mnist_lenet_classification.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# MNIST数据集使用LeNet进行图像分类\n",
    "本示例教程演示如何在MNIST数据集上用LeNet进行图像分类。\n",
    "手写数字的MNIST数据集，包含60,000个用于训练的示例和10,000个用于测试的示例。这些数字已经过尺寸标准化并位于图像中心，图像是固定大小(28x28像素)，其值为0到1。该数据集的官方地址为：http://yann.lecun.com/exdb/mnist/"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 环境\n",
    "本教程基于paddle-2.0-beta编写，如果您的环境不是本版本，请先安装paddle-2.0-beta版本。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "2.0.0-beta0\n"
     ]
    }
   ],
   "source": [
    "import paddle\n",
    "print(paddle.__version__)\n",
    "paddle.disable_static()\n",
    "# 开启动态图"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 加载数据集\n",
    "我们使用飞桨自带的paddle.dataset完成mnist数据集的加载。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "download training data and load training data\n",
      "load finished\n"
     ]
    }
   ],
   "source": [
    "print('download training data and load training data')\n",
    "train_dataset = paddle.vision.datasets.MNIST(mode='train')\n",
    "test_dataset = paddle.vision.datasets.MNIST(mode='test')\n",
    "print('load finished')"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "取训练集中的一条数据看一下。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "train_data0 label is: [5]\n"
     ]
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 144x144 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "import numpy as np\n",
    "import matplotlib.pyplot as plt\n",
    "train_data0, train_label_0 = train_dataset[0][0],train_dataset[0][1]\n",
    "train_data0 = train_data0.reshape([28,28])\n",
    "plt.figure(figsize=(2,2))\n",
    "plt.imshow(train_data0, cmap=plt.cm.binary)\n",
    "print('train_data0 label is: ' + str(train_label_0))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 组网\n",
    "用paddle.nn下的API，如`Conv2d`、`MaxPool2d`、`Linear`完成LeNet的构建。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [],
   "source": [
    "import paddle\n",
    "import paddle.nn.functional as F\n",
    "class LeNet(paddle.nn.Layer):\n",
    "    def __init__(self):\n",
    "        super(LeNet, self).__init__()\n",
    "        self.conv1 = paddle.nn.Conv2d(in_channels=1, out_channels=6, kernel_size=5, stride=1, padding=2)\n",
    "        self.max_pool1 = paddle.nn.MaxPool2d(kernel_size=2,  stride=2)\n",
    "        self.conv2 = paddle.nn.Conv2d(in_channels=6, out_channels=16, kernel_size=5, stride=1)\n",
    "        self.max_pool2 = paddle.nn.MaxPool2d(kernel_size=2, stride=2)\n",
    "        self.linear1 = paddle.nn.Linear(in_features=16*5*5, out_features=120)\n",
    "        self.linear2 = paddle.nn.Linear(in_features=120, out_features=84)\n",
    "        self.linear3 = paddle.nn.Linear(in_features=84, out_features=10)\n",
    "\n",
    "    def forward(self, x):\n",
    "        x = self.conv1(x)\n",
    "        x = F.relu(x)\n",
    "        x = self.max_pool1(x)\n",
    "        x = F.relu(x)\n",
    "        x = self.conv2(x)\n",
    "        x = self.max_pool2(x)\n",
    "        x = paddle.flatten(x, start_axis=1,stop_axis=-1)\n",
    "        x = self.linear1(x)\n",
    "        x = F.relu(x)\n",
    "        x = self.linear2(x)\n",
    "        x = F.relu(x)\n",
    "        x = self.linear3(x)\n",
    "        x = F.softmax(x)\n",
    "        return x"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 训练方式一\n",
    "组网后，开始对模型进行训练，先构建`train_loader`，加载训练数据，然后定义`train`函数，设置好损失函数后，按batch加载数据，完成模型的训练。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "epoch: 0, batch_id: 0, loss is: [2.3037894], acc is: [0.140625]\n",
      "epoch: 0, batch_id: 100, loss is: [1.6175328], acc is: [0.9375]\n",
      "epoch: 0, batch_id: 200, loss is: [1.5388051], acc is: [0.96875]\n",
      "epoch: 0, batch_id: 300, loss is: [1.5251061], acc is: [0.96875]\n",
      "epoch: 0, batch_id: 400, loss is: [1.4678856], acc is: [1.]\n",
