fixed backward pass in cont layer to accumulate gradients this will pass the layer test now also removed compile warnings and changed some comments

dlib/dnn/cpu_dlib.h

@@ -355,14 +355,16 @@ namespace dlib
             ) {}
 
             void setup(
-                const tensor& data,
-                const tensor& filters,
+                const tensor& data,    /* not used but required for interface */
+                const tensor& filters, /* not used but required for interface */
                 int stride_y,
                 int stride_x,
                 int padding_y,
                 int padding_x
             ) 
             {
+                (void)data;    /* silence compiler */
+                (void)filters; /* silence compiler */
                 last_stride_y = stride_y;
                 last_stride_x = stride_x;
                 last_padding_y = padding_y;

dlib/dnn/layers.h

@@ -454,9 +454,9 @@ namespace dlib
             temp.copy_size(sub.get_gradient_input());
             auto filt = filters(params,0);           
             conv(temp,gradient_input, filt);
-            tt::copy_tensor(sub.get_gradient_input(),0,temp,0,sub.get_gradient_input().k());
+            // need to add the new gradients on top of the previous ones
+            tt::add(1,sub.get_gradient_input(),1,temp);
             // no point computing the parameter gradients if they won't be used.
-            
             if (learning_rate_multiplier != 0)
             {
                 auto filt = filters(params_grad,0);                
@@ -582,8 +582,6 @@ namespace dlib
         double bias_learning_rate_multiplier;
         double bias_weight_decay_multiplier;
 
-        // These are here only because older versions of con (which you might encounter
-        // serialized to disk) used different padding settings.
         int padding_y_;
         int padding_x_;
 

dlib/dnn/layers_abstract.h

@@ -850,16 +850,17 @@ namespace dlib
 
             WHAT THIS OBJECT REPRESENTS
                 This is an implementation of the EXAMPLE_COMPUTATIONAL_LAYER_ interface
-                defined above.  In particular, it defines a convolution layer that takes an
-                input tensor (nominally representing an image) and convolves it with a set
-                of filters and then outputs the results. 
-
+                defined above.  In particular, it defines a transposed convolution layer 
+                that takes an input tensor (nominally representing an image) and 
+                transpose convolves (deconvolves) it with a set of filters and then outputs the results. 
+                This is basically a convolutional layer with reversed forward/backward passes
+                
                 The dimensions of the tensors output by this layer are as follows (letting
                 IN be the input tensor and OUT the output tensor):
                     - OUT.num_samples() == IN.num_samples()
                     - OUT.k()  == num_filters()
-                    - OUT.nr() == 1+(IN.nr() + 2*padding_y() - nr())/stride_y()
-                    - OUT.nc() == 1+(IN.nc() + 2*padding_x() - nc())/stride_x()
+                    - OUT.nr() == stride_y * (IN.nr() -1) + nr) - 2*padding_y
+                    - OUT.nc() == stride_x * (IN.nc() -1) + nc) - 2*padding_x
         !*/
 
     public:

dlib/test/dnn.cpp

@@ -1410,6 +1410,18 @@ namespace
             auto res = test_layer(l);
             DLIB_TEST_MSG(res, res);
         }
+        {
+            print_spinner();
+            cont_<3,3,3,1,1> l;
+            auto res = test_layer(l);
+            DLIB_TEST_MSG(res, res);
+        }
+        {
+            print_spinner();
+            cont_<3,2,2,2,2> l;
+            auto res = test_layer(l);
+            DLIB_TEST_MSG(res, res);
+        }
         {
             print_spinner();
             con_<3,2,2,2,2> l;