Merge pull request #42 from cvjena/releasing_v2.0.0

Pull Changes for v2.0.0

.travis.yml

@@ -19,7 +19,7 @@ install:
   - if [ "$TRAVIS_OS_NAME" == "linux" ]; then if [ "$TEST_SUITE" = "opencl" ]; then mkdir -p cl12/CL; fi; fi
   - if [ "$TRAVIS_OS_NAME" == "linux" ]; then if [ "$TEST_SUITE" = "opencl" ]; then cd cl12/CL; fi; fi
   - if [ "$TRAVIS_OS_NAME" == "linux" ]; then if [ "$TEST_SUITE" = "opencl" ]; then wget https://www.khronos.org/registry/cl/api/1.2/cl.h; fi; fi
-  - if [ "$TRAVIS_OS_NAME" == "linux" ]; then if [ "$TEST_SUITE" = "opencl" ]; then wget https://www.khronos.org/registry/cl/api/1.2/cl.hpp; fi; fi
+#  - if [ "$TRAVIS_OS_NAME" == "linux" ]; then if [ "$TEST_SUITE" = "opencl" ]; then wget https://www.khronos.org/registry/cl/api/1.2/cl.hpp; fi; fi
   - if [ "$TRAVIS_OS_NAME" == "linux" ]; then if [ "$TEST_SUITE" = "opencl" ]; then wget https://www.khronos.org/registry/cl/api/1.2/cl_d3d10.h; fi; fi
   - if [ "$TRAVIS_OS_NAME" == "linux" ]; then if [ "$TEST_SUITE" = "opencl" ]; then wget https://www.khronos.org/registry/cl/api/1.2/cl_d3d11.h; fi; fi
   - if [ "$TRAVIS_OS_NAME" == "linux" ]; then if [ "$TEST_SUITE" = "opencl" ]; then wget https://www.khronos.org/registry/cl/api/1.2/cl_dx9_media_sharing.h; fi; fi
@@ -49,7 +49,10 @@ before_script:
   - if [ "$TEST_SUITE" = "reference" ]; then cmake -DCMAKE_BUILD_TYPE=Release ..; fi
   - if [ "$TEST_SUITE" = "opencl" ]; then cmake -DCMAKE_BUILD_TYPE=Release -DCN24_BUILD_OPENCL:BOOL=ON -DCN24_BUILD_OPENCL_CLBLAS:BOOL=ON ..; fi
 
-script: make
+script:
+  - make
+# OpenCL on Mac OS X supports a maximum work group size of (1,1,1) on CPUs, so we only do a sanity check in that case
+  - if [ "$TEST_SUITE" == "opencl" ] && [ "$TRAVIS_OS_NAME" == "osx" ]; then ./testOpenCL; else ./runBenchmark --ci;fi
 
 os:
   - linux
@@ -62,3 +65,8 @@ compiler:
 env:
   - TEST_SUITE=opencl
   - TEST_SUITE=reference
+
+matrix:
+  exclude:
+    - os: linux
+      env: TEST_SUITE=opencl

CMakeLists.txt

@@ -1,7 +1,9 @@
 cmake_minimum_required(VERSION 2.8)
 project(CN24 C CXX)
 
-set(CMAKE_BUILD_TYPE "Debug" CACHE STRING "Build type")
+if(NOT CMAKE_BUILD_TYPE)
+  set(CMAKE_BUILD_TYPE "RelWithDebInfo" CACHE STRING "Build type: Select either Debug, RelWithDebInfo or Release" FORCE)
+endif()
 
 set(CMAKE_MODULE_PATH ${CMAKE_MODULE_PATH} "${CMAKE_SOURCE_DIR}/cmake/")
 
@@ -197,16 +199,17 @@ if(CN24_BUILD_ACCELERATE)
   message(STATUS "Using Accelerate include directory: ${ACCELERATE_INCLUDE_DIR}")
   include_directories(${ACCELERATE_INCLUDE_DIR})
   set(CN24_LIBS ${CN24_LIBS} ${ACCELERATE_BLAS})
+  set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -flax-vector-conversions")
   add_definitions("-DBUILD_BLAS")
   add_definitions("-DBLAS_ACCELERATE")
 endif()
 
