tensorflow.js

Author: csxiaoyaojianxian

20-tensorflow.js/04- normalize/index.html renamed to 20-tensorflow.js/04-normalize/index.html

@@ -1,34 +1,40 @@
+/*
+ * @Author: victorsun
+ * @Date: 2019-12-04 20:15:29
+ * @LastEditors: victorsun - csxiaoyao
+ * @LastEditTime: 2020-03-21 18:34:40
+ * @Description: sunjianfeng@csxiaoyao.com
+ */
 export function getData(numSamples) {
-    let points = [];
-  
-    function genGauss(cx, cy, label) {
-      for (let i = 0; i < numSamples / 2; i++) {
-        let x = normalRandom(cx);
-        let y = normalRandom(cy);
-        points.push({ x, y, label });
-      }
+  let points = [];
+  // 生成高斯分布(正态分布)的点
+  function genGauss(cx, cy, label) {
+    // numSamples 分成两拨
+    for (let i = 0; i < numSamples / 2; i++) {
+      let x = normalRandom(cx);
+      let y = normalRandom(cy);
+      points.push({ x, y, label });
     }
-  
-    genGauss(2, 2, 1);
-    genGauss(-2, -2, 0);
-    return points;
   }
-  
-  /**
-   * Samples from a normal distribution. Uses the seedrandom library as the
-   * random generator.
-   *
-   * @param mean The mean. Default is 0.
-   * @param variance The variance. Default is 1.
-   */
-  function normalRandom(mean = 0, variance = 1) {
-    let v1, v2, s;
-    do {
-      v1 = 2 * Math.random() - 1;
-      v2 = 2 * Math.random() - 1;
-      s = v1 * v1 + v2 * v2;
-    } while (s > 1);
-  
-    let result = Math.sqrt(-2 * Math.log(s) / s) * v1;
-    return mean + Math.sqrt(variance) * result;
-  }
+  genGauss(2, 2, 1);
+  genGauss(-2, -2, 0);
+  return points;
+}
+
+/**
+ * Box-Muller transform 算法
+ * @param mean 正态分布中心值
+ * @param variance 密集程度
+ */
+function normalRandom(mean = 0, variance = 1) {
+  let v1, v2, s;
+  // Math.random() 0-1均匀分布
+  // 原正态分布公式使用了 sin cos，性能较低，此处进行了优化
+  do {
+    v1 = 2 * Math.random() - 1;
+    v2 = 2 * Math.random() - 1;
+    s = v1 * v1 + v2 * v2;
+  } while (s > 1);
+  let result = Math.sqrt(-2 * Math.log(s) / s) * v1;
+  return mean + Math.sqrt(variance) * result;
+}

20-tensorflow.js/04- normalize/script.js renamed to 20-tensorflow.js/04-normalize/script.js

@@ -2,7 +2,7 @@
  * @Author: victorsun
  * @Date: 2019-12-04 20:15:29
  * @LastEditors: victorsun - csxiaoyao
- * @LastEditTime: 2020-03-21 13:27:25
+ * @LastEditTime: 2020-03-22 00:05:54
  * @Description: sunjianfeng@csxiaoyao.com
  */
 import * as tf from '@tensorflow/tfjs';
@@ -31,16 +31,16 @@ window.onload = async () => {
 
     // 2. 初始化神经网络模型
     const model = tf.sequential();
-    // 添加层，dense： y=ax+b，设置激活函数(防止输入超过100%，对过大过小值收敛，保证数据在 0 - 1 之间)，常用 sigmoid
+    // 添加层，dense： y=ax+b，设置激活函数sigmoid(防止输入超过100%，对过大过小值收敛，保证数据在 0 - 1 之间)
     model.add(tf.layers.dense({
         units: 1, // 输出值为一个概率值，1个神经元即可
         inputShape: [2], // 坐标 x,y 两个值，特征数量为2
-        activation: 'sigmoid' // 设置激活函数 sigmoid
+        activation: 'sigmoid' // 设置激活函数 sigmoid 0-1
     }));
     // 设置损失函数和优化器
     model.compile({
-        loss: tf.losses.logLoss,
-        optimizer: tf.train.adam(0.1) // adam自动调节学习速率
+        loss: tf.losses.logLoss, // 损失函数，log损失，用于逻辑回归问题
+        optimizer: tf.train.adam(0.1) // adam自动调节学习速率，初始化学习速率0.1
     });
 
