JOCLを使う

.gitignore

@@ -1,3 +1,4 @@
 /target/
 /nb-configuration.xml
-/nbactions*.xml
+/nbactions*.xml
+/*.log

pom.xml

@@ -14,13 +14,23 @@
         <dependency>
             <groupId>com.fasterxml.jackson.core</groupId>
             <artifactId>jackson-databind</artifactId>
-            <version>2.9.7</version>
+            <version>2.10.0</version>
         </dependency>
         <dependency>
             <groupId>org.projectlombok</groupId>
             <artifactId>lombok</artifactId>
             <version>1.18.2</version>
         </dependency>
+        <dependency>
+            <groupId>org.jogamp.gluegen</groupId>
+            <artifactId>gluegen-rt-main</artifactId>
+            <version>2.3.2</version>
+        </dependency>
+        <dependency>
+            <groupId>org.jogamp.jocl</groupId>
+            <artifactId>jocl-main</artifactId>
+            <version>2.3.2</version>
+        </dependency>
     </dependencies>
     <build>
         <plugins>
@@ -40,4 +50,4 @@
         <maven.compiler.source>1.8</maven.compiler.source>
         <maven.compiler.target>1.8</maven.compiler.target>
     </properties>
-</project>
+</project>

src/main/java/kishida/cnn/ConvolutionalNet.java

@@ -8,6 +8,8 @@
 import java.awt.Image;
 import java.awt.image.BufferedImage;
 import java.io.IOException;
+import java.io.InputStream;
+import java.io.InputStreamReader;
 import java.io.Reader;
 import java.io.UncheckedIOException;
 import java.io.Writer;
@@ -59,6 +61,8 @@ public class ConvolutionalNet {
     public static final String AVERAGE_PNG = "average.png";
     private static final String FILENAME = "C:\\Users\\naoki\\Desktop\\alexnet.json.txt";
     private static final String RESOURCE_NAME = "/alexnet_def.json";
+    //private static final String FILENAME = "C:\\Users\\naoki\\Desktop\\tinynet.json.txt";
+    //private static final String RESOURCE_NAME = "/tinynet_def.json";
 
     static class Img{
 
@@ -89,8 +93,9 @@ BufferedImage readImage(){
     static List<Double> historyData = new ArrayList<>();
     static LinkedList<Integer> rateData = new LinkedList<>();
 
+    @SuppressWarnings({"ThrowableInstanceNotThrown", "ThrowableInstanceNeverThrown"})
     public static void main(String[] args) throws IOException {
-        System.setProperty("com.amd.aparapi.enableShowGeneratedOpenCL", "false");
+        System.setProperty("com.aparapi.enableShowGeneratedOpenCL", "false");
         String def = "C:\\Users\\naoki\\Desktop\\sampleimg288";
         Path dir = Paths.get(args.length > 0 ? args[0] : def);
         List<String> categories = Files.list(dir)
@@ -184,14 +189,15 @@ public static void main(String[] args) throws IOException {
 
         NeuralNetwork nn;
 
-        /*
-        try(InputStream is = ConvolutionalNet.class.getResourceAsStream(RESOURCE_NAME);
-                InputStreamReader isr = new InputStreamReader(is)){
-            nn = NeuralNetwork.readFromJson(isr);
-        }*/
-
-        try(Reader r = Files.newBufferedReader(Paths.get(FILENAME))){
-            nn = NeuralNetwork.readFromJson(r);
+        if(true){
+            try(InputStream is = ConvolutionalNet.class.getResourceAsStream(RESOURCE_NAME);
+                    InputStreamReader isr = new InputStreamReader(is)){
+                nn = NeuralNetwork.readFromJson(isr);
+            }
+        }else{
+            try(Reader r = Files.newBufferedReader(Paths.get(FILENAME))){
+                nn = NeuralNetwork.readFromJson(r);
+            }
         }
 
