Changed tensor layout from NHWC to NCHW since this layout is much faster for

cuDNN.

dlib/dnn/core.h

@@ -1362,7 +1362,7 @@ namespace dlib
 
     namespace timpl
     {
-        void fill_with_gassuan_random_numbers (
+        inline void fill_with_gassuan_random_numbers (
             tensor& t,
             dlib::rand& rnd,
             double sigma = 1
@@ -1383,12 +1383,12 @@ namespace dlib
                 // Output and gradient_input have to have the same dimensions in each
                 // layer.
                 const long num_samples = rnd.get_random_32bit_number()%4+3;
+                const long k  = rnd.get_random_32bit_number()%4+2;
                 const long nr = rnd.get_random_32bit_number()%4+2;
                 const long nc = rnd.get_random_32bit_number()%4+2;
-                const long k  = rnd.get_random_32bit_number()%4+2;
 
-                output.set_size(num_samples, nr, nc, k);
-                gradient_input.set_size(num_samples, nr, nc, k);
+                output.set_size(num_samples, k, nr, nc);
+                gradient_input.set_size(num_samples, k, nr, nc);
 
                 // Use a non-zero initial gradient to make sure the layers add to it
                 // rather than assign and blow away the initial value.
@@ -1447,7 +1447,8 @@ namespace dlib
         };
 
 
-        void print_tensor(
+        // TODO, remove?
+        inline void print_tensor(
             const tensor& a
         )
         {

dlib/dnn/cudnn_dlibapi.cpp

@@ -85,9 +85,9 @@ namespace dlib
         void tensor_descriptor::
         set_size(
             int n, 
+            int k,
             int nr, 
-            int nc, 
-            int k
+            int nc 
         )
         {
             if (n == 0 || nr == 0 || nc == 0 || k == 0)
@@ -105,7 +105,7 @@ namespace dlib
                 handle = h;
 
                 check(cudnnSetTensor4dDescriptor((cudnnTensorDescriptor_t)handle,
-                        CUDNN_TENSOR_NHWC,
+                        CUDNN_TENSOR_NCHW,
                         CUDNN_DATA_FLOAT,
                         n,
                         k,
@@ -117,9 +117,9 @@ namespace dlib
         void tensor_descriptor::
         get_size (
             int& n, 
-            int& nr, 
-            int& nc, 
-            int& k
+            int& k,
+            int& nr,
+            int& nc
         ) const
         {
             if (handle)
@@ -140,9 +140,9 @@ namespace dlib
             else
             {
                 n = 0;
+                k = 0;
                 nr = 0;
                 nc = 0;
-                k = 0;
             }
         }
 
@@ -254,7 +254,7 @@ namespace dlib
                         &out_nc));
 
                 tensor_descriptor dest_desc;
-                dest_desc.set_size(out_num_samples,out_nr,out_nc,out_k);
+                dest_desc.set_size(out_num_samples,out_k,out_nr,out_nc);
 
                 cudnnConvolutionFwdAlgo_t forward_best_algo;
                 check(cudnnGetConvolutionForwardAlgorithm(
@@ -299,7 +299,7 @@ namespace dlib
             const tensor& filters
         )
         {
-            output.set_size(out_num_samples, out_nr, out_nc, out_k);
+            output.set_size(out_num_samples, out_k, out_nr, out_nc);
 
 
             // TODO, remove

dlib/dnn/cudnn_dlibapi.h

@@ -37,9 +37,9 @@ namespace dlib
 
             void set_size(
                 int n, 
+                int k,
                 int nr, 
-                int nc, 
-                int k
+                int nc 
             );
             /*!
                 ensures
@@ -48,9 +48,9 @@ namespace dlib
 
             void get_size (
                 int& n, 
-                int& nr, 
-                int& nc, 
-                int& k
+                int& k,
+                int& nr,
+                int& nc 
             ) const;
 
             const void* get_handle (
@@ -209,9 +209,9 @@ namespace dlib
 
             // dimensions of the output tensor from operator()
             int out_num_samples;
+            int out_k;
             int out_nr;
             int out_nc;
-            int out_k;
 
             int forward_algo;
             size_t forward_workspace_size_in_bytes;

dlib/dnn/input.h

@@ -56,8 +56,9 @@ namespace dlib
 
             
             // initialize data to the right size to contain the stuff in the iterator range.
-            data.set_size(std::distance(ibegin,iend), nr, nc, pixel_traits<T>::num);
+            data.set_size(std::distance(ibegin,iend), pixel_traits<T>::num, nr, nc);
 
