Merge branch 'master' of git://github.com/OranjeeGeneral/dlib into OranjeeGeneral-master

dlib/dnn/cpu_dlib.cpp

@@ -1739,43 +1739,32 @@ namespace dlib
             }
         }
 
+
         void tensor_conv::operator() (
             resizable_tensor& output,
             const tensor& data,
-            const tensor& filters,
-            int stride_y,
-            int stride_x,
-            int padding_y,
-            int padding_x
+            const tensor& filters
         )
         {
             DLIB_CASSERT(is_same_object(output,data) == false);
             DLIB_CASSERT(is_same_object(output,filters) == false);
             DLIB_CASSERT(filters.k() == data.k());
-            DLIB_CASSERT(stride_y > 0 && stride_x > 0);
-            DLIB_CASSERT(0 <= padding_y && padding_y < filters.nr());
-            DLIB_CASSERT(0 <= padding_x && padding_x < filters.nc());
-            DLIB_CASSERT(filters.nr() <= data.nr() + 2*padding_y,
+            DLIB_CASSERT(filters.nr() <= data.nr() + 2*last_padding_y,
                 "Filter windows must be small enough to fit into the padded image.");
-            DLIB_CASSERT(filters.nc() <= data.nc() + 2*padding_x,
+            DLIB_CASSERT(filters.nc() <= data.nc() + 2*last_padding_x,
                 "Filter windows must be small enough to fit into the padded image.");
 
             output.set_size(data.num_samples(),
                             filters.num_samples(),
-                            1+(data.nr()+2*padding_y-filters.nr())/stride_y,
-                            1+(data.nc()+2*padding_x-filters.nc())/stride_x);
+                            1+(data.nr()+2*last_padding_y-filters.nr())/last_stride_y,
+                            1+(data.nc()+2*last_padding_x-filters.nc())/last_stride_x);
 
             matrix<float> temp;
             for (long n = 0; n < data.num_samples(); ++n)
             {
-                img2col(temp, data, n, filters.nr(), filters.nc(), stride_y, stride_x, padding_y, padding_x);
+                img2col(temp, data, n, filters.nr(), filters.nc(), last_stride_y, last_stride_x, last_padding_y, last_padding_x);
                 output.set_sample(n, mat(filters)*trans(temp));
             }
-
-            last_stride_y = stride_y;
-            last_stride_x = stride_x;
-            last_padding_y = padding_y;
-            last_padding_x = padding_x;
         }
 
     // ------------------------------------------------------------------------------------
@@ -1824,7 +1813,7 @@ namespace dlib
                     filters_gradient += gi*temp;
             }
         }
-    // ------------------------------------------------------------------------------------
+     // ------------------------------------------------------------------------------------
     void copy_tensor(
             tensor& dest,
             size_t dest_k_offset,

dlib/dnn/cpu_dlib.h

@@ -368,14 +368,29 @@ namespace dlib
             void clear(
             ) {}
 
-            void operator() (
-                resizable_tensor& output,
-                const tensor& data,
-                const tensor& filters,
+            void setup(
+                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 */
+                DLIB_CASSERT(stride_y > 0 && stride_x > 0);
+                DLIB_CASSERT(0 <= padding_y && padding_y < filters.nr());
+                DLIB_CASSERT(0 <= padding_x && padding_x < filters.nc());
+                last_stride_y = stride_y;
+                last_stride_x = stride_x;
+                last_padding_y = padding_y;
+                last_padding_x = padding_x;            
+            }
+
+             void operator() (
+                resizable_tensor& output,
+                const tensor& data,
+                const tensor& filters
             );
 
             void get_gradient_for_data (

dlib/dnn/cudnn_dlibapi.cpp

@@ -953,11 +953,7 @@ namespace dlib
         void tensor_conv::operator() (
             resizable_tensor& output,
             const tensor& data,
-            const tensor& filters,
-            int stride_y,
-            int stride_x,
-            int padding_y,
-            int padding_x
+            const tensor& filters
         )
         {
             DLIB_CASSERT(is_same_object(output,data) == false);
@@ -978,8 +974,6 @@ namespace dlib
                 );
 
 
-            setup(data,filters,stride_y,stride_x,padding_y,padding_x);
-
             output.set_size(out_num_samples, out_k, out_nr, out_nc);
 
