Add input_tensor input type (#2951)

dlib/cuda/tensor.h

@@ -680,6 +680,14 @@ namespace dlib
 
 // ----------------------------------------------------------------------------------------
 
+    inline void memcpy (
+        alias_tensor_instance&& dest, 
+        const tensor& src
+    ) 
+    {
+        memcpy(static_cast<tensor&>(dest), src);
+    }
+
 }
 
 #endif // DLIB_DNn_TENSOR_H_

dlib/cuda/tensor_abstract.h

@@ -607,6 +607,14 @@ namespace dlib
         ); 
     };
 
+    inline void memcpy (
+        alias_tensor_instance&& dest, 
+        const tensor& src
+    ) { memcpy(static_cast<tensor&>(dest), src); }
+    /*!
+        A convenient overload for copying from src to dest when you have a temporary alias tensor.
+    !*/
+
     class alias_tensor_const_instance 
     {
         /*!

dlib/dnn/input.h

@@ -1082,6 +1082,93 @@ namespace dlib
         float avg_blue;
     };
 
+// ----------------------------------------------------------------------------------------
+
+    class input_tensor
+    {
+    public:
+        typedef tensor input_type;
+
+        input_tensor() {}
+        input_tensor(const input_tensor&) {}
+
+        template<typename forward_iterator>
+        void to_tensor(
+            forward_iterator ibegin,
+            forward_iterator iend,
+            resizable_tensor& data
+        ) const
+        {
+            DLIB_CASSERT(std::distance(ibegin, iend) > 0);
+            const auto k = ibegin->k();
+            const auto nr = ibegin->nr();
+            const auto nc = ibegin->nc();
+            // make sure all the input tensors have the same dimensions
+            for (auto i = ibegin; i != iend; ++i)
+            {
+                DLIB_CASSERT(i->k() == k && i->nr() == nr && i->nc() == nc,
+                    "\t input_tensor::to_tensor()"
+                    << "\n\t All tensor objects given to to_tensor() must have the same dimensions."
+                    << "\n\t k: " << k
+                    << "\n\t nr: " << nr
+                    << "\n\t nc: " << nc
+                    << "\n\t i->k(): " << i->k()
+                    << "\n\t i->nr(): " << i->nr()
+                    << "\n\t i->nc(): " << i->nc()
+                );
+            }
+
+            const auto num_samples = count_samples(ibegin, iend);
+            // initialize data to the right size to contain the stuff in the iterator range.
+            data.set_size(num_samples, k, nr, nc);
+
+            const size_t stride = k * nr * nc;
+            size_t offset = 0;
+            for (auto i = ibegin; i != iend; ++i)
+            {
+                alias_tensor slice(i->num_samples(), k, nr, nc);
+                memcpy(slice(data, offset), *i);
+                offset += slice.num_samples() * stride;
+            }
+        }
+
+        friend void serialize(const input_tensor&, std::ostream& out)
+        {
+            serialize("input_tensor", out);
+        }
+
+        friend void deserialize(input_tensor&, std::istream& in)
+        {
+            std::string version;
+            deserialize(version, in);
+            if (version != "input_tensor")
+                throw serialization_error("Unexpected version found while deserializing dlib::input_tensor.");
+        }
+
+        friend std::ostream& operator<<(std::ostream& out, const input_tensor&)
+        {
+            out << "input_tensor";
+            return out;
+        }
+
+        friend void to_xml(const input_tensor&, std::ostream& out)
+        {
+            out << "<input_tensor/>\n";
+        }
+
+    private:
+
+        template<typename forward_iterator>
+        long long count_samples(
+            forward_iterator ibegin,
+            forward_iterator iend
+        ) const
+        {
+            return std::accumulate(ibegin, iend, 0,
+                [](long long a, const auto& b) { return a + b.num_samples(); });
+        }
+    };
+
 // ----------------------------------------------------------------------------------------
 
 }

dlib/dnn/input_abstract.h

@@ -719,6 +719,57 @@ namespace dlib
 
 // ----------------------------------------------------------------------------------------
 
+    class input_tensor
+    {
+        /*!
+            WHAT THIS OBJECT REPRESENTS
+                This input layer works with dlib::tensor objects. It is very similar to
+                the dlib::input layer except that it allows for concatenating data that
+                already resides in GPU memory.
+        !*/
+
+    public:
+        typedef tensor input_type;
+
+        input_tensor(
+        );
+        /*!
+            ensures
+                - input_tensor objects are default constructable
+        !*/
+
+        input_tensor(
+            const input_tensor& item
+        );
+        /*!
+            ensures
+                - input_tensor objects are copy constructable
+        !*/
+
+        template <typename forward_iterator>
+        void to_tensor(
+            forward_iterator ibegin,
+            forward_iterator iend,
+            resizable_tensor& data
+        ) const;
+        /*!
+            requires
+                - [ibegin, iend) is an iterator range over input_type objects.
+                - std::distance(ibegin,iend) > 0
+                - The input range should contain tensor objects that all have the same
+                  dimensions.
+            ensures
+                - Copies the iterator range into #data.  In particular, if the input tensors
+                  have R rows, C columns, and K channels then we will have:
+                    - #data.num_samples() == count_samples(ibegin,iend)
+                    - #data.nr() == R
+                    - #data.nc() == C
+                    - #data.k() == K
+                  This results in a tensor concatenation along the sample dimension.
+        !*/
+    };
+
+// ----------------------------------------------------------------------------------------
 }
 
 #endif // DLIB_DNn_INPUT_ABSTRACT_H_

dlib/test/dnn.cpp

@@ -4276,6 +4276,38 @@ namespace
 #endif
     }
 
+    void test_input_tensor()
+    {
+        using namespace dlib::tt;
+        print_spinner();
+        tt::tensor_rand rnd;
+        std::vector<resizable_tensor> tensors(3);
+
+        for (auto& t : tensors) {
+            t.set_size(1, 3, 224, 224);
+            rnd.fill_gaussian(t);
+        }
+
+        resizable_tensor out;
+        input_tensor input_layer;
+
+        input_layer.to_tensor(tensors.begin(), tensors.end(), out);
+
+        DLIB_TEST(out.num_samples() == 3);
+        DLIB_TEST(out.k() == 3);
+        DLIB_TEST(out.nr() == 224);
+        DLIB_TEST(out.nc() == 224);
+        size_t stride = out.k() * out.nr() * out.nc();
+        size_t offset = 0;
+        int error = 0;
+
+        for (auto& t : tensors) {
+            error = memcmp(out.host() + offset, t.host(), sizeof(float) * t.size());
+            DLIB_TEST(error == 0);
+            offset += stride;
+        }
+    }
+
 // ----------------------------------------------------------------------------------------
 
     class dnn_tester : public tester
@@ -4386,6 +4418,7 @@ namespace
             test_input_ouput_mappers();
             test_fuse_layers();
             test_reorg();
+            test_input_tensor();
         }
 
         void perform_test()