      "epoch: 0, batch_id: 500, loss is: [1.4944503], acc is: [0.984375]\n",
      "epoch: 0, batch_id: 600, loss is: [1.5365536], acc is: [0.96875]\n",
      "epoch: 0, batch_id: 700, loss is: [1.4885054], acc is: [0.984375]\n",
      "epoch: 0, batch_id: 800, loss is: [1.4872254], acc is: [0.984375]\n",
      "epoch: 0, batch_id: 900, loss is: [1.4884174], acc is: [0.984375]\n",
      "epoch: 1, batch_id: 0, loss is: [1.4776722], acc is: [1.]\n",
      "epoch: 1, batch_id: 100, loss is: [1.4751343], acc is: [1.]\n",
      "epoch: 1, batch_id: 200, loss is: [1.4772581], acc is: [1.]\n",
      "epoch: 1, batch_id: 300, loss is: [1.4918218], acc is: [0.984375]\n",
      "epoch: 1, batch_id: 400, loss is: [1.5038397], acc is: [0.96875]\n",
      "epoch: 1, batch_id: 500, loss is: [1.5088196], acc is: [0.96875]\n",
      "epoch: 1, batch_id: 600, loss is: [1.4961376], acc is: [0.984375]\n",
      "epoch: 1, batch_id: 700, loss is: [1.4755756], acc is: [1.]\n",
      "epoch: 1, batch_id: 800, loss is: [1.4921497], acc is: [0.984375]\n",
      "epoch: 1, batch_id: 900, loss is: [1.4944404], acc is: [1.]\n"
     ]
    }
   ],
   "source": [
    "import paddle\n",
    "train_loader = paddle.io.DataLoader(train_dataset, places=paddle.CPUPlace(), batch_size=64, shuffle=True)\n",
    "# 加载训练集 batch_size 设为 64\n",
    "def train(model):\n",
    "    model.train()\n",
    "    epochs = 2\n",
    "    optim = paddle.optimizer.Adam(learning_rate=0.001, parameters=model.parameters())\n",
    "    # 用Adam作为优化函数\n",
    "    for epoch in range(epochs):\n",
    "        for batch_id, data in enumerate(train_loader()):\n",
    "            x_data = data[0]\n",
    "            y_data = data[1]\n",
    "            predicts = model(x_data)\n",
    "            loss = paddle.nn.functional.cross_entropy(predicts, y_data)\n",
    "            # 计算损失\n",
    "            acc = paddle.metric.accuracy(predicts, y_data, k=2)\n",
    "            avg_loss = paddle.mean(loss)\n",
    "            avg_acc = paddle.mean(acc)\n",
    "            avg_loss.backward()\n",
    "            if batch_id % 100 == 0:\n",
    "                print(\"epoch: {}, batch_id: {}, loss is: {}, acc is: {}\".format(epoch, batch_id, avg_loss.numpy(), avg_acc.numpy()))\n",
    "            optim.step()\n",
    "            optim.clear_grad()\n",
    "model = LeNet()\n",
    "train(model)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 对模型进行验证\n",
    "训练完成后，需要验证模型的效果，此时，加载测试数据集，然后用训练好的模对测试集进行预测，计算损失与精度。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "batch_id: 0, loss is: [1.4915928], acc is: [1.]\n",
      "batch_id: 20, loss is: [1.4818308], acc is: [1.]\n",
      "batch_id: 40, loss is: [1.5006062], acc is: [0.984375]\n",
      "batch_id: 60, loss is: [1.521233], acc is: [1.]\n",
      "batch_id: 80, loss is: [1.4772738], acc is: [1.]\n",
      "batch_id: 100, loss is: [1.4755945], acc is: [1.]\n",
      "batch_id: 120, loss is: [1.4746133], acc is: [1.]\n",
      "batch_id: 140, loss is: [1.4786345], acc is: [1.]\n"
     ]
    }
   ],
   "source": [
    "import paddle\n",
    "test_loader = paddle.io.DataLoader(test_dataset, places=paddle.CPUPlace(), batch_size=64)\n",
    "# 加载测试数据集\n",
    "def test(model):\n",
    "    model.eval()\n",
    "    batch_size = 64\n",
    "    for batch_id, data in enumerate(test_loader()):\n",
    "        x_data = data[0]\n",
    "        y_data = data[1]\n",
    "        predicts = model(x_data)\n",
    "        # 获取预测结果\n",
    "        loss = paddle.nn.functional.cross_entropy(predicts, y_data)\n",
    "        acc = paddle.metric.accuracy(predicts, y_data, k=2)\n",
    "        avg_loss = paddle.mean(loss)\n",
    "        avg_acc = paddle.mean(acc)\n",
    "        avg_loss.backward()\n",
    "        if batch_id % 20 == 0:\n",
    "            print(\"batch_id: {}, loss is: {}, acc is: {}\".format(batch_id, avg_loss.numpy(), avg_acc.numpy()))\n",