             DLIB_ASSERT(output.num_samples() == data.num_samples(),out_num_samples << "  " << data.num_samples());

dlib/dnn/cudnn_dlibapi.h

@@ -205,11 +205,7 @@ namespace dlib
             void operator() (
                 resizable_tensor& output,
                 const tensor& data,
-                const tensor& filters,
-                int stride_y,
-                int stride_x,
-                int padding_y,
-                int padding_x
+                const tensor& filters
             );
             /*!
                 requires
@@ -270,9 +266,7 @@ namespace dlib
                       and assigns this gradient to filters_gradient.
             !*/
 
-        private:
-
-            void setup(
+           void setup(
                 const tensor& data,
                 const tensor& filters,
                 int stride_y,
@@ -280,6 +274,9 @@ namespace dlib
                 int padding_y,
                 int padding_x
             );
+
+        private:
+
             /*!
                 requires
                     - filters.k() == data.k()

dlib/dnn/layers.h

@@ -147,14 +147,16 @@ namespace dlib
         template <typename SUBNET>
         void forward(const SUBNET& sub, resizable_tensor& output)
         {
+            conv.setup(sub.get_output(),
+                       filters(params,0),
+                       _stride_y,
+                       _stride_x,
+                       padding_y_,
+                       padding_x_);
+            
             conv(output,
                 sub.get_output(),
-                filters(params,0),
-                _stride_y,
-                _stride_x,
-                padding_y_,
-                padding_x_
-                );
+                filters(params,0));
 
             tt::add(1,output,1,biases(params,filters.size()));
         } 
@@ -306,6 +308,297 @@ namespace dlib
         >
     using con = add_layer<con_<num_filters,nr,nc,stride_y,stride_x>, SUBNET>;
 
