import io.leavesfly.tinyai.ndarr.NdArray; import io.leavesfly.tinyai.ndarr.Shape; // 创建数组 NdArray a = NdArray.of(new float[][]{{1, 2}, {3, 4}}); NdArray b = NdArray.zeros(Shape.of(2, 3)); NdArray c = NdArray.randn(Shape.of(100, 50)); // 数学运算 NdArray result = a.add(b) // 加法 .mul(c) // 对应元素相乘 .dot(d) // 矩阵乘法 .sigmoid() // Sigmoid激活 .transpose(); // 转置

import io.leavesfly.tinyai.func.Variable; // 创建变量 Variable x = new Variable(NdArray.of(new float[][]{{1, 2}, {3, 4}}), "x"); Variable y = new Variable(NdArray.of(new float[][]{{2, 3}, {4, 5}}), "y"); // 计算 Variable z = x.add(y); Variable w = z.sigmoid(); // 反向传播 w.backward(); // 获取梯度 NdArray xGrad = x.getGrad(); System.out.println("x的梯度: " + xGrad);

import io.leavesfly.tinyai.nnet.block.*; import io.leavesfly.tinyai.nnet.v1.dnn.layer.LinearLayer; import io.leavesfly.tinyai.nnet.v1.activate.layer.ReLuLayer; // 构建MLP MlpBlock mlp = new MlpBlock("classifier", 784, new int[]{128, 64, 10}); // 或者手动构建 SequentialBlock model = new SequentialBlock("manual_model"); model.addLayer(new LinearLayer("fc1", 784, 256)); model.addLayer(new ReLuLayer("relu1")); model.addLayer(new LinearLayer("fc2", 256, 10)); // 前向传播 Variable output = model.blockForward(input);

import io.leavesfly.tinyai.ml.*; // 准备数据 DataSet dataSet = new ArrayDataset(xData, yData); // 创建模型 MlpBlock mlpBlock = new MlpBlock("classifier", 784, new int[]{128, 64, 10}); Model model = new Model("mnist_classifier", mlpBlock); // 配置训练器 Monitor monitor = new Monitor(); Evaluator evaluator = new Evaluator(); Trainer trainer = new Trainer(100, monitor, evaluator, true, 4); // 并行训练 // 初始化训练器 trainer.init(dataSet, model, new SoftmaxCrossEntropyLoss(), new Adam(0.001)); // 开始训练 trainer.train(true);

import io.leavesfly.tinyai.rl.*; // 创建环境 CartPoleEnvironment env = new CartPoleEnvironment(); // 创建DQN智能体 DQNAgent agent = new DQNAgent( "CartPole-DQN", stateDim, actionDim, hiddenSizes, learningRate, epsilon, gamma, batchSize, bufferSize, updateFreq ); // 训练循环 for (int episode = 0; episode < maxEpisodes; episode++) { State state = env.reset(); while (!done) { Action action = agent.selectAction(state); StepResult result = env.step(action); agent.store(state, action, result.reward, result.nextState, result.done); agent.train(); state = result.nextState; } }

import io.leavesfly.tinyai.nl.*; // 词嵌入 WordEmbedding embedding = new WordEmbedding("embedding", vocabSize, embeddingDim); Variable embedded = embedding.layerForward(inputIds); // LSTM LSTM lstm = new LSTM("lstm", inputSize, hiddenSize, numLayers, dropout); Variable output = lstm.layerForward(embedded); // 注意力机制 Attention attention = new Attention("attention", hiddenSize); Variable context = attention.layerForward(query, key, value);

import io.leavesfly.tinyai.nnet.Layer; public class CustomLayer extends Layer { @Override public Variable layerForward(Variable... inputs) { // 实现自定义前向传播 return customComputation(inputs[0]); } @Override public void init() { // 参数初始化 addParameter("weight", NdArray.randn(inputShape, outputShape)); addParameter("bias", NdArray.zeros(outputShape)); } }

import io.leavesfly.tinyai.ml.Optimizer; public class CustomOptimizer extends Optimizer { @Override public void step() { // 实现自定义参数更新逻辑 for (Map.Entry<String, Variable> entry : parameters.entrySet()) { Variable param = entry.getValue(); NdArray grad = param.getGrad(); // 自定义更新规则 updateParameter(param, grad); } } }

import io.leavesfly.tinyai.ml.LossFunction; public class CustomLoss implements LossFunction { @Override public Variable computeLoss(Variable predicted, Variable target) { // 实现自定义损失计算 return customLossComputation(predicted, target); } }