    "test(model)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 训练方式一结束\n",
    "以上就是训练方式一，通过这种方式，可以清楚的看到训练和测试中的每一步过程。但是，这种方式句法比较复杂。因此，我们提供了训练方式二，能够更加快速、高效的完成模型的训练与测试。"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 3.训练方式二\n",
    "通过paddle提供的`Model` 构建实例，使用封装好的训练与测试接口，快速完成模型训练与测试。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "metadata": {},
   "outputs": [],
   "source": [
    "import paddle\n",
    "from paddle.static import InputSpec\n",
    "from paddle.metric import Accuracy\n",
    "inputs = InputSpec([None, 784], 'float32', 'x')\n",
    "labels = InputSpec([None, 10], 'float32', 'x')\n",
    "model = paddle.Model(LeNet(), inputs, labels)\n",
    "optim = paddle.optimizer.Adam(learning_rate=0.001, parameters=model.parameters())\n",
    "\n",
    "model.prepare(\n",
    "    optim,\n",
    "    paddle.nn.loss.CrossEntropyLoss(),\n",
    "    Accuracy(topk=(1, 2))\n",
    "    )"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 使用model.fit来训练模型"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 16,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Epoch 1/2\n",
      "step 200/938 - loss: 1.5219 - acc_top1: 0.9829 - acc_top2: 0.9965 - 14ms/step\n",
      "step 400/938 - loss: 1.4765 - acc_top1: 0.9825 - acc_top2: 0.9958 - 13ms/step\n",
      "step 600/938 - loss: 1.4624 - acc_top1: 0.9823 - acc_top2: 0.9953 - 13ms/step\n",
      "step 800/938 - loss: 1.4768 - acc_top1: 0.9829 - acc_top2: 0.9955 - 13ms/step\n",
      "step 938/938 - loss: 1.4612 - acc_top1: 0.9836 - acc_top2: 0.9956 - 13ms/step\n",
      "Epoch 2/2\n",
      "step 200/938 - loss: 1.4705 - acc_top1: 0.9834 - acc_top2: 0.9959 - 13ms/step\n",
      "step 400/938 - loss: 1.4620 - acc_top1: 0.9833 - acc_top2: 0.9960 - 13ms/step\n",
      "step 600/938 - loss: 1.4613 - acc_top1: 0.9830 - acc_top2: 0.9960 - 13ms/step\n",
      "step 800/938 - loss: 1.4763 - acc_top1: 0.9831 - acc_top2: 0.9960 - 13ms/step\n",
      "step 938/938 - loss: 1.4924 - acc_top1: 0.9834 - acc_top2: 0.9959 - 13ms/step\n"
     ]
    }
   ],
   "source": [
    "model.fit(train_dataset,\n",
    "        epochs=2,\n",
    "        batch_size=64,\n",
    "        log_freq=200\n",
    "        )"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 使用model.evaluate来预测模型"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 17,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Eval begin...\n",
      "step  20/157 - loss: 1.5246 - acc_top1: 0.9773 - acc_top2: 0.9969 - 6ms/step\n",
      "step  40/157 - loss: 1.4622 - acc_top1: 0.9758 - acc_top2: 0.9961 - 6ms/step\n",
      "step  60/157 - loss: 1.5241 - acc_top1: 0.9763 - acc_top2: 0.9951 - 6ms/step\n",
      "step  80/157 - loss: 1.4612 - acc_top1: 0.9787 - acc_top2: 0.9959 - 6ms/step\n",
      "step 100/157 - loss: 1.4612 - acc_top1: 0.9823 - acc_top2: 0.9967 - 5ms/step\n",
      "step 120/157 - loss: 1.4612 - acc_top1: 0.9835 - acc_top2: 0.9966 - 5ms/step\n",
      "step 140/157 - loss: 1.4612 - acc_top1: 0.9844 - acc_top2: 0.9969 - 5ms/step\n",
      "step 157/157 - loss: 1.4612 - acc_top1: 0.9838 - acc_top2: 0.9966 - 5ms/step\n",
      "Eval samples: 10000\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "{'loss': [1.4611504], 'acc_top1': 0.9838, 'acc_top2': 0.9966}"
      ]
     },
     "execution_count": 17,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "model.evaluate(test_dataset, log_freq=20, batch_size=64)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 训练方式二结束\n",
    "以上就是训练方式二，可以快速、高效的完成网络模型训练与预测。"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 总结\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "以上就是用LeNet对手写数字数据及MNIST进行分类。本示例提供了两种训练模型的方式，一种可以快速完成模型的组建与预测，非常适合新手用户上手。另一种则需要多个步骤来完成模型的训练，适合进阶用户使用。"
   ]
  }
 ],
 "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": "3.7.3"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 4
}