+// ----------------------------------------------------------------------------------------
+
+    template <
+        long _num_filters,
+        long _nr,
+        long _nc,
+        int _stride_y,
+        int _stride_x,
+        int _padding_y = _stride_y!=1? 0 : _nr/2,
+        int _padding_x = _stride_x!=1? 0 : _nc/2
+        >
+    class cont_
+    {
+    public:
+
+        static_assert(_num_filters > 0, "The number of filters must be > 0");
+        static_assert(_nr > 0, "The number of rows in a filter must be > 0");
+        static_assert(_nc > 0, "The number of columns in a filter must be > 0");
+        static_assert(_stride_y > 0, "The filter stride must be > 0");
+        static_assert(_stride_x > 0, "The filter stride must be > 0");
+        static_assert(0 <= _padding_y && _padding_y < _nr, "The padding must be smaller than the filter size.");
+        static_assert(0 <= _padding_x && _padding_x < _nc, "The padding must be smaller than the filter size.");
+
+        cont_(
+        ) : 
+            learning_rate_multiplier(1),
+            weight_decay_multiplier(1),
+            bias_learning_rate_multiplier(1),
+            bias_weight_decay_multiplier(0),
+            padding_y_(_padding_y),
+            padding_x_(_padding_x)
+        {
+        }
+
+        long num_filters() const { return _num_filters; }
+        long nr() const { return _nr; }
+        long nc() const { return _nc; }
+        long stride_y() const { return _stride_y; }
+        long stride_x() const { return _stride_x; }
+        long padding_y() const { return padding_y_; }
+        long padding_x() const { return padding_x_; }
+
+        double get_learning_rate_multiplier () const  { return learning_rate_multiplier; }
+        double get_weight_decay_multiplier () const   { return weight_decay_multiplier; }
+        void set_learning_rate_multiplier(double val) { learning_rate_multiplier = val; }
+        void set_weight_decay_multiplier(double val)  { weight_decay_multiplier  = val; }
+
+        double get_bias_learning_rate_multiplier () const  { return bias_learning_rate_multiplier; }
+        double get_bias_weight_decay_multiplier () const   { return bias_weight_decay_multiplier; }
+        void set_bias_learning_rate_multiplier(double val) { bias_learning_rate_multiplier = val; }
+        void set_bias_weight_decay_multiplier(double val)  { bias_weight_decay_multiplier  = val; }
+
+        inline point map_output_to_input (
+            point p
+        ) const
+        {
+            p.x() = (p.x()+padding_x()-nc()/2)/stride_x();
+            p.y() = (p.y()+padding_y()-nr()/2)/stride_y();
+            return p;
+        }
+
+        inline point map_input_to_output (
+            point p
+        ) const
+        {
+            p.x() = p.x()*stride_x() - padding_x() + nc()/2;
+            p.y() = p.y()*stride_y() - padding_y() + nr()/2;
+            return p;
+        }
+
+        cont_ (
+            const cont_& item
+        ) : 
+            params(item.params),
+            filters(item.filters),
+            biases(item.biases),
+            learning_rate_multiplier(item.learning_rate_multiplier),
+            weight_decay_multiplier(item.weight_decay_multiplier),
+            bias_learning_rate_multiplier(item.bias_learning_rate_multiplier),
+            bias_weight_decay_multiplier(item.bias_weight_decay_multiplier),
+            padding_y_(item.padding_y_),
+            padding_x_(item.padding_x_)
+        {
+            // this->conv is non-copyable and basically stateless, so we have to write our
+            // own copy to avoid trying to copy it and getting an error.
+        }
+
+        cont_& operator= (
+            const cont_& item
+        )
+        {
+            if (this == &item)
+                return *this;
+
+            // this->conv is non-copyable and basically stateless, so we have to write our
+            // own copy to avoid trying to copy it and getting an error.
+            params = item.params;
+            filters = item.filters;
+            biases = item.biases;
+            padding_y_ = item.padding_y_;
+            padding_x_ = item.padding_x_;
+            learning_rate_multiplier = item.learning_rate_multiplier;
+            weight_decay_multiplier = item.weight_decay_multiplier;
+            bias_learning_rate_multiplier = item.bias_learning_rate_multiplier;
+            bias_weight_decay_multiplier = item.bias_weight_decay_multiplier;
+            return *this;
+        }
+
+        template <typename SUBNET>
+        void setup (const SUBNET& sub)
+        {
+            long num_inputs = _nr*_nc*sub.get_output().k();
+            long num_outputs = _num_filters;
+            // allocate params for the filters and also for the filter bias values.
+            params.set_size(num_inputs*_num_filters + _num_filters);
+
+            dlib::rand rnd(std::rand());
+            randomize_parameters(params, num_inputs+num_outputs, rnd);
+
+            filters = alias_tensor(sub.get_output().k(),_num_filters, _nr, _nc);
+            biases = alias_tensor(1,_num_filters);
+
+            // set the initial bias values to zero
+            biases(params,_num_filters) = 0;
+        }
+
+        template <typename SUBNET>
+        void forward(const SUBNET& sub, resizable_tensor& output)
+        {
+            auto filt = filters(params,0);
+            unsigned int gnr = _stride_y * (sub.get_output().nr() - 1) + filt.nr() - 2 * padding_y_;
+            unsigned int gnc = _stride_x * (sub.get_output().nc() - 1) + filt.nc() - 2 * padding_x_;
+            unsigned int gnsamps = sub.get_output().num_samples();
+            unsigned int gk = filt.k();
+            output.set_size(gnsamps,gk,gnr,gnc);
+            output = 0;
+            conv.setup(output,filt,_stride_y,_stride_x,padding_y_,padding_x_);
+            conv.get_gradient_for_data(sub.get_output(),filt,output);            
+            tt::add(1,output,1,biases(params,filters.size()));
+        } 
+
+        template <typename SUBNET>
+        void backward(const tensor& gradient_input, SUBNET& sub, tensor& params_grad)
+        {
+            resizable_tensor temp;
