Add bias correction of adam optimizer

add bias correction according to the epoch.

**Self evaluation:**
1. Build test:	 [X]Passed [ ]Failed [ ]Skipped
2. Run test:	 [X]Passed [ ]Failed [ ]Skipped

Signed-off-by: jijoong.moon <jijoong.moon@samsung.com>

diff --git a/DeepQ/jni/CMakeLists.txt b/DeepQ/jni/CMakeLists.txt
index fbb1454..19e1970 100644 (file)
--- a/DeepQ/jni/CMakeLists.txt
+++ b/DeepQ/jni/CMakeLists.txt
@@ -4,7 +4,7 @@ project(DeepQ)
 set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -Wall -Werror -g")
 set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wall -Werror -g -std=c++11")
 
-# ADD_DEFINITIONS(-DUSING_CUSTOM_ENV)
+ADD_DEFINITIONS(-DUSING_CUSTOM_ENV)
 
 find_package(PkgConfig REQUIRED)
 

diff --git a/DeepQ/jni/main.cpp b/DeepQ/jni/main.cpp
index da66652..88db8d4 100644 (file)
--- a/DeepQ/jni/main.cpp
+++ b/DeepQ/jni/main.cpp
@@ -147,7 +147,6 @@ int main(int argc, char **argv) {
 
   mainNet.init(input_size, HIDDEN_LAYER_SIZE, output_size, MINI_BATCH,
                LEARNING_RATE, "tanh", true);
-
   mainNet.setOptimizer("adam", LEARNING_RATE, 0.9, 0.999, 1e-8);
 
   targetNet.init(input_size, HIDDEN_LAYER_SIZE, output_size, MINI_BATCH,

diff --git a/NeuralNet/layers.cpp b/NeuralNet/layers.cpp
index 27b2451..cf88f78 100644 (file)
--- a/NeuralNet/layers.cpp
+++ b/NeuralNet/layers.cpp
@@ -126,6 +126,8 @@ Matrix FullyConnectedLayer::backwarding(Matrix derivative, int iteration) {
     M = M.multiply(opt.beta1).add(dJdW.average().multiply(1 - opt.beta1));
     V = V.multiply(opt.beta2).add(
         (dJdW.average().multiply(dJdW.average())).multiply(1 - opt.beta2));
+    M.divide(1 - pow(opt.beta1, iteration + 1));
+    V.divide(1 - pow(opt.beta2, iteration + 1));
     Weight = Weight.subtract((M.divide(V.applyFunction(sqrt).add(opt.epsilon)))
                                  .multiply(opt.learning_rate));
     break;
@@ -231,6 +233,8 @@ Matrix OutputLayer::backwarding(Matrix label, int iteration) {
     M = M.multiply(opt.beta1).add(dJdW.average().multiply(1 - opt.beta1));
     V = V.multiply(opt.beta2).add(
         (dJdW.average().multiply(dJdW.average())).multiply(1 - opt.beta2));
+    M.divide(1 - pow(opt.beta1, iteration + 1));
+    V.divide(1 - pow(opt.beta2, iteration + 1));
     Weight = Weight.subtract((M.divide(V.applyFunction(sqrt).add(opt.epsilon)))
                                  .multiply(opt.learning_rate));
     break;