         nn.init();
@@ -285,23 +291,13 @@ public static void main(String[] args) throws IOException {
                 Image lineGraph = createLineGraph(500, 200,
                         historyData, 1, 0);
                 historyLabel.setIcon(new ImageIcon(lineGraph));
-                //一段目のフィルタの表示
-                //全結合一段の表示
-                firstFc.setIcon(new ImageIcon(createGraph(256, 128, fc1.getResult())));
-                //全結合二段の表示
-                lastResult.setIcon(new ImageIcon(createGraph(256, 128, output)));
-
-                firstBias.setIcon(new ImageIcon(createGraph(500, 128, conv1.getBias())));
-                secondBias.setIcon(new ImageIcon(createGraph(500, 128,
-                        conv2.getBias())));
-                fc1Bias.setIcon(new ImageIcon(createGraph(500, 128, fc1.getBias())));
-                fc2Bias.setIcon(new ImageIcon(createGraph(500, 128, fc2.getBias())));
 
                 //System.out.println(Arrays.stream(output).mapToObj(d -> String.format("%.2f", d)).collect(Collectors.joining(",")));
 
                 count[0]++;
                 nn.setImageIndex(nn.getImageIndex() + 1);
                 if(count[0] >= MINI_BATCH){
+
                     nn.joinBatch();
                     batchCount[0]++;
                     System.out.printf("%5d %4d %.2f/m %s %s%n", batchCount[0],
@@ -319,7 +315,7 @@ public static void main(String[] args) throws IOException {
                             System.out.printf("weight: %.2f～%.2f average %.2f ",
                                     ws.getMin(), ws.getMax(), ws.getAverage());
                             DoubleSummaryStatistics bs = ((LerningLayer)layer).getBiasStatistics();
-                            System.out.printf("bias: %.2f～%.2f average %.2f ",
+                            System.out.printf("bias: %.8f～%.8f average %.2f ",
                                     bs.getMin(), bs.getMax(), bs.getAverage());
                         }
                         System.out.println();
@@ -329,6 +325,18 @@ public static void main(String[] args) throws IOException {
                     pStart[0] = System.currentTimeMillis();
                     nn.prepareBatch();
 
+                    //一段目のフィルタの表示
+                    //全結合一段の表示
+                    firstFc.setIcon(new ImageIcon(createGraph(256, 128, fc1.getResult())));
+                    //全結合二段の表示
+                    lastResult.setIcon(new ImageIcon(createGraph(256, 128, output)));
+
+                    firstBias.setIcon(new ImageIcon(createGraph(500, 128, conv1.getBias())));
+                    secondBias.setIcon(new ImageIcon(createGraph(500, 128,
+                            conv2.getBias())));
+                    fc1Bias.setIcon(new ImageIcon(createGraph(500, 128, fc1.getBias())));
+                    fc2Bias.setIcon(new ImageIcon(createGraph(500, 128, fc2.getBias())));
+
                     // 1時間に一回保存する
                     int hour = LocalTime.now().getHour();
                     if(lastHour[0] != hour){