+            temp.copy_size(sub.get_gradient_input());
+            auto filt = filters(params,0);           
+            conv(temp,gradient_input, filt);
+            // 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);                
+                conv.get_gradient_for_filters (sub.get_output(),gradient_input, filt);
+                auto b = biases(params_grad, filters.size());
+                tt::assign_conv_bias_gradient(b, gradient_input);
+            }
+        }
+
+        const tensor& get_layer_params() const { return params; }
+        tensor& get_layer_params() { return params; }
+
+        friend void serialize(const cont_& item, std::ostream& out)
+        {
+            serialize("cont_1", out);
+            serialize(item.params, out);
+            serialize(_num_filters, out);
+            serialize(_nr, out);
+            serialize(_nc, out);
+            serialize(_stride_y, out);
+            serialize(_stride_x, out);
+            serialize(item.padding_y_, out);
+            serialize(item.padding_x_, out);
+            serialize(item.filters, out);
+            serialize(item.biases, out);
+            serialize(item.learning_rate_multiplier, out);
+            serialize(item.weight_decay_multiplier, out);
+            serialize(item.bias_learning_rate_multiplier, out);
+            serialize(item.bias_weight_decay_multiplier, out);
+        }
+
+        friend void deserialize(cont_& item, std::istream& in)
+        {
+            std::string version;
+            deserialize(version, in);
+            long num_filters;
+            long nr;
+            long nc;
+            int stride_y;
+            int stride_x;
+            if (version == "cont_1")
+            {
+                deserialize(item.params, in);
+                deserialize(num_filters, in);
+                deserialize(nr, in);
+                deserialize(nc, in);
+                deserialize(stride_y, in);
+                deserialize(stride_x, in);
+                deserialize(item.padding_y_, in);
+                deserialize(item.padding_x_, in);
+                deserialize(item.filters, in);
+                deserialize(item.biases, in);
+                deserialize(item.learning_rate_multiplier, in);
+                deserialize(item.weight_decay_multiplier, in);
+                deserialize(item.bias_learning_rate_multiplier, in);
+                deserialize(item.bias_weight_decay_multiplier, in);
+                if (item.padding_y_ != _padding_y) throw serialization_error("Wrong padding_y found while deserializing dlib::con_");
+                if (item.padding_x_ != _padding_x) throw serialization_error("Wrong padding_x found while deserializing dlib::con_");
+                if (num_filters != _num_filters) 
+                {
+                    std::ostringstream sout;
+                    sout << "Wrong num_filters found while deserializing dlib::con_" << std::endl;
+                    sout << "expected " << _num_filters << " but found " << num_filters << std::endl;
+                    throw serialization_error(sout.str());
+                }
+
+                if (nr != _nr) throw serialization_error("Wrong nr found while deserializing dlib::con_");
+                if (nc != _nc) throw serialization_error("Wrong nc found while deserializing dlib::con_");
+                if (stride_y != _stride_y) throw serialization_error("Wrong stride_y found while deserializing dlib::con_");
+                if (stride_x != _stride_x) throw serialization_error("Wrong stride_x found while deserializing dlib::con_");
+            }
+            else
+            {
+                throw serialization_error("Unexpected version '"+version+"' found while deserializing dlib::con_.");
+            }
+        }
+
+
+        friend std::ostream& operator<<(std::ostream& out, const cont_& item)
+        {
+            out << "cont\t ("
+                << "num_filters="<<_num_filters
+                << ", nr="<<_nr
+                << ", nc="<<_nc
+                << ", stride_y="<<_stride_y
+                << ", stride_x="<<_stride_x
+                << ", padding_y="<<item.padding_y_
+                << ", padding_x="<<item.padding_x_
+                << ")";
+            out << " learning_rate_mult="<<item.learning_rate_multiplier;
+            out << " weight_decay_mult="<<item.weight_decay_multiplier;
+            out << " bias_learning_rate_mult="<<item.bias_learning_rate_multiplier;
+            out << " bias_weight_decay_mult="<<item.bias_weight_decay_multiplier;
+            return out;
+        }
+
+        friend void to_xml(const cont_& item, std::ostream& out)
+        {
+            out << "<cont"
+                << " num_filters='"<<_num_filters<<"'"
+                << " nr='"<<_nr<<"'"
+                << " nc='"<<_nc<<"'"
+                << " stride_y='"<<_stride_y<<"'"
+                << " stride_x='"<<_stride_x<<"'"
+                << " padding_y='"<<item.padding_y_<<"'"
+                << " padding_x='"<<item.padding_x_<<"'"
+                << " learning_rate_mult='"<<item.learning_rate_multiplier<<"'"
+                << " weight_decay_46mult='"<<item.weight_decay_multiplier<<"'"
+                << " bias_learning_rate_mult='"<<item.bias_learning_rate_multiplier<<"'"
+                << " bias_weight_decay_mult='"<<item.bias_weight_decay_multiplier<<"'>\n";
+            out << mat(item.params);
+            out << "</cont>";
+        }
+
+    private:
+
+        resizable_tensor params;
+        alias_tensor filters, biases;
+
+        tt::tensor_conv conv;
+        double learning_rate_multiplier;
+        double weight_decay_multiplier;
+        double bias_learning_rate_multiplier;
+        double bias_weight_decay_multiplier;
+
+        int padding_y_;
+        int padding_x_;
+
+    };
+
+    template <
+        long num_filters,
+        long nr,
+        long nc,
+        int stride_y,
+        int stride_x,
+        typename SUBNET
+        >
+    using cont = add_layer<cont_<num_filters,nr,nc,stride_y,stride_x>, SUBNET>;
+
 // ----------------------------------------------------------------------------------------
 