+            const size_t offset = nr*nc;
             auto ptr = data.host();
             for (auto i = ibegin; i != iend; ++i)
             {
@@ -66,10 +67,15 @@ namespace dlib
                     for (long c = 0; c < nc; ++c)
                     {
                         auto temp = pixel_to_vector<float>((*i)(r,c));
+                        auto p = ptr++;
                         for (long j = 0; j < temp.size(); ++j)
-                            *ptr++ = temp(j);
+                        {
+                            *p = temp(j);
+                            p += offset;
+                        }
                     }
                 }
+                ptr += offset*(data.k()-1);
             }
 
         }
@@ -123,8 +129,9 @@ namespace dlib
 
             
             // initialize data to the right size to contain the stuff in the iterator range.
-            data.set_size(std::distance(ibegin,iend), nr, nc, pixel_traits<T>::num);
+            data.set_size(std::distance(ibegin,iend), pixel_traits<T>::num, nr, nc);
 
+            const size_t offset = nr*nc;
             auto ptr = data.host();
             for (auto i = ibegin; i != iend; ++i)
             {
@@ -133,10 +140,15 @@ namespace dlib
                     for (long c = 0; c < nc; ++c)
                     {
                         auto temp = pixel_to_vector<float>((*i)[r][c]);
+                        auto p = ptr++;
                         for (long j = 0; j < temp.size(); ++j)
-                            *ptr++ = temp(j);
+                        {
+                            *p = temp(j);
+                            p += offset;
+                        }
                     }
                 }
+                ptr += offset*(data.k()-1);
             }
 
         }

dlib/dnn/layers.h

@@ -83,7 +83,7 @@ namespace dlib
         template <typename SUBNET>
         void forward(const SUBNET& sub, resizable_tensor& output)
         {
-            output.set_size(sub.get_output().num_samples(), 1,1,num_outputs);
+            output.set_size(sub.get_output().num_samples(), num_outputs);
 
             output = mat(sub.get_output())*mat(params);
         } 

dlib/dnn/tensor.h

@@ -22,30 +22,22 @@ namespace dlib
 
         tensor (
         ) : 
-            m_n(0), m_nr(0), m_nc(0), m_k(0)
+            m_n(0), m_k(0), m_nr(0), m_nc(0)
         {
         }
 
         inline virtual ~tensor() = 0;
 
         long num_samples() const { return m_n; }
+        long k() const { return m_k; }
         long nr() const { return m_nr; }
         long nc() const { return m_nc; }
-        long k() const { return m_k; }
         size_t size() const { return data.size(); }
 
         void async_copy_to_device() 
         {
             data.async_copy_to_device();
         }
-        /*!
-            ensures
-                - begin asynchronously copying this tensor to the GPU.
-
-                NOTE that the "get device pointer" routine in this class
-                will have to do some kind of synchronization that ensures
-                the copy is finished.
-        !*/
 
