Complete example

example/.gitignore

 train.log
 target/
-data/vGG16/mnist.pkl.gz
-data/vGG16/st.h5
+resources/training_data
 model/
+build/
\ No newline at end of file

example/build.sh

+rm -r target/; 
+java -jar embedded-montiarc-emadl-generator-0.2.5-SNAPSHOT-jar-with-dependencies.jar -m src/emadl/models/ -r cNNCalculator.Connector -o target -b MXNET -p /home/christopher/anaconda3/bin/python
+
 rm -rf build
 mkdir build && cd build
 
 echo "Building DigitClassifier.."
 cmake ..
-cp -r /home/christopher/Documents/Master/Semester2/Praktikum/mxnet/include/mxnet/ ../target/
+#cp -r /home/christopher/Documents/Master/Semester2/Praktikum/mxnet/include/mxnet/ ../target/
 make
\ No newline at end of file

example/prepare.sh

+echo "Creating HDF5 dataset from image files.."
+cd resources/
+wget https://git.rwth-aachen.de/thomas.timmermanns/EMADL-Demo/raw/master/src/resources/data.tar.gz
+tar xf data.tar.gz
+rm data.tar.gz
+python imgDir-to-h5.py --in_port data --out_port softmax --data_path data --target_path training_data
+rm -r data
\ No newline at end of file

example/resources/imgDir-to-h5.py

+from __future__ import print_function
+from __future__ import division
+import h5py
+import numpy as np
+import cv2
+import os
+import argparse
+import errno
+import random
+import sys
+
+
+def create_img_list(name, data_path):
+    dir_name = data_path + "/" + name
+
+    image_paths = []
+    image_class_indices = []
+    print(dir_name)
+    for class_index_name in os.listdir(dir_name):
+        class_dir_path = dir_name + "/" + class_index_name
+        if os.path.isdir(class_dir_path):
+            for image_name in os.listdir(class_dir_path):
+                image_path = class_dir_path + "/" + image_name
+                image_paths.append(image_path)
+                class_index = float(class_index_name)
+                image_class_indices.append(class_index)
+    return image_paths, image_class_indices
+
+def create_h5_from_list(image_paths, image_class_indices, target_dir, target_file_name, input_port_name, output_port_name, shuffle=True):
+    img = cv2.imread(image_paths[0])
+    t_img = np.transpose(img, (2,0,1)).astype(np.float32)
+    #t_img = t_img[-1:,:,:]
+    #print(t_img)
+    channels = t_img.shape[0]
+    height = t_img.shape[1]
+    width = t_img.shape[2]
+    data_size = len(image_paths)
+    
+    target_file = target_dir + "/" + target_file_name + ".h5"
+    if os.path.isfile(target_file):
+        print("File", target_file, "already exists. Skipping data file creation.")
+        return
+    try:
+        os.makedirs(target_dir)
+    except OSError as e:
+        if e.errno != errno.EEXIST:
+            raise
+
+    if shuffle:
+        combined = list(zip(image_paths, image_class_indices))
+        random.shuffle(combined)
+        image_paths[:], image_class_indices[:] = zip(*combined)
+    
+    print("Creating " + target_file + " (images:" + str(data_size) + ", channels:" + str(channels) + ", height:" + str(height) + ", width:" + str(width) + "):")
+    with h5py.File(target_file, "w") as ofile:
+        in_dset = ofile.create_dataset(input_port_name, (data_size,channels, height, width), dtype=np.float32)
+        out_dset = ofile.create_dataset(output_port_name + "_label", (data_size,), dtype=np.float32)
+        for i in range(data_size):
+            img = cv2.imread(image_paths[i])
+            t_img = np.transpose(img, (2,0,1)).astype(np.float32)
+            #t_img = t_img[-1:,:,:]
+            in_dset[i] = t_img
+            out_dset[i] = image_class_indices[i]
+
+            #print progress
+            if i % 100 == 0:
+                percentage = 100*i / data_size
+                sys.stdout.write("\r{:0.1f}%".format(percentage))
+                sys.stdout.flush()
+    sys.stdout.write("\r100.0%\n")
+    sys.stdout.flush()
+
+
+            
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description='Translate image directories into hdf5 training sets for EMADL.')
+    parser.add_argument("--in_port", action="store", dest="in_port", default="data")
+    parser.add_argument("--out_port", action="store", dest="out_port", default="softmax")
+    parser.add_argument("--data_path", action="store", dest="data_path", default=".")
+    parser.add_argument("--target_path", action="store", dest="target_path", default=".")
+    args = parser.parse_args()
+    for file_name in os.listdir(args.data_path):
+        if file_name == "train":
+            image_paths, image_class_indices = create_img_list(file_name, args.data_path)
+            create_h5_from_list(image_paths, image_class_indices, args.target_path, file_name, args.in_port, args.out_port)
+        if file_name == "test":
+            image_paths, image_class_indices = create_img_list(file_name, args.data_path)
+            create_h5_from_list(image_paths, image_class_indices, args.target_path, file_name, args.in_port, args.out_port)