     template <

dlib/dnn/layers_abstract.h

@@ -840,6 +840,217 @@ namespace dlib
         >
     using con = add_layer<con_<num_filters,nr,nc,stride_y,stride_x>, SUBNET>;
 
+// ----------------------------------------------------------------------------------------
+
+// ----------------------------------------------------------------------------------------
+
+    template <
+        long _num_filters,
+        long _nr,
+        long _nc,
+        int _stride_y,
+        int _stride_x,
+        int _padding_y = _stride_y!=1? 0 : _nr/2,
+        int _padding_x = _stride_x!=1? 0 : _nc/2
+        >
+    class cont_
+    {
+        /*!
+            REQUIREMENTS ON TEMPLATE ARGUMENTS
+                All of them must be > 0.
+                Also, we require that:
+                    - 0 <= _padding_y && _padding_y < _nr
+                    - 0 <= _padding_x && _padding_x < _nc
+
+            WHAT THIS OBJECT REPRESENTS
+                This is an implementation of the EXAMPLE_COMPUTATIONAL_LAYER_ interface
+                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() == stride_y * (IN.nr() -1) + nr) - 2*padding_y
+                    - OUT.nc() == stride_x * (IN.nc() -1) + nc) - 2*padding_x
+        !*/
+
+    public:
+        cont_(
+        );
+        /*!
+            ensures
+                - #num_filters() == _num_filters
+                - #nr() == _nr
+                - #nc() == _nc
+                - #stride_y() == _stride_y
+                - #stride_x() == _stride_x
+                - #padding_y() == _padding_y
+                - #padding_x() == _padding_x
+                - #get_learning_rate_multiplier()      == 1
+                - #get_weight_decay_multiplier()       == 1
+                - #get_bias_learning_rate_multiplier() == 1
+                - #get_bias_weight_decay_multiplier()  == 0
+        !*/
+
+        long num_filters(
+        ) const; 
+        /*!
+            ensures
+                - returns the number of filters contained in this layer.  The k dimension
+                  of the output tensors produced by this layer will be equal to the number
+                  of filters.
+        !*/
+
+        long nr(
+        ) const; 
+        /*!
+            ensures
+                - returns the number of rows in the filters in this layer.
+        !*/
+
+        long nc(
+        ) const;
+        /*!
+            ensures
+                - returns the number of columns in the filters in this layer.
+        !*/
+
+        long stride_y(
+        ) const; 
+        /*!
+            ensures
+                - returns the vertical stride used when convolving the filters over an
+                  image.  That is, each filter will be moved stride_y() pixels down at a
+                  time when it moves over the image.
+        !*/
+
+        long stride_x(
+        ) const;
+        /*!
+            ensures
+                - returns the horizontal stride used when convolving the filters over an
+                  image.  That is, each filter will be moved stride_x() pixels right at a
+                  time when it moves over the image.
+        !*/
+
+        long padding_y(
+        ) const; 
+        /*!
+            ensures
+                - returns the number of pixels of zero padding added to the top and bottom
+                  sides of the image.
+        !*/
+
+        long padding_x(
+        ) const; 
+        /*!
+            ensures
+                - returns the number of pixels of zero padding added to the left and right 
+                  sides of the image.
+        !*/
+
+        double get_learning_rate_multiplier(
+        ) const;  
+        /*!
+            ensures
+                - returns a multiplier number.  The interpretation is that this object is
+                  requesting that the learning rate used to optimize its parameters be
+                  multiplied by get_learning_rate_multiplier().
+        !*/
+
+        double get_weight_decay_multiplier(
+        ) const; 
+        /*!
+            ensures
+                - returns a multiplier number.  The interpretation is that this object is
+                  requesting that the weight decay used to optimize its parameters be
+                  multiplied by get_weight_decay_multiplier().
+        !*/
+
+        void set_learning_rate_multiplier(
+            double val
+        );
+        /*!
+            requires
+                - val >= 0
+            ensures
+                - #get_learning_rate_multiplier() == val
+        !*/
+
+        void set_weight_decay_multiplier(
+            double val
+        ); 
+        /*!
+            requires
+                - val >= 0
+            ensures
+                - #get_weight_decay_multiplier() == val
+        !*/
+
+        double get_bias_learning_rate_multiplier(
+        ) const; 
+        /*!
+            ensures
+                - returns a multiplier number.  The interpretation is that this object is
+                  requesting that the learning rate used to optimize its bias parameters be
+                  multiplied by get_learning_rate_multiplier()*get_bias_learning_rate_multiplier().
+        !*/
+
+        double get_bias_weight_decay_multiplier(
+        ) const; 
+        /*!
+            ensures
+                - returns a multiplier number.  The interpretation is that this object is
+                  requesting that the weight decay used to optimize its bias parameters be
+                  multiplied by get_weight_decay_multiplier()*get_bias_weight_decay_multiplier().
+        !*/
+
+        void set_bias_learning_rate_multiplier(
+            double val
+        ); 
+        /*!
+            requires
+                - val >= 0
+            ensures
+                - #get_bias_learning_rate_multiplier() == val
+        !*/
+
+        void set_bias_weight_decay_multiplier(
+            double val
+        ); 
+        /*!
+            requires
+                - val >= 0
+            ensures
+                - #get_bias_weight_decay_multiplier() == val
+        !*/
+
+        template <typename SUBNET> void setup (const SUBNET& sub);
+        template <typename SUBNET> void forward(const SUBNET& sub, resizable_tensor& output);
+        template <typename SUBNET> void backward(const tensor& gradient_input, SUBNET& sub, tensor& params_grad);
+        point map_input_to_output(point p) const;
+        point map_output_to_input(point p) const;
+        const tensor& get_layer_params() const; 
+        tensor& get_layer_params(); 
+        /*!
+            These functions are implemented as described in the EXAMPLE_COMPUTATIONAL_LAYER_ interface.
+        !*/
+
+    };
+
+    template <
+        long num_filters,
+        long nr,
+        long nc,
+        int stride_y,
+        int stride_x,
+        typename SUBNET
+        >
+    using cont = add_layer<cont_<num_filters,nr,nc,stride_y,stride_x>, SUBNET>;
+
 // ----------------------------------------------------------------------------------------
 