 set(CN24_BUILD_OPENCL OFF CACHE BOOL "Build CN24 with OpenCL support")
 if(CN24_BUILD_OPENCL)
   find_library(OPENCL_ICDL NAMES OpenCL libOpenCL PATHS $ENV{CUDA_PATH}/lib/x64
-    $ENV{INTELOCLSDKROOT}/lib/x64)
+    $ENV{INTELOCLSDKROOT}/lib/x64 $ENV{AMDAPPSDKROOT}/lib/x86_64)
   find_path(OPENCL_INCLUDE_DIR CL/cl.h cl.h PATHS $ENV{CUDA_PATH}/include
-    $ENV{INTELOCLSDKROOT}/include)
+    $ENV{INTELOCLSDKROOT}/include $ENV{AMDAPPSDKROOT}/include)
   message(STATUS "Using OpenCL library: ${OPENCL_ICDL}")
   message(STATUS "Using OpenCL include directory: ${OPENCL_INCLUDE_DIR}")
   include_directories(${OPENCL_INCLUDE_DIR})
@@ -219,8 +222,8 @@ endif()
 set(CN24_BUILD_OPENCL_CLBLAS OFF CACHE BOOL "Build CN24 with OpenCL/clBLAS")
 if(CN24_BUILD_OPENCL_CLBLAS)
   if(CN24_BUILD_OPENCL)
-    find_library(CLBLAS_LIBRARY clBLAS libclBLAS)
-    find_path(CLBLAS_INCLUDE_DIR clBLAS.h)
+    find_library(CLBLAS_LIBRARY clBLAS libclBLAS PATHS $ENV{AMDAPPSDKROOT}/lib64/import)
+    find_path(CLBLAS_INCLUDE_DIR clBLAS.h PATHS $ENV{AMDAPPSDKROOT}/include)
     message(STATUS "Using OpenCL/clBLAS library: ${CLBLAS_LIBRARY}")
     message(STATUS "Using OpenCL/clBLAS include directory: ${CLBLAS_INCLUDE_DIR}")
     include_directories(${CLBLAS_INCLUDE_DIR})

include/cn24.h

@@ -17,14 +17,22 @@
 #include "cn24/util/Config.h"
 #include "cn24/util/Dataset.h"
 #include "cn24/util/Tensor.h"
+#include "cn24/util/CompressedTensor.h"
 #include "cn24/util/TensorViewer.h"
 #include "cn24/util/CombinedTensor.h"
+#include "cn24/util/TensorStream.h"
+#include "cn24/util/CompressedTensorStream.h"
+#include "cn24/util/FloatTensorStream.h"
 #include "cn24/util/PNGUtil.h"
 #include "cn24/util/JPGUtil.h"
 #include "cn24/util/Log.h"
 #include "cn24/util/KITTIData.h"
 #include "cn24/util/Init.h"
 #include "cn24/util/GradientTester.h"
+#include "cn24/util/StatAggregator.h"
+#include "cn24/util/StatSink.h"
+#include "cn24/util/ConsoleStatSink.h"
+#include "cn24/util/CSVStatSink.h"
 
 #include "cn24/math/TensorMath.h"
 
@@ -38,6 +46,7 @@
 #include "cn24/net/ConvolutionLayer.h"
 #include "cn24/net/MaxPoolingLayer.h"
 #include "cn24/net/AdvancedMaxPoolingLayer.h"
+#include "cn24/net/InputDownSamplingLayer.h"
 #include "cn24/net/LocalResponseNormalizationLayer.h"
 #include "cn24/net/UpscaleLayer.h"
 #include "cn24/net/LossFunctionLayer.h"
@@ -48,11 +57,13 @@
 #include "cn24/net/SpatialPriorLayer.h"
 #include "cn24/net/ConcatenationLayer.h"
 #include "cn24/net/GradientAccumulationLayer.h"
+#include "cn24/net/SumLayer.h"
 #include "cn24/net/Net.h"
 #include "cn24/net/Trainer.h"
 #include "cn24/net/NetGraph.h"
 #include "cn24/net/NetStatus.h"
 
 #include "cn24/factory/ConfigurableFactory.h"
+#include "cn24/factory/SkipLayerNetworkFactory.h"
 
 #endif

include/cn24/factory/ConfigurableFactory.h

@@ -24,8 +24,20 @@
 #include "../util/Log.h"
 
 namespace Conv {
+
+class Factory {
+public: 
+  virtual int AddLayers(Net& net, Connection data_layer_connection, const unsigned int output_classes, bool add_loss_layer = false, std::ostream& graph_output = std::cout) = 0;
+  virtual bool AddLayers(NetGraph& graph, NetGraphConnection data_layer_connection, const unsigned int output_classes, bool add_loss_layer = false) = 0;
+  virtual int patchsizex() = 0;
+  virtual int patchsizey() = 0;
+  virtual Layer* CreateLossLayer(const unsigned int output_classes, const datum loss_weight = 1.0) = 0;
+  virtual void InitOptimalSettings() = 0;
+  virtual TrainerSettings optimal_settings() const = 0;
+  virtual Method method() const = 0;
+};
 