diff --git a/Training/jni/demo.cpp b/Training/jni/demo.cpp
deleted file mode 100644 (file)
index 168fc0e..0000000
--- a/Training/jni/demo.cpp
+++ /dev/null
@@ -1,215 +0,0 @@
-#include <stdio.h>
-#include "tensorflow/contrib/lite/interpreter.h"
-#include "tensorflow/contrib/lite/kernels/register.h"
-#include "tensorflow/contrib/lite/string_util.h"
-#include "tensorflow/contrib/lite/string.h"
-#include "tensorflow/contrib/lite/model.h"
-#include "tensorflow/contrib/lite/tools/gen_op_registration.h"
-#include "bitmap_helpers.h"
-#include <iostream>
-#include <fstream>
-#include "math.h"
-
-#define TOTAL_DATA_SIZE 5
-#define TOTAL_LABEL_SIZE 3
-#define TOTAL_TEST_SIZE 8
-
-int KNN(float out[3][5][128], float *test){
-  int ret=0;
-  float dist[15];
-
-  int count =0;
-  float sum=0.0;
-  float max=100000.0;
-  int id = 0;
-  for(int i=0;i<TOTAL_LABEL_SIZE;i++){
-    for(int j=0;j<TOTAL_DATA_SIZE;j++){
-      float d;
-      for(int k=0;k<128;k++){
-       sum += (out[i][j][k] - test[k])*(out[i][j][k] - test[k]);
-      }
-      d=sqrt(sum);
-      dist[count++]=d;
-      if(d < max){
-       max=d;
-       id=i;
-      }
-      printf("id %d, dist %f\n", id, d);
-      sum=0.0;
-    }
-  }
-  ret = id;
-
-  return ret;
-}
-
-
-int main(){
-  int tensor_size;
-  int node_size;
-  int input_size;
-  int output_size;
-  int *output_idx_list;
-  int *input_idx_list;
-  int inputDim[4];
-  int outputDim[4];
-  int input_idx_list_len=0;
-  int output_idx_list_len = 0;
-  float out[TOTAL_LABEL_SIZE][TOTAL_DATA_SIZE][128];
-
-  char *total_label[TOTAL_LABEL_SIZE]={"happy","sad","soso"};
-  char *data_path="/sdcard/Transfer-Learning/";
-  
-  std::unique_ptr<tflite::FlatBufferModel> model= tflite::FlatBufferModel::BuildFromFile("/sdcard/Transfer-Learning/ssd_mobilenet_v2_coco_feature.tflite");
-
-  if(!model){
-    printf("Failed to mmap mdoel\n");
-    exit(0);
-  }
-
-  tflite::ops::builtin::BuiltinOpResolver resolver;
-  std::unique_ptr<tflite::Interpreter> interpreter;
-  tflite::InterpreterBuilder(*model.get(), resolver)(&interpreter);
-
-  tensor_size = interpreter->tensors_size();
-  node_size = interpreter->nodes_size();
-  input_size = interpreter->inputs().size();
-  output_size = interpreter->outputs().size();
-
-  input_idx_list = new int[input_size];
-  output_idx_list = new int[output_size];
-
-  int t_size = interpreter->tensors_size();
-  for(int i=0;i<t_size;i++){
-    for(int j=0;j<input_size;j++){
-      if(strcmp(interpreter->tensor(i)->name, interpreter->GetInputName(j)) == 0)
-       input_idx_list[input_idx_list_len++]=i;
-    }
-    for(int j=0;j<output_size;j++){
-      if(strcmp(interpreter->tensor(i)->name, interpreter->GetOutputName(j))==0)
-       output_idx_list[output_idx_list_len++] = i;
-    }
-  }
-
-  for(int i=0;i<4;i++){
-    inputDim[i]=1;
-    outputDim[i]=1;
-  }
-    
-  int len = interpreter->tensor(input_idx_list[0])->dims->size;
-  std::reverse_copy(interpreter->tensor(input_idx_list[0])->dims->data,
-                   interpreter->tensor(input_idx_list[0])->dims->data+len, inputDim);
-  
-  len = interpreter->tensor(output_idx_list[0])->dims->size;
-  std::reverse_copy(interpreter->tensor(output_idx_list[0])->dims->data,
-                   interpreter->tensor(output_idx_list[0])->dims->data+len, outputDim);
-
-  printf("input %d %d %d %d\n",inputDim[0], inputDim[1], inputDim[2], inputDim[3]);
-  printf("output %d %d %d %d\n",outputDim[0], outputDim[1], outputDim[2], outputDim[3]);
-  
-  int output_number_of_pixels = 1;
-
-  int wanted_channels = inputDim[0];
-  int wanted_height = inputDim[1];
-  int wanted_width = inputDim[2];