example/resources/sort.sh

+mkdir 0
+mv *-num0.png 0/.
+
+mkdir 1
+mv *-num1.png 1/.
+
+
+mkdir 2
+mv *-num2.png 2/.
+
+
+mkdir 3
+mv *-num3.png 3/.
+
+mkdir 4
+mv *-num4.png 4/.
+
+
+mkdir 5
+mv *-num5.png 5/.
+
+
+mkdir 6
+mv *-num6.png 6/.
+
+
+mkdir 7
+mv *-num7.png 7/.
+
+
+mkdir 8
+mv *-num8.png 8/.
+
+
+mkdir 9
+mv *-num9.png 9/.
+
+
+

example/src/cpp/DigitCalculator.cpp

@@ -10,52 +10,63 @@
 #include <map>
 
 int main(int argc, char* argv[]) {
-    if(argc < 2){ //Note: argc=1 if no arguments are provided
-        std::cout << "Missing argument: Path to test image must be provided " << std::endl;
+    if(argc < 7){ //Note: argc=1 if no arguments are provided
+        std::cout << "Missing argument: Path to 6 test images must be provided " << std::endl;
         exit(1);
     }
-    std::string filePath = argv[1];
+    
 
-    if (!std::ifstream(filePath).good()) {
-        std::cerr << "Image loading failure, test image '" << filePath << "' does not exist." << std::endl;
-        exit(1);
-    }
+    cNNCalculator_connector connector;
+    connector.init();
+
+    for(int n = 0; n < 6; n++) {
+        std::string filePath = argv[n+1];
 
-    cv::Mat img = cv::imread(filePath);
-    std::cout << "== original image size: " << img.size() << " ==" << std::endl;
+        if (!std::ifstream(filePath).good()) {
+            std::cerr << "Image loading failure, test image '" << filePath << "' does not exist." << std::endl;
+            exit(1);
+        }
 
-    // scale image to fit
-    cv::Size scale(28,28);
-    cv::resize(img, img, scale);
-    std::cout << "== simply resize: " << img.size() << " ==" << std::endl;
+        cv::Mat img = cv::imread(filePath);
+        std::cout << "== original image size: " << img.size() << " ==" << std::endl;
 
-    size_t channels = 1;
-    size_t height = img.rows;
-    size_t width = img.cols;
-    vector<float> data(channels*height*width);
+        // scale image to fit
+        cv::Size scale(32,32);
+        cv::resize(img, img, scale);
+        std::cout << "== simply resize: " << img.size() << " ==" << std::endl;
 
-    for(size_t j=0; j<height; j++){
-        for(size_t k=0; k<width; k++){
-            cv::Vec3b intensity = img.at<cv::Vec3b>(j, k);
-            for(size_t i=0; i<channels; i++){
-                data[i*height*width + j*height + k] = (float) intensity[i];
+        size_t channels = 3;
+        size_t height = img.rows;
+        size_t width = img.cols;
+        vector<float> data(channels*height*width);
+
+        for(size_t j=0; j<height; j++){
+            for(size_t k=0; k<width; k++){
+                cv::Vec3b intensity = img.at<cv::Vec3b>(j, k);
+                for(size_t i=0; i<channels; i++){
+                    data[i*height*width + j*height + k] = (float) intensity[i];
+                }
             }
         }
+
+        if(n == 0)
+            connector.image1 = conv_to< icube >::from( CNNTranslator::translateToCube(data, vector<size_t> {channels,height,width}) );
+        if(n == 1)
+            connector.image2 = conv_to< icube >::from( CNNTranslator::translateToCube(data, vector<size_t> {channels,height,width}) );
+        if(n == 2)
+            connector.image3 = conv_to< icube >::from( CNNTranslator::translateToCube(data, vector<size_t> {channels,height,width}) );
+        if(n == 3)
+            connector.image4 = conv_to< icube >::from( CNNTranslator::translateToCube(data, vector<size_t> {channels,height,width}) );
+        if(n == 4)
+            connector.image5 = conv_to< icube >::from( CNNTranslator::translateToCube(data, vector<size_t> {channels,height,width}) );
+        if(n == 5)
+            connector.image6 = conv_to< icube >::from( CNNTranslator::translateToCube(data, vector<size_t> {channels,height,width}) );
     }
 