-class ConfigurableFactory {
+class ConfigurableFactory : public Factory {
 public:
   /**
 	* @brief Builds a ConfigurableFactory using an input stream and a random seed

include/cn24/factory/SkipLayerNetworkFactory.h

@@ -0,0 +1,28 @@
+#ifndef CONV_SKIPLAYERNETWORKFACTORY_H
+#define CONV_SKIPLAYERNETWORKFACTORY_H
+
+#include <iostream>
+
+#include "../net/Net.h"
+#include "../net/NetGraph.h"
+#include "../net/Trainer.h"
+#include "../util/Dataset.h"
+#include "../util/Log.h"
+#include "ConfigurableFactory.h"
+
+namespace Conv {
+
+class SkipLayerNetworkFactory : public Factory {
+  int AddLayers(Net& net, Connection data_layer_connection, const unsigned int output_classes, bool add_loss_layer = false, std::ostream& graph_output = std::cout);
+  bool AddLayers(NetGraph& graph, NetGraphConnection data_layer_connection, const unsigned int output_classes, bool add_loss_layer = false);
+  int patchsizex();
+  int patchsizey();
+  Layer* CreateLossLayer(const unsigned int output_classes, const datum loss_weight = 1.0);
+  void InitOptimalSettings();
+  TrainerSettings optimal_settings() const;
+  Method method() const;
+};
+
+}
+
+#endif

include/cn24/math/TensorMath.h

@@ -100,6 +100,25 @@ class TensorMath {
   static void SMS2(
     const Tensor& source,
     Tensor& target);
+
+  static void DOWN(
+    const Tensor& source,
+    Tensor& target,
+    const int region_width,
+    const int region_height,
+    const datum target_factor);
+
+  static void UP(
+    const Tensor& source,
+    Tensor& target,
+    const int region_width,
+    const int region_height,
+    const datum target_factor);
+
+  static void ADD(
+    const Tensor& source_a,
+    const Tensor& source_b,
+    Tensor& target);
 };
 
 }

include/cn24/net/BinaryStatLayer.h

@@ -19,6 +19,7 @@
 
 #include "Layer.h"
 #include "StatLayer.h"	
+#include "../util/StatAggregator.h"
 
 namespace Conv {
 
@@ -33,6 +34,8 @@ class BinaryStatLayer: public Layer, public StatLayer {
 	*/
   BinaryStatLayer(unsigned int thresholds = 24, const datum min_t = -0.458333,
                   const datum max_t = 0.5);
+
+  void UpdateAll();
 
   /**
 	* @brief Prints the current statistics
@@ -80,6 +83,13 @@ class BinaryStatLayer: public Layer, public StatLayer {
   datum* false_negatives_ = nullptr;
 
   bool disabled_ = false;
+
+  StatDescriptor* stat_fpr_ = nullptr;
+  StatDescriptor* stat_fnr_ = nullptr;
+  StatDescriptor* stat_pre_ = nullptr;
+  StatDescriptor* stat_rec_ = nullptr;
+  StatDescriptor* stat_acc_ = nullptr;
+  StatDescriptor* stat_f1_ = nullptr;
 };
 
 }

include/cn24/net/ConfusionMatrixLayer.h

@@ -19,7 +19,9 @@
 #include <vector>
 
 #include "Layer.h"
-#include "StatLayer.h"	
+#include "StatLayer.h"
+
+#include "../util/StatAggregator.h"
 
 namespace Conv {
 
@@ -34,10 +36,11 @@ class ConfusionMatrixLayer: public Layer, public StatLayer {
   explicit ConfusionMatrixLayer(std::vector<std::string> names,
                                 const unsigned int classes);
 
+  void UpdateAll();
   /**
 	* @brief Prints the current statistics
 	*
-	* @param prefix This is printed before every line ouf output
+	* @param prefix This is printed before every line of output
 	* @param training Whether the net is currently training. Affects output color
 	*/
   void Print (std::string prefix, bool training);
@@ -80,6 +83,11 @@ class ConfusionMatrixLayer: public Layer, public StatLayer {
   long double total_ = 0;
   long double right_ = 0;
   long double* per_class_ = nullptr;
+
+  StatDescriptor* stat_orr_ = nullptr;
+  StatDescriptor* stat_arr_ = nullptr;
+  StatDescriptor* stat_iou_ = nullptr;
+
 };
 