-  
-  for(int k=0;k<4;k++)
-    output_number_of_pixels *= inputDim[k];
-
-  int input=interpreter->inputs()[0];  
-  
-  for(int i=0;i<TOTAL_LABEL_SIZE;i++){
-    std::string path = data_path;
-    path += total_label[i];
-    printf("\n[%s]\n", path.c_str());
-    
-    for(int j=0;j<TOTAL_DATA_SIZE;j++){
-      std::string img=path+"/";
-      img += total_label[i]+std::to_string(j+1)+".bmp";
-      printf("%s\n",img.c_str());
-
-      uint8_t *in;
-      float *output;
-      in = tflite::label_image::read_bmp(img, &wanted_width, &wanted_height, &wanted_channels);
-
-      if(interpreter->AllocateTensors() != kTfLiteOk){
-       std::cout << "Failed to allocate tnesors!" <<std::endl;
-       return -2;
-      }
-
-      for(int l=0;l<output_number_of_pixels;l++){
-       (interpreter->typed_tensor<float>(input))[l] =
-         ((float) in[l]-127.5f)/127.5f;
-      }
-
-      if(interpreter->Invoke() != kTfLiteOk) {
-       std::cout << "Failed to invoke!" <<std::endl;   
-       return -3;
-      }
-
-      output = interpreter->typed_output_tensor<float>(0);
-
-      std::copy(output, output+128, out[i][j]);
-      
-    }
-  }
-
-  for(int i=0;i<TOTAL_LABEL_SIZE;i++){
-    for(int j=0;j<TOTAL_DATA_SIZE;j++){
-      std::string out_file="/sdcard/Transfer-Learning/";
-      out_file += total_label[i]+std::to_string(j+1)+".txt";
-      printf("%s\n",out_file.c_str());
-      std::ofstream writeFile(out_file.data());
-      if(writeFile.is_open()){
-       for(int k=0;k<128;k++)
-         writeFile << out[i][j][k]<<std::endl;
-       writeFile.close();
-      }
-    }
-  }
-
-  
-  float testout[TOTAL_TEST_SIZE][128];
-  
-  for(int i=0;i<TOTAL_TEST_SIZE;i++){
-    std::string path = data_path;
-    path += "testset";
-    printf("\n[%s]\n", path.c_str());
-
-    std::string img = path+"/";
-    img += "test" + std::to_string(i+1)+".bmp";
-    printf("%s\n", img.c_str());
-
-    uint8_t *in;
-    float *output;
-    in=tflite::label_image::read_bmp(img, &wanted_width, &wanted_height, &wanted_channels);
-
-    if(interpreter->AllocateTensors() != kTfLiteOk){
-      std::cout << "Failed to allocate tnesors!" <<std::endl;
-      return -2;
-    }
-
-    for(int l=0;l<output_number_of_pixels;l++){
-      (interpreter->typed_tensor<float>(input))[l] =
-       ((float) in[l]-127.5f)/127.5f;
-    }
-
-    if(interpreter->Invoke() != kTfLiteOk) {
-      std::cout << "Failed to invoke!" <<std::endl;
-      return -3;
-    }
-
-    output = interpreter->typed_output_tensor<float>(0);
-    std::copy(output, output+128, testout[i]);
-    
-    int ret=0;
-    
-    ret=KNN(out, testout[i]);
-    printf("class %d\n", ret);
-  }
-
-  delete[] input_idx_list;
-  delete[] output_idx_list;
-
-  return 0;
-}

diff --git a/Training/jni/main.cpp b/Training/jni/main.cpp
index f7de4db..8ddc46d 100644 (file)
--- a/Training/jni/main.cpp
+++ b/Training/jni/main.cpp
@@ -166,15 +166,16 @@ int main(int argc, char *argv[]) {
   ExtractFeatures(data_path, inputVector, outputVector);
 
   Network::NeuralNetwork NN;
-  Network::NeuralNetwork NN2;
 
   NN.init(128, 20, TOTAL_LABEL_SIZE, 1, LEARNING_RATE, "sigmoid", true);
 
   NN.setOptimizer("sgd", LEARNING_RATE, 0.9, 0.999, 1e-8);
 
+  // NN.readModel("model.bin");
+
   for (int i = 0; i < ITERATION; i++) {
     for (unsigned int j = 0; j < inputVector.size(); j++) {
-      NN.backwarding(Matrix({inputVector[j]}), Matrix({outputVector[j]}), j);
+      NN.backwarding(Matrix({inputVector[j]}), Matrix({outputVector[j]}), i);
     }
     cout << "#" << i + 1 << "/" << ITERATION << " - Loss : " << NN.getLoss()
          << endl;