{ "cells": [ { "cell_type": "markdown", "metadata": {}, "source": [ "# PaddleSlim Distillation知识蒸馏简介与实验\n", "\n", "一般情况下，模型参数量越多，结构越复杂，其性能越好，但参数也越冗余，运算量和资源消耗也越大。**知识蒸馏**就是一种将大模型学习到的有用信息（Dark Knowledge）压缩进更小更快的模型，而获得可以匹敌大模型结果的方法。\n", "\n", "在本文中性能强劲的大模型被称为teacher, 性能稍逊但体积较小的模型被称为student。示例包含以下步骤：\n", "\n", "1. 导入依赖\n", "2. 定义student_program和teacher_program\n", "3. 选择特征图\n", "4. 合并program (merge)并添加蒸馏loss\n", "5. 模型训练\n", "\n", "\n", "## 1. 导入依赖\n", "PaddleSlim依赖Paddle1.7版本，请确认已正确安装Paddle，然后按以下方式导入Paddle、PaddleSlim以及其他依赖:" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "import paddle\n", "import paddle.fluid as fluid\n", "import paddleslim as slim\n", "import sys\n", "sys.path.append(\"../\")\n", "import models" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "## 2. 定义student_program和teacher_program\n", "\n", "本教程在MNIST数据集上进行知识蒸馏的训练和验证，输入图片尺寸为`[1, 28, 28]`，输出类别数为10。\n", "选择`ResNet50`作为teacher对`MobileNet`结构的student进行蒸馏训练。" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "scrolled": true }, "outputs": [], "source": [ "model = models.__dict__['MobileNet']()\n", "student_program = fluid.Program()\n", "student_startup = fluid.Program()\n", "with fluid.program_guard(student_program, student_startup):\n", " image = fluid.data(\n", " name='image', shape=[None] + [1, 28, 28], dtype='float32')\n", " label = fluid.data(name='label', shape=[None, 1], dtype='int64')\n", " out = model.net(input=image, class_dim=10)\n", " cost = fluid.layers.cross_entropy(input=out, label=label)\n", " avg_cost = fluid.layers.mean(x=cost)\n", " acc_top1 = fluid.layers.accuracy(input=out, label=label, k=1)\n", " acc_top5 = fluid.layers.accuracy(input=out, label=label, k=5)" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "teacher_model = models.__dict__['ResNet50']()\n", "teacher_program = fluid.Program()\n", "teacher_startup = fluid.Program()\n", "with fluid.program_guard(teacher_program, teacher_startup):\n", " with fluid.unique_name.guard():\n", " image = fluid.data(\n", " name='image', shape=[None] + [1, 28, 28], dtype='float32')\n", " predict = teacher_model.net(image, class_dim=10)\n", "exe = fluid.Executor(fluid.CPUPlace())\n", "exe.run(teacher_startup)" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "## 3. 选择特征图\n", "我们可以用student_的list_vars方法来观察其中全部的Variables，从中选出一个或多个变量（Variable）来拟合teacher相应的变量。" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "# get all student variables\n", "student_vars = []\n", "for v in student_program.list_vars():\n", " student_vars.append((v.name, v.shape))\n", "#uncomment the following lines to observe student's variables for distillation\n", "#print(\"=\"*50+\"student_model_vars\"+\"=\"*50)\n", "#print(student_vars)\n", "\n", "# get all teacher variables\n", "teacher_vars = []\n", "for v in teacher_program.list_vars():\n", " teacher_vars.append((v.name, v.shape))\n", "#uncomment the following lines to observe teacher's variables for distillation\n", "#print(\"=\"*50+\"teacher_model_vars\"+\"=\"*50)\n", "#print(teacher_vars)" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "经过筛选我们可以看到，teacher_program中的'bn5c_branch2b.output.1.tmp_3'和student_program的'depthwise_conv2d_11.tmp_0'尺寸一致，可以组成蒸馏损失函数。" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "## 4. 合并program (merge)并添加蒸馏loss\n", "merge操作将student_program和teacher_program中的所有Variables和Op都将被添加到同一个Program中，同时为了避免两个program中有同名变量会引起命名冲突，merge也会为teacher_program中的Variables添加一个同一的命名前缀name_prefix，其默认值是'teacher_'\n", "\n", "为了确保teacher网络和student网络输入的数据是一样的，merge操作也会对两个program的输入数据层进行合并操作，所以需要指定一个数据层名称的映射关系data_name_map，key是teacher的输入数据名称，value是student的" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "data_name_map = {'image': 'image'}\n", "main = slim.dist.merge(teacher_program, student_program, data_name_map, fluid.CPUPlace())\n", "with fluid.program_guard(student_program, student_startup):\n", " l2_loss = slim.dist.l2_loss('teacher_bn5c_branch2b.output.1.tmp_3', 'depthwise_conv2d_11.tmp_0', student_program)\n", " loss = l2_loss + avg_cost\n", " opt = fluid.optimizer.Momentum(0.01, 0.9)\n", " opt.minimize(loss)\n", "exe.run(student_startup)" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "## 5. 模型训练\n", "\n", "为了快速执行该示例，我们选取简单的MNIST数据，Paddle框架的`paddle.dataset.mnist`包定义了MNIST数据的下载和读取。\n", "代码如下：" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "train_reader = paddle.fluid.io.batch(\n", " paddle.dataset.mnist.train(), batch_size=128, drop_last=True)\n", "train_feeder = fluid.DataFeeder(['image', 'label'], fluid.CPUPlace(), student_program)" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "for data in train_reader():\n", " acc1, acc5, loss_np = exe.run(student_program, feed=train_feeder.feed(data), fetch_list=[acc_top1.name, acc_top5.name, loss.name])\n", " print(\"Acc1: {:.6f}, Acc5: {:.6f}, Loss: {:.6f}\".format(acc1.mean(), acc5.mean(), loss_np.mean()))" ] } ], "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": "3.7.5" } }, "nbformat": 4, "nbformat_minor": 1 }