     class dropout_

dlib/dnn/tensor_tools.h

@@ -879,12 +879,8 @@ namespace dlib { namespace tt
         void operator() (
             resizable_tensor& output,
             const tensor& data,
-            const tensor& filters,
-            int stride_y,
-            int stride_x,
-            int padding_y,
-            int padding_x
-        ) { impl(output,data,filters,stride_y,stride_x,padding_y,padding_x); }
+            const tensor& filters
+        ) { impl(output,data,filters); }
         /*!
             requires
                 - stride_y > 0
@@ -947,6 +943,16 @@ namespace dlib { namespace tt
                   this gradient to filters_gradient.
         !*/
 
+ 
+       void setup(
+            const tensor& data,
+            const tensor& filters,
+            int stride_y,
+            int stride_x,
+            int padding_y,
+            int padding_x
+        ) {impl.setup(data,filters,stride_y,stride_x,padding_y,padding_x); }
+       
     private:
 #ifdef DLIB_USE_CUDA
         cuda::tensor_conv impl;

dlib/test/dnn.cpp

@@ -805,8 +805,10 @@ namespace
                 padding_y = (filters.nr()-data.nr()+1)/2;
             if (!(filters.nc() <= data.nc() + 2*padding_x))
                 padding_x = (filters.nc()-data.nc()+1)/2;
-            conv1(output1, data, filters, stride_y,stride_x, padding_y, padding_x);
-            conv2(output2, data, filters, stride_y,stride_x, padding_y, padding_x);
+            conv1.setup(data,filters,stride_y,stride_x,padding_y,padding_x);
+            conv1(output1, data, filters);
+            conv2.setup(data,filters,stride_y,stride_x,padding_y,padding_x);
+            conv2(output2, data, filters);
             dlog << LINFO << "forward error: "<< max(abs(mat(output1)-mat(output2)));
             DLIB_TEST_MSG(max(abs(mat(output1)-mat(output2))) < 1e-3, max(abs(mat(output1)-mat(output2)))
                  <<"\n\t padding_y: "<< padding_y 
@@ -1473,6 +1475,24 @@ namespace
             auto res = test_layer(l);
             DLIB_TEST_MSG(res, res);
         }
+        {
+            print_spinner();
+            cont_<3,3,3,2,2> l;
+            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;