diff --git a/ch3_DL_basics/3.9_mlp-scratch.ipynb b/ch3_DL_basics/3.9_mlp-scratch.ipynb new file mode 100644 index 0000000..d3b0d49 --- /dev/null +++ b/ch3_DL_basics/3.9_mlp-scratch.ipynb @@ -0,0 +1,262 @@ +{ + "cells": [ + { + "cell_type": "markdown", + "metadata": {}, + "source": [ + "3.9 多层感知机的从零开始实现\n", + "我们已经从上一节里了解了多层感知机的原理。下面，我们一起来动手实现一个多层感知机。首先导入实现所需的包或模块。" + ] + }, + { + "cell_type": "code", + "execution_count": 41, + "metadata": { + "pycharm": { + "is_executing": false + } + }, + "outputs": [ + { + "name": "stdout", + "output_type": "stream", + "text": [ + "2.0.0\n" + ] + } + ], + "source": [ + "import tensorflow as tf\n", + "import numpy as np\n", + "import sys\n", + "print(tf.__version__)" + ] + }, + { + "cell_type": "markdown", + "metadata": {}, + "source": [ + "3.9.1 获取和读取数据\n", + "这里继续使用Fashion-MNIST数据集。我们将使用多层感知机对图像进行分类" + ] + }, + { + "cell_type": "code", + "execution_count": 42, + "metadata": { + "collapsed": true + }, + "outputs": [], + "source": [ + "from tensorflow.keras.datasets import fashion_mnist\n", + "(x_train, y_train), (x_test, y_test) = fashion_mnist.load_data()\n", + "batch_size = 256\n", + "x_train = tf.cast(x_train, tf.float32)\n", + "x_test = tf.cast(x_test, tf.float32)\n", + "x_train = x_train/255.0\n", + "x_test = x_test/255.0\n", + "train_iter = tf.data.Dataset.from_tensor_slices((x_train, y_train)).batch(batch_size)\n", + "test_iter = tf.data.Dataset.from_tensor_slices((x_test, y_test)).batch(batch_size)" + ] + }, + { + "cell_type": "markdown", + "metadata": {}, + "source": [ + "3.9.2 定义模型参数\n", + "我们在3.6节（softmax回归的从零开始实现）里已经介绍了，Fashion-MNIST数据集中图像形状为 28×28，类别数为10。本节中我们依然使用长度为 28×28=784 的向量表示每一张图像。因此，输入个数为784，输出个数为10。实验中，我们设超参数隐藏单元个数为256。" + ] + }, + { + "cell_type": "code", + "execution_count": 43, + "metadata": {}, + "outputs": [], + "source": [ + "num_inputs, num_outputs, num_hiddens = 784, 10, 256\n", + "\n", + "w1 = tf.Variable(tf.random.truncated_normal([num_inputs, num_hiddens], stddev=0.1))\n", + "b1 = tf.Variable(tf.random.truncated_normal([num_hiddens], stddev=0.1))\n", + "w2 = tf.Variable(tf.random.truncated_normal([num_hiddens, num_outputs], stddev=0.1))\n", + "b2=tf.Variable(tf.random.truncated_normal([num_outputs], stddev=0.1))\n" + ] + }, + { + "cell_type": "markdown", + "metadata": {}, + "source": [ + "3.9.3 定义激活函数\n", + "这里我们使用基础的max函数来实现ReLU，而非直接调用relu函数。" + ] + }, + { + "cell_type": "code", + "execution_count": 44, + "metadata": { + "collapsed": true + }, + "outputs": [], + "source": [ + "def relu(x):\n", + " return tf.math.maximum(x,0)" + ] + }, + { + "cell_type": "code", + "execution_count": 45, + "metadata": { + "collapsed": true + }, + "outputs": [], + "source": [ + "def net(x,w1,b1,w2,b2):\n", + " x = tf.reshape(x,shape=[-1,num_inputs])\n", + " h = relu(tf.matmul(x,w1) + b1 )\n", + " y = tf.math.softmax( tf.matmul(h,w2) + b2 )\n", + " return y" + ] + }, + { + "cell_type": "markdown", + "metadata": {}, + "source": [ + "3.9.5. 定义损失函数¶\n", + "为了得到更好的数值稳定性，我们直接使用Tensorflow提供的包括softmax运算和交叉熵损失计算的函数。" + ] + }, + { + "cell_type": "code", + "execution_count": 46, + "metadata": { + "collapsed": true + }, + "outputs": [], + "source": [ + "def loss(y_hat,y_true):\n", + " return tf.losses.sparse_categorical_crossentropy(y_true,y_hat)" + ] + }, + { + "cell_type": "markdown", + "metadata": {}, + "source": [ + "3.9.6. 训练模型" + ] + }, + { + "cell_type": "code", + "execution_count": 47, + "metadata": { + "collapsed": true + }, + "outputs": [], + "source": [ + "def acc(y_hat,y):\n", + " return np.mean((tf.argmax(y_hat,axis=1) == y))" + ] + }, + { + "cell_type": "code", + "execution_count": 48, + "metadata": { + "collapsed": true + }, + "outputs": [], + "source": [ + "num_epochs, lr = 5, 0.5" + ] + }, + { + "cell_type": "code", + "execution_count": 49, + "metadata": {}, + "outputs": [ + { + "name": "stdout", + "output_type": "stream", + "text": [ + "0 loss: 0.7799275\n", + "0 test_acc: 0.875\n", + "1 loss: 0.72887945\n", + "1 test_acc: 0.9375\n", + "2 loss: 0.72454\n", + "2 test_acc: 0.8125\n", + "3 loss: 0.5607478\n", + "3 test_acc: 0.875\n", + "4 loss: 0.5008962\n", + "4 test_acc: 0.9375\n" + ] + } + ], + "source": [ + "for epoch in range(num_epochs):\n", + " loss_all = 0\n", + " for x,y in train_iter:\n", + " with tf.GradientTape() as tape:\n", + " y_hat = net(x,w1,b1,w2,b2)\n", + " l = tf.reduce_mean(loss(y_hat,y))\n", + " loss_all += l.numpy()\n", + " grads = tape.gradient(l, [w1, b1, w2, b2])\n", + " w1.assign_sub(grads[0])\n", + " b1.assign_sub(grads[1])\n", + " w2.assign_sub(grads[2])\n", + " b2.assign_sub(grads[3])\n", + " print(epoch, 'loss:', l.numpy())\n", + " total_correct, total_number = 0, 0\n", + "\n", + " for x,y in test_iter:\n", + " with tf.GradientTape() as tape:\n", + " y_hat = net(x,w1,b1,w2,b2)\n", + " y=tf.cast(y,'int64')\n", + " correct=acc(y_hat,y)\n", + " print(epoch,\"test_acc:\", correct)" + ] + }, + { + "cell_type": "code", + "execution_count": null, + "metadata": {}, + "outputs": [], + "source": [] + }, + { + "cell_type": "code", + "execution_count": null, + "metadata": { + "collapsed": true + }, + "outputs": [], + "source": [] + }, + { + "cell_type": "code", + "execution_count": null, + "metadata": { + "collapsed": true + }, + "outputs": [], + "source": [] + } + ], + "metadata": { + "kernelspec": { + "display_name": "Python 3", + "language": "python", + "name": "python3" + }, + "language_info": { + "codemirror_mode": { + "name": "ipython", + "version": 3 + }, + "file_extension": ".py", + "mimetype": "text/x-python", + "name": "python", + "nbconvert_exporter": "python", + "pygments_lexer": "ipython3", + "version": "3.6.1" + } + }, + "nbformat": 4, + "nbformat_minor": 2 +}