From 18695b7b4b1b73730ed001b690cebd5ff7c96437 Mon Sep 17 00:00:00 2001
From: Davis King <davis@dlib.net>
Date: Wed, 23 Dec 2015 08:23:46 -0500
Subject: [PATCH] Made the default input layer automatically normalize unsigned
 char pixel values to the range [0,1].

---
 dlib/dnn/input.h          | 13 +++++++++++--
 dlib/dnn/input_abstract.h |  4 ++++
 2 files changed, 15 insertions(+), 2 deletions(-)

diff --git a/dlib/dnn/input.h b/dlib/dnn/input.h
index a4bfa479c..7603aa0af 100644
--- a/dlib/dnn/input.h
+++ b/dlib/dnn/input.h
@@ -58,6 +58,8 @@ namespace dlib
             // initialize data to the right size to contain the stuff in the iterator range.
             data.set_size(std::distance(ibegin,iend), pixel_traits<T>::num, nr, nc);
 
+            typedef typename pixel_traits<T>::basic_pixel_type bptype;
+
             const size_t offset = nr*nc;
             auto ptr = data.host();
             for (auto i = ibegin; i != iend; ++i)
@@ -70,7 +72,10 @@ namespace dlib
                         auto p = ptr++;
                         for (long j = 0; j < temp.size(); ++j)
                         {
-                            *p = temp(j);
+                            if (is_same_type<bptype,unsigned char>::value)
+                                *p = temp(j)/256.0;
+                            else
+                                *p = temp(j);
                             p += offset;
                         }
                     }
@@ -130,6 +135,7 @@ namespace dlib
             
             // initialize data to the right size to contain the stuff in the iterator range.
             data.set_size(std::distance(ibegin,iend), pixel_traits<T>::num, nr, nc);
+            typedef typename pixel_traits<T>::basic_pixel_type bptype;
 
             const size_t offset = nr*nc;
             auto ptr = data.host();
@@ -143,7 +149,10 @@ namespace dlib
                         auto p = ptr++;
                         for (long j = 0; j < temp.size(); ++j)
                         {
-                            *p = temp(j);
+                            if (is_same_type<bptype,unsigned char>::value)
+                                *p = temp(j)/256.0;
+                            else
+                                *p = temp(j);
                             p += offset;
                         }
                     }
diff --git a/dlib/dnn/input_abstract.h b/dlib/dnn/input_abstract.h
index ceab0de9f..68e8b1f67 100644
--- a/dlib/dnn/input_abstract.h
+++ b/dlib/dnn/input_abstract.h
@@ -135,6 +135,10 @@ namespace dlib
                   For example, a matrix<float,3,3> would turn into a tensor with 3 rows, 3
                   columns, and k()==1.  Or a matrix<rgb_pixel,4,5> would turn into a tensor
                   with 4 rows, 5 columns, and k()==3 (since rgb_pixels have 3 channels).
+                - If the input data contains pixels of type unsigned char, rgb_pixel, or
+                  other pixel types with a basic_pixel_type of unsigned char then each
+                  value written to the output tensor is first divided by 256.0 so that the
+                  resulting outputs are all in the range [0,1].
         !*/
     };