 }

include/cn24/net/InputDownSamplingLayer.h

@@ -0,0 +1,70 @@
+/*
+ * This file is part of the CN24 semantic segmentation software,
+ * copyright (C) 2015 Clemens-Alexander Brust (ikosa dot de at gmail dot com).
+ *
+ * For licensing information, see the LICENSE file included with this project.
+ */  
+/**
+ * @file InputDownSamplingLayer.h
+ * @class InputDownSamplingLayer
+ * @brief Layer that scales input down
+ * 
+ * @author Clemens-Alexander Brust (ikosa dot de at gmail dot com)
+ */
+
+#ifndef CONV_INPUTDOWNSAMPLINGLAYER_H
+#define CONV_INPUTDOWNSAMPLINGLAYER_H
+
+#include <string>
+#include <sstream>
+
+#include "SimpleLayer.h"
+
+
+namespace Conv {
+
+class InputDownSamplingLayer : public SimpleLayer {
+public:
+  /**
+   * @brief Constructs a max-pooling Layer.
+   * 
+   * @param region_width Width of the pooling regions
+   * @param region_height Height of the pooling regions
+   */
+  InputDownSamplingLayer(const unsigned int region_width,
+                  const unsigned int region_height);
+
+  // Implementations for SimpleLayer
+  bool CreateOutputs (const std::vector< CombinedTensor* >& inputs, std::vector< CombinedTensor* >& outputs);
+  bool Connect (const CombinedTensor* input, CombinedTensor* output);
+  void FeedForward();
+  void BackPropagate();
+
+  inline unsigned int Gain() {
+    return gain / (region_width_ * region_height_);
+  }
+
+	inline std::string GetLayerDescription() {
+		std::ostringstream ss;
+		ss << "Input Down-Sampling Layer (" << region_width_ << "x" << region_height_ << ")";
+		return ss.str();
+	}
+
+  bool IsOpenCLAware();
+private:
+  // Settings
+  unsigned int region_width_ = 0;
+  unsigned int region_height_ = 0;
+
+  // Feature map dimensions
+  unsigned int input_width_ = 0;
+  unsigned int input_height_ = 0;
+  unsigned int output_width_ = 0;
+  unsigned int output_height_ = 0;
+
+  unsigned int maps_ = 0;
+};
+
+}
+
+#endif

include/cn24/net/NetGraph.h

@@ -20,6 +20,8 @@
 #include "NetStatus.h"
 #include "../util/TensorViewer.h"
 
+#include "StatLayer.h"
+
 #include <vector>
 
 namespace Conv {
@@ -106,6 +108,12 @@ class NetGraph : public NetStatus {
 	// Output
 	void PrintGraph(std::ostream& graph_output);
   void SetLayerViewEnabled(bool enabled) { layerview_enabled_ = enabled; }
+  void SetStatLayersEnabled(bool enabled) {
+    for (unsigned int n = 0; n < GetStatNodes().size(); n++) {
+      StatLayer* stat_layer = dynamic_cast<StatLayer*>(GetStatNodes()[n]->layer);
+      stat_layer->SetDisabled(!enabled);
+    } 
+  }
 	datum AggregateLoss();
 
 	// Status

include/cn24/net/NetStatus.h

@@ -15,6 +15,9 @@
 #ifndef CONV_NETSTATUS_H
 #define CONV_NETSTATUS_H
 
+#include "../util/Init.h"
+#include "../util/StatAggregator.h"
+
 namespace Conv {
 
 class NetStatus{
@@ -29,7 +32,10 @@ class NetStatus{
    * 
    * @param is_testing The new testing status
    */
-  inline void SetIsTesting(bool is_testing) { is_testing_ = is_testing; }
+  inline void SetIsTesting(bool is_testing) { 
+    is_testing_ = is_testing;
+    System::stat_aggregator->hardcoded_stats_.is_training = !is_testing;
+  }
 private:
 	bool is_testing_ = false;
 };

include/cn24/net/StatLayer.h

@@ -23,6 +23,7 @@ namespace Conv {
 
 class StatLayer {
 public:
+  virtual void UpdateAll() = 0;
   virtual void Print(std::string prefix, bool training) = 0;
   virtual void Reset() = 0;
 	virtual void SetDisabled(bool disabled) = 0;