     // 3. 训练数据转tensor

20-tensorflow.js/05-logistic-regression/data.js

@@ -0,0 +1,44 @@
+/*
+ * @Author: victorsun
+ * @Date: 2019-12-04 20:15:29
+ * @LastEditors: victorsun - csxiaoyao
+ * @LastEditTime: 2020-03-21 18:53:29
+ * @Description: sunjianfeng@csxiaoyao.com
+ */
+// 参考05逻辑回归中data
+export function getData(numSamples) {
+  let points = [];
+
+  function genGauss(cx, cy, label) {
+    for (let i = 0; i < numSamples / 2; i++) {
+      let x = normalRandom(cx);
+      let y = normalRandom(cy);
+      points.push({ x, y, label });
+    }
+  }
+
+  genGauss(2, 2, 0);
+  genGauss(-2, -2, 0);
+  genGauss(-2, 2, 1);
+  genGauss(2, -2, 1);
+  return points;
+}
+
+/**
+ * Samples from a normal distribution. Uses the seedrandom library as the
+ * random generator.
+ *
+ * @param mean The mean. Default is 0.
+ * @param variance The variance. Default is 1.
+ */
+function normalRandom(mean = 0, variance = 1) {
+  let v1, v2, s;
+  do {
+    v1 = 2 * Math.random() - 1;
+    v2 = 2 * Math.random() - 1;
+    s = v1 * v1 + v2 * v2;
+  } while (s > 1);
+
+  let result = Math.sqrt(-2 * Math.log(s) / s) * v1;
+  return mean + Math.sqrt(variance) * result;
+}

20-tensorflow.js/05-logistic-regression/script.js

@@ -0,0 +1,6 @@
+<script src="script.js"></script>
+<form action="" onsubmit="predict(this);return false;">
+    x: <input type="text" name="x">
+    y: <input type="text" name="y">
+    <button type="submit">预测</button>
+</form>

20-tensorflow.js/06-xor/data.js

@@ -0,0 +1,70 @@
+/*
+ * @Author: victorsun
+ * @Date: 2019-12-04 20:15:29
+ * @LastEditors: victorsun - csxiaoyao
+ * @LastEditTime: 2020-03-22 00:06:21
+ * @Description: sunjianfeng@csxiaoyao.com
+ */
+import * as tf from '@tensorflow/tfjs';
+import * as tfvis from '@tensorflow/tfjs-vis';
+import { getData } from './data.js';
+
+/**
+ * XOR 异或(非线性)逻辑回归
+ * 分割为四个象限
+ */
+window.onload = async () => {
+    // 1. 准备训练数据
+    const data = getData(400);
+    // 散点图
+    tfvis.render.scatterplot(
+        { name: 'XOR 训练数据' },
+        {
+            values: [
+                data.filter(p => p.label === 1),
+                data.filter(p => p.label === 0),
+            ]
+        }
+    );
+    // 2. 初始化神经网络
+    const model = tf.sequential();
+
+    // 3. 添加全连接层
+    // 3.1 隐藏层，核心
+    model.add(tf.layers.dense({
+        units: 4, // 神经元个数，比如此案例，4个象限
+        inputShape: [2], // 输入特征，坐标 x y，inputShape为2
+        // 激活函数 relu 让神经网络拥有非线性拟合能力，这是复杂神经网络拟合的核心
+        // 如果不加这个激活函数，结果只能为线性，不能得到非线性的结果
+        activation: 'relu'
+    }));
+    // 3.2 输出层
+    model.add(tf.layers.dense({
+        units: 1, // 神经元个数，最终输出的是一个概率值
+        // inputShape: [4], // 不需要设置，因为上层的 units 已经确定了
+        activation: 'sigmoid' // 选 sigmoid 输出 0-1 之间的值
+    }));
+    // 3.3 设置损失函数和优化器
+    model.compile({
+        loss: tf.losses.logLoss, // 损失函数，log损失，本质也是逻辑回归
+        optimizer: tf.train.adam(0.1) // 优化器，adam，初始化学习速率0.1
+    });
+
+    // 4. 准备训练数据
+    const inputs = tf.tensor(data.map(p => [p.x, p.y]));
+    const labels = tf.tensor(data.map(p => p.label));
+
+    // 5. 训练
+    await model.fit(inputs, labels, {
+        epochs: 10, // 10轮
+        callbacks: tfvis.show.fitCallbacks(
+            { name: '训练效果' },
+            ['loss']
+        )
+    });
+
+    window.predict = (form) => {
+        const pred = model.predict(tf.tensor([[form.x.value * 1, form.y.value * 1]]));
+        alert(`预测结果：${pred.dataSync()[0]}`);
+    };
+};