-    cNNCalculator_connector connector;
-    connector.init();
-    connector.image1 = conv_to< icube >::from( CNNTranslator::translateToCube(data, vector<size_t> {channels,height,width}) );
-    connector.image2 = conv_to< icube >::from( CNNTranslator::translateToCube(data, vector<size_t> {channels,height,width}) );
-    connector.image3 = conv_to< icube >::from( CNNTranslator::translateToCube(data, vector<size_t> {channels,height,width}) );
-    connector.image4 = conv_to< icube >::from( CNNTranslator::translateToCube(data, vector<size_t> {channels,height,width}) );
-    connector.image5 = conv_to< icube >::from( CNNTranslator::translateToCube(data, vector<size_t> {channels,height,width}) );
-    connector.image6 = conv_to< icube >::from( CNNTranslator::translateToCube(data, vector<size_t> {channels,height,width}) );
     connector.execute();
 
     int classIndex = (int)connector.res;
     std::cout << "== SUM: " << classIndex << std::endl;
 
     return 0;
-
 }

example/src/emadl/models/cNNCalculator/Calculator.emadl

@@ -22,12 +22,12 @@ component Calculator {
         instance Add add;
         
 
-        connect in1_1 -> number1_ones.inputVector;
+        connect in1_1 -> number1_hundreds.inputVector;
         connect in1_2 -> number1_tens.inputVector;
-        connect in1_3 -> number1_hundreds.inputVector;
-        connect in2_1 -> number2_ones.inputVector;
+        connect in1_3 -> number1_ones.inputVector;
+        connect in2_1 -> number2_hundreds.inputVector;
         connect in2_2 -> number2_tens.inputVector;
-        connect in2_3 -> number2_hundreds.inputVector;
+        connect in2_3 -> number2_ones.inputVector;
 
         connect number1_ones.maxIndex -> number1.ones;
         connect number1_tens.maxIndex -> number1.tens;

example/src/emadl/models/cNNCalculator/Connector.emadl

@@ -9,27 +9,29 @@ component Connector {
           in Z(0:255)^{1, 28, 28} image6,
           out Z(0:1998) res;
 
-    instance VGG16 predictor1;
-    instance VGG16 predictor2;
-    instance VGG16 predictor3;
-    instance VGG16 predictor4;
-    instance VGG16 predictor5;
-    instance VGG16 predictor6;
+    instance Network<10> predictor1;
+    instance Network<10> predictor2;
+    instance Network<10> predictor3;
+    instance Network<10> predictor4;
+    instance Network<10> predictor5;
+    instance Network<10> predictor6;
     instance Calculator cal;
 
-    connect image1 -> predictor1.image;
-    connect image2 -> predictor2.image;
-    connect image3 -> predictor3.image;
-    connect image4 -> predictor4.image;
-    connect image5 -> predictor5.image;
-    connect image6 -> predictor6.image;
+    instance ArgMax<10> maxi;
 
-    connect predictor1.predictions -> cal.in1_1;
-    connect predictor2.predictions -> cal.in1_2;
-    connect predictor3.predictions -> cal.in1_3;
-    connect predictor4.predictions -> cal.in2_1;
-    connect predictor5.predictions -> cal.in2_2;
-    connect predictor6.predictions -> cal.in2_3;
+    connect image1 -> predictor1.data;
+    connect image2 -> predictor2.data;
+    connect image3 -> predictor3.data;
+    connect image4 -> predictor4.data;
+    connect image5 -> predictor5.data;
+    connect image6 -> predictor6.data;
+
+    connect predictor1.softmax -> cal.in1_1;
+    connect predictor2.softmax -> cal.in1_2;
+    connect predictor3.softmax -> cal.in1_3;
+    connect predictor4.softmax -> cal.in2_1;
+    connect predictor5.softmax -> cal.in2_2;
+    connect predictor6.softmax -> cal.in2_3;
 
     connect cal.out1 -> res;
 

example/src/emadl/models/cNNCalculator/Network.cnnt

+configuration Network{
+     num_epoch:5
+     batch_size:100
+     normalize:true
+     context:gpu
+     load_checkpoint:false
+     optimizer:sgd{
+         learning_rate:0.2
+         learning_rate_decay:0.85
+         step_size:1000
+         weight_decay:0.0
+     }
+}

example/src/emadl/models/cNNCalculator/Network.emadl

+package cNNCalculator;
+
+component Network<Z(2:oo) classes = 10>{
+    ports in Z(0:255)^{3, 32, 32} data,
+         out Q(0:1)^{classes} softmax;
+
+    implementation CNN {
+        def conv(channels, kernel=1, stride=1){
+            Convolution(kernel=(kernel,kernel),channels=channels) ->
+            Relu() ->
+            Pooling(pool_type="max", kernel=(2,2), stride=(stride,stride))
+        }
+
+        data ->
+        conv(kernel=5, channels=20, stride=2) ->
+        conv(kernel=5, channels=50, stride=2) ->
+        FullyConnected(units=500) ->
+        Relu() ->
+        Dropout() ->
+        FullyConnected(units=classes) ->
+        Softmax() ->
+        softmax
+    }
+}

example/src/emadl/models/data_paths.txt

-cNNCalculator.VGG16 src/resources/data/vGG16
+cNNCalculator.Network resources/training_data