, nc};
+            }
+            else if (i->detail_name == "max_pool" || i->detail_name == "avg_pool")
+            {
+                long filter_nc = i->attribute("nc");
+                long filter_nr = i->attribute("nr");
+                long stride_x = i->attribute("stride_x");
+                long stride_y = i->attribute("stride_y");
+                long padding_x = i->attribute("padding_x");
+                long padding_y = i->attribute("padding_y");
+                long nr = 1+(input_shape(2) + 2*padding_y - filter_nr)/stride_y;
+                long nc = 1+(input_shape(3) + 2*padding_x - filter_nc)/stride_x;
+                i->output_tensor_shape = {input_shape(0), input_shape(1), nr, nc};
+            }
+            else if (i->detail_name == "add_prev")
+            {
+                auto aux_shape = find_layer(i, i->attribute("tag")).output_tensor_shape;
+                for (long j = 0; j < input_shape.size(); ++j)
+                    i->output_tensor_shape(j) = std::max(input_shape(j), aux_shape(j));
+            }
+            else
+            {
+                i->output_tensor_shape = input_shape;
+            }
+        }
+        else
+        {
+            i->output_tensor_shape = input_shape;
+        }
+
+    }
+}
+
+// ----------------------------------------------------------------------------------------
+
 std::vector<layer> parse_dlib_xml(
+    const matrix<long,4,1>& input_tensor_shape, 
     const string& xml_filename
 )
 {
@@ -561,6 +688,8 @@ std::vector<layer> parse_dlib_xml(
     if (dh.layers.back().type != "input")
         throw dlib::error("The network in the XML file is missing an input layer!");
 
+    compute_output_tensor_shapes(input_tensor_shape, dh.layers);
+
     return dh.layers;
 }