Add File

src/main/java/org/dromara/easyai/transFormer/TransFormerManager.java

@@ -0,0 +1,135 @@
+package org.dromara.easyai.transFormer;
+
+import org.dromara.easyai.config.TfConfig;
+import org.dromara.easyai.matrixTools.Matrix;
+import org.dromara.easyai.transFormer.model.CodecBlockModel;
+import org.dromara.easyai.transFormer.model.TransFormerModel;
+import org.dromara.easyai.transFormer.model.TransWordVectorModel;
+import org.dromara.easyai.transFormer.nerve.SensoryNerve;
+
+import java.util.ArrayList;
+import java.util.List;
+
+public class TransFormerManager {
+    private final List<CodecBlock> encoderBlocks = new ArrayList<>();//编码器模块
+    private final List<CodecBlock> decoderBlocks = new ArrayList<>();//解码器模块
+    private SensoryNerve sensoryNerve;//感知神经元
+    private FirstDecoderBlock firstDecoderBlock;//第一个解码器模块
+    private LineBlock lineBlock;//线性分类层
+    private TransWordVector transWordVector;//内置词向量
+
+    public TransWordVector getTransWordVector() {
+        return transWordVector;
+    }
+
+    public SensoryNerve getSensoryNerve() {
+        return sensoryNerve;
+    }
+
+    public TransFormerModel getModel() throws Exception {
+        TransFormerModel transFormerModel = new TransFormerModel();
+        transFormerModel.setTransWordVectorModel(transWordVector.getModel());
+        List<CodecBlockModel> encoderBlockModels = new ArrayList<>();
+        List<CodecBlockModel> decoderBlockModels = new ArrayList<>();
+        for (int i = 0; i < encoderBlocks.size(); i++) {
+            encoderBlockModels.add(encoderBlocks.get(i).getModel());
+            decoderBlockModels.add(decoderBlocks.get(i).getModel());
+        }
+        transFormerModel.setEncoderBlockModels(encoderBlockModels);
+        transFormerModel.setDecoderBlockModels(decoderBlockModels);
+        transFormerModel.setFirstDecoderBlockModel(firstDecoderBlock.getModel());
+        transFormerModel.setLineBlockModel(lineBlock.getModel());
+        return transFormerModel;
+    }
+
+
+    public void insertModel(TransFormerModel transFormerModel, TfConfig tfConfig) throws Exception {
+        init(tfConfig, null, transFormerModel.getTransWordVectorModel());
+        List<CodecBlockModel> encoderBlockModels = transFormerModel.getEncoderBlockModels();
+        List<CodecBlockModel> decoderBlockModels = transFormerModel.getDecoderBlockModels();
+        int minSize = Math.min(encoderBlocks.size(), encoderBlockModels.size());
+        for (int i = 0; i < minSize; i++) {
+            encoderBlocks.get(i).insertModel(encoderBlockModels.get(i));
+            decoderBlocks.get(i).insertModel(decoderBlockModels.get(i));
+        }
+        firstDecoderBlock.insertModel(transFormerModel.getFirstDecoderBlockModel());
+        lineBlock.insertModel(transFormerModel.getLineBlockModel());
+    }
+
+    public void init(TfConfig tfConfig, List<String> sentenceList) throws Exception {
+        if (transWordVector == null) {
+            init(tfConfig, sentenceList, null);
+        } else {
+            transWordVector.init(sentenceList);
+        }
+    }
+
+    /**
+     * 初始化神经元参数
+     *
+     * @param tfConfig             配置参数
+     * @param sentenceList         样本语句
+     * @param transWordVectorModel 词向量模型
+     * @throws Exception 如果参数错误则抛异常
+     */
+    private void init(TfConfig tfConfig, List<String> sentenceList, TransWordVectorModel transWordVectorModel) throws Exception {
+        transWordVector = new TransWordVector(tfConfig);
+        int typeNumber = tfConfig.getTypeNumber();
+        if (transWordVectorModel == null) {
+            transWordVector.init(sentenceList);
+        } else {
+            transWordVector.insertModel(transWordVectorModel);
+        }
+        if (tfConfig.isNorm()) {
+            typeNumber = transWordVector.getWordSize();
+        }
+        int multiNumber = tfConfig.getMultiNumber();
+        int featureDimension = tfConfig.getFeatureDimension();
+        if (featureDimension % 2 != 0) {
+            throw new Exception("TransFormer 词向量维度必须为偶数");
+        }
+        int allDepth = tfConfig.getAllDepth();
+        float studyPoint = tfConfig.getStudyRate();
+        boolean showLog = tfConfig.isShowLog();
+        int regularModel = tfConfig.getRegularModel();
+        float regular = tfConfig.getRegular();
+        if (multiNumber > 1 && featureDimension > 0 && allDepth > 0 && typeNumber > 1) {
+            for (int i = 0; i < allDepth; i++) {
+                CodecBlock encoderBlock = new CodecBlock(multiNumber, featureDimension, studyPoint,
+                        i + 1, true, regularModel, regular, tfConfig.getCoreNumber(), transWordVector);
+                encoderBlocks.add(encoderBlock);
+            }
+            CodecBlock lastEnCoderBlock = encoderBlocks.get(encoderBlocks.size() - 1);//最后一层编码器
+            for (int i = 0; i < allDepth; i++) {
+                CodecBlock decoderBlock = new CodecBlock(multiNumber, featureDimension, studyPoint,
+                        i + 2, false, regularModel, regular, tfConfig.getCoreNumber(), transWordVector);
+                decoderBlock.setLastEncoderBlock(lastEnCoderBlock);//放入最优一层编码器
+                decoderBlocks.add(decoderBlock);
+            }
+            CodecBlock lastDecoderBlock = decoderBlocks.get(decoderBlocks.size() - 1);
+            connectCodecBlock(encoderBlocks);
+            connectCodecBlock(decoderBlocks);
+            lineBlock = new LineBlock(typeNumber, featureDimension, studyPoint, lastDecoderBlock, showLog, regularModel
+                    , regular, tfConfig.getCoreNumber(), tfConfig.getTimePunValue());
+            lastDecoderBlock.setLineBlock(lineBlock);
+            firstDecoderBlock = new FirstDecoderBlock(multiNumber, featureDimension, studyPoint, decoderBlocks.get(0),
+                    tfConfig.getCoreNumber(), transWordVector);
+            firstDecoderBlock.setLastEncoderBlock(lastEnCoderBlock);
+            decoderBlocks.get(0).setFirstDecoderBlock(firstDecoderBlock);
+            sensoryNerve = new SensoryNerve(encoderBlocks.get(0), firstDecoderBlock, transWordVector);
+        } else {
+            throw new Exception("param is null,typeNumber:" + typeNumber + ",featureDimension:" + featureDimension);
+        }
+    }
+
+    private void connectCodecBlock(List<CodecBlock> codecBlocks) {
+        int size = codecBlocks.size();
+        for (int i = 0; i < size - 1; i++) {
+            CodecBlock encoderBlock = codecBlocks.get(i);
+            CodecBlock beforeBlock = codecBlocks.get(i + 1);
+            encoderBlock.setBeforeEncoderBlock(beforeBlock);
+            beforeBlock.setAfterEncoderBlock(encoderBlock);
+        }
+    }
+
+}