src/main/java/kishida/cnn/NeuralNetwork.java

@@ -10,11 +10,13 @@
 import com.fasterxml.jackson.annotation.JsonProperty;
 import com.fasterxml.jackson.databind.ObjectMapper;
 import com.fasterxml.jackson.databind.SerializationFeature;
+import com.jogamp.opencl.CLBuffer;
 import java.io.IOException;
 import java.io.Reader;
 import java.io.StringReader;
 import java.io.StringWriter;
 import java.io.Writer;
+import java.nio.FloatBuffer;
 import java.util.ArrayList;
 import java.util.Arrays;
 import java.util.List;
@@ -23,6 +25,7 @@
 import java.util.stream.IntStream;
 import kishida.cnn.activation.LogisticFunction;
 import kishida.cnn.layers.ConvolutionLayer;
+import kishida.cnn.layers.FullGpuEnabled;
 import kishida.cnn.layers.FullyConnect;
 import kishida.cnn.layers.InputLayer;
 import kishida.cnn.layers.MaxPoolingLayer;
@@ -165,8 +168,23 @@ public float[] forward(float[] readData, float[] correctData){
             delta[idx] = correctData[idx] - output[idx];
         }
         //逆伝播
+        CLBuffer<FloatBuffer> bufDelta = null;
         for(int i = layers.size() - 1; i >= 1; --i){
-            delta = layers.get(i).backward(delta);
+            FullGpuEnabled layer = layers.get(i) instanceof FullGpuEnabled ?
+                    (FullGpuEnabled) layers.get(i) : null;
+            FullGpuEnabled pre = layers.get(i).getPreLayer() instanceof FullGpuEnabled ?
+                    (FullGpuEnabled)layers.get(i).getPreLayer() : null;
+            if(true && layer != null && pre != null && layer.isUseGpu()){
+                if(bufDelta == null){
+                    bufDelta = layer.backwardBuf(pre.getBufResult(), delta);
+                }else{
+                    bufDelta = layer.backwardBuf(pre.getBufResult(), bufDelta);
+                }
+                delta = null;
+            }else{
+                delta = layers.get(i).backward(delta);
+                bufDelta = null;
+            }
         }
 
         return output;

src/main/java/kishida/cnn/activation/ActivationFunction.java

@@ -5,6 +5,7 @@
  */
 package kishida.cnn.activation;
 
+import com.fasterxml.jackson.annotation.JsonIgnore;
 import com.fasterxml.jackson.annotation.JsonSubTypes;
 import com.fasterxml.jackson.annotation.JsonTypeInfo;
 
@@ -28,4 +29,6 @@ public void applyAfter(float[] values) {
     /** 微分 */
     public abstract float diff(float value);
 
+    @JsonIgnore
+    public abstract String getName();
 }

src/main/java/kishida/cnn/activation/LimitedRectifiedLinear.java

@@ -23,4 +23,9 @@ public float diff(float value) {
         return value >= 0 && value <= 2 ? 1 : 0;
     }
 
+    @Override
+    public String getName() {
+        return "limitrelu";
+    }
+
 }

src/main/java/kishida/cnn/activation/LinearFunction.java

@@ -20,5 +20,9 @@ public float apply(float value) {
     public float diff(float value) {
         return 1;
     }
+    @Override
+    public String getName() {
+        return "linear";
+    }
 
 }

src/main/java/kishida/cnn/activation/LogisticFunction.java

@@ -17,5 +17,9 @@ public float apply(float value) {
     public float diff(float value) {
         return value * (1 - value);
     }
+    @Override
+    public String getName() {
+        return "logistic";
+    }
 
 }

src/main/java/kishida/cnn/activation/RectifiedLinear.java

@@ -17,5 +17,9 @@ public float apply(float value) {
     public float diff(float value) {
         return value >= 0 ? 1 : 0;
     }
+    @Override
+    public String getName() {
+        return "relu";
+    }
 
 }

src/main/java/kishida/cnn/activation/SoftMaxFunction.java

@@ -30,4 +30,9 @@ public float diff(float value) {
         return value * (1 - value);
     }
 
+    @Override
+    public String getName() {
+        return "softmax";
+    }
+
 }

src/main/java/kishida/cnn/kernels/ConvolutionBackwordDeltaKernel.java

@@ -51,7 +51,6 @@ private void proc(int chxxyy) {
         }
         newDelta[chxxyy] = tempDelta;
     }
-    float[] input;
     float[] result;
     int inputChannels;
     int inputWidth;
@@ -65,11 +64,10 @@ private void proc(int chxxyy) {
     float[] delta;
     float[] newDelta;
 