         const float* host() const { return data.host(); }
         float*       host()       { return data.host(); }
@@ -135,13 +127,13 @@ namespace dlib
         tensor& operator= (const tensor& item) 
         {
             m_n  = item.m_n;
+            m_k  = item.m_k;
             m_nr = item.m_nr;
             m_nc = item.m_nc;
-            m_k  = item.m_k;
             data.set_size(item.data.size());
             std::memcpy(data.host(), item.data.host(), data.size()*sizeof(float));
 #ifdef DLIB_USE_CUDA
-            cudnn_descriptor.set_size(m_n,m_nr,m_nc,m_k);
+            cudnn_descriptor.set_size(m_n,m_k,m_nr,m_nc);
 #endif
             return *this;
         }
@@ -159,9 +151,9 @@ namespace dlib
         void swap(tensor& item)
         {
             std::swap(m_n,    item.m_n);
+            std::swap(m_k,    item.m_k);
             std::swap(m_nr,   item.m_nr);
             std::swap(m_nc,   item.m_nc);
-            std::swap(m_k,    item.m_k);
             std::swap(data, item.data);
 #ifdef DLIB_USE_CUDA
             std::swap(cudnn_descriptor, item.cudnn_descriptor);
@@ -170,9 +162,9 @@ namespace dlib
 
 
         long m_n;
+        long m_k;
         long m_nr;
         long m_nc;
-        long m_k;
         gpu_data data;
 #ifdef DLIB_USE_CUDA
         cuda::tensor_descriptor cudnn_descriptor;
@@ -227,9 +219,9 @@ namespace dlib
     )
     {
         return a.num_samples() == b.num_samples() &&
+               a.k()  == b.k() &&
                a.nr() == b.nr() &&
-               a.nc() == b.nc() &&
-               a.k()  == b.k();
+               a.nc() == b.nc();
     }
 
 // ----------------------------------------------------------------------------------------
@@ -242,10 +234,10 @@ namespace dlib
         {}
 
         explicit resizable_tensor(
-            long n_, long nr_ = 1, long nc_ = 1, long k_ = 1
+            long n_, long k_ = 1, long nr_ = 1, long nc_ = 1
         ) 
         {
-            set_size(n_,nr_,nc_,k_);
+            set_size(n_,k_,nr_,nc_);
         }
 
         resizable_tensor(const resizable_tensor&) = default;
@@ -265,7 +257,7 @@ namespace dlib
                 - resizes *this so that: have_same_dimensions(#*this, item)==true
         !*/
         {
-            set_size(item.num_samples(), item.nr(), item.nc(), item.k());
+            set_size(item.num_samples(), item.k(), item.nr(), item.nc());
         }
 
         resizable_tensor& operator= (float val)
@@ -323,16 +315,16 @@ namespace dlib
         }
 
         void set_size(
-            long n_, long nr_ = 1, long nc_ = 1, long k_ = 1
+            long n_, long k_ = 1, long nr_ = 1, long nc_ = 1
         )
         {
             m_n = n_;
+            m_k = k_;
             m_nr = nr_;
             m_nc = nc_;
-            m_k = k_;
-            data.set_size(m_n*m_nr*m_nc*m_k);
+            data.set_size(m_n*m_k*m_nr*m_nc);
 #ifdef DLIB_USE_CUDA
-            cudnn_descriptor.set_size(m_n,m_nr,m_nc,m_k);
+            cudnn_descriptor.set_size(m_n,m_k,m_nr,m_nc);
 #endif
         }
     };
@@ -342,9 +334,9 @@ namespace dlib
         int version = 1;
         serialize(version, out);
         serialize(item.num_samples(), out);
+        serialize(item.k(), out);
         serialize(item.nr(), out);
         serialize(item.nc(), out);
-        serialize(item.k(), out);
         auto data = item.host();
         for (size_t i = 0; i < item.size(); ++i)
             serialize(data[i], out);
@@ -357,12 +349,12 @@ namespace dlib
         if (version != 1)
             throw serialization_error("Unexpected version found while deserializing dlib::resizable_tensor.");
 
-        long num_samples=0, nr=0, nc=0, k=0;
+        long num_samples=0, k=0, nr=0, nc=0;
         deserialize(num_samples, in);
+        deserialize(k, in);
         deserialize(nr, in);
         deserialize(nc, in);
-        deserialize(k, in);
-        item.set_size(num_samples, nr, nc, k);
+        item.set_size(num_samples, k, nr, nc);
         auto data = item.host();
         for (size_t i = 0; i < item.size(); ++i)
             deserialize(data[i], in);