-    public float[] backword(float[] input, float[] delta, float[] result,
+    public float[] backword(float[] delta, float[] result,
             int inputChannels, int inputWidth, int inputHeight,
             float[] filter, int outputChannels, int outputWidth, int outputHeight,
-            int filterSize, int stride, boolean useGpu) {
-        this.input = input;
+            int filterSize, int stride, float[] newDelta, boolean useGpu) {
         this.delta = delta;
         this.inputChannels = inputChannels;
         this.inputWidth = inputWidth;
@@ -81,12 +79,11 @@ public float[] backword(float[] input, float[] delta, float[] result,
         this.filterSize = filterSize;
         this.stride = stride;
         this.result = result;
-        this.newDelta = new float[inputChannels * inputWidth * inputHeight];
+        this.newDelta = newDelta;
         if (useGpu) {
             put(filter);
             put(delta);
             put(result);
-            put(input);
             execute(inputChannels * inputWidth * inputHeight);
             get(newDelta);
         } else {

src/main/java/kishida/cnn/kernels/ConvolutionBackwordFilterKernel.java

@@ -44,15 +44,15 @@ private void proc(int fchij) {
                 }
             }
         }
-        filter[fchij] += df;
+        filterDelta[fchij] += df;
     }
     float[] input;
     float[] result;
     float[] delta;
     int inputChannels;
     int inputWidth;
     int inputHeight;
-    float[] filter;
+    float[] filterDelta;
     int outputChannels;
     int outputWidth;
     int outputHeight;
@@ -62,14 +62,14 @@ private void proc(int fchij) {
 
     public void backword(float[] delta, float[] result,
             float[] input, int inputChannels, int inputWidth, int inputHeight,
-            float[] filter, int outputChannels, int outputWidth, int outputHeight,
+            float[] filterDelta, int outputChannels, int outputWidth, int outputHeight,
             int filterSize, int stride, float learningRate, boolean useGpu) {
         this.input = input;
         this.delta = delta;
         this.inputChannels = inputChannels;
         this.inputWidth = inputWidth;
         this.inputHeight = inputHeight;
-        this.filter = filter;
+        this.filterDelta = filterDelta;
         this.outputChannels = outputChannels;
         this.outputWidth = outputWidth;
         this.outputHeight = outputHeight;
@@ -79,11 +79,11 @@ public void backword(float[] delta, float[] result,
         this.learningRate = learningRate;// / outputWidth;// * outputHeight);
         if (useGpu) {
             put(delta);
-            put(filter);
+            put(filterDelta);
             put(input);
             put(result);
             execute(outputChannels * inputChannels * filterSize * filterSize);
-            get(filter);
+            get(filterDelta);
         } else {
             IntStream.range(0, outputChannels).parallel().forEach((f) -> {
                 for (int chij = 0; chij < inputChannels * filterSize * filterSize; ++chij) {

src/main/java/kishida/cnn/kernels/ConvolutionBackwordKernel.java

@@ -62,7 +62,7 @@ private void proc(int fxy) {
     int outputHeight;
     int filterSize;
     int stride;
-    float[] bias;
+    //float[] bias;
     float[] delta;
     float learningRate;
     float[] tempDelta;
@@ -74,7 +74,7 @@ public float[] backward(float[] delta, float[] result,
             float[] input, int inputChannels, int inputWidth, int inputHeight,
             float[] filter, int outputChannels, int outputWidth, int outputHeight,
             float[] filterDelta, float[] biasDelta,
-            int filterSize, int stride, float[] bias, float learningRate, boolean useGpu) {
+            int filterSize, int stride, float learningRate, boolean useGpu) {
         this.delta = delta;
         this.input = input;
         this.inputChannels = inputChannels;
@@ -86,7 +86,7 @@ public float[] backward(float[] delta, float[] result,
         this.outputHeight = outputHeight;
         this.filterSize = filterSize;
         this.stride = stride;
-        this.bias = bias;
+        //this.bias = bias;
         this.result = result;
         this.tempDelta = new float[outputChannels * inputChannels * inputWidth * inputHeight];
         this.learningRate = learningRate;// / (outputWidth * outputHeight);