图像分类模型知识蒸馏-快速开始¶

该教程以图像分类模型MobileNetV1为例，说明如何快速使用PaddleSlim的知识蒸馏接口。该示例包含以下步骤：

导入依赖
定义student_program和teacher_program
选择特征图
合并program(merge)并添加蒸馏loss
模型训练

以下章节依次介绍每个步骤的内容。

1. 导入依赖¶

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和PaddleSlim:

import paddle
import paddle.fluid as fluid
import paddleslim as slim

2. 定义student_program和teacher_program¶

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。选择ResNet50作为teacher对MobileNet结构的student进行蒸馏训练。

model = models.__dict__['MobileNet']()
student_program = fluid.Program()
student_startup = fluid.Program()
with fluid.program_guard(student_program, student_startup):
    image = fluid.data(
        name='image', shape=[None] + [1, 28, 28], dtype='float32')
    label = fluid.data(name='label', shape=[None, 1], dtype='int64')
    out = model.net(input=image, class_dim=10)
    cost = fluid.layers.cross_entropy(input=out, label=label)
    avg_cost = fluid.layers.mean(x=cost)
    acc_top1 = fluid.layers.accuracy(input=out, label=label, k=1)
    acc_top5 = fluid.layers.accuracy(input=out, label=label, k=5)

teacher_model = models.__dict__['ResNet50']()
teacher_program = fluid.Program()
teacher_startup = fluid.Program()
with fluid.program_guard(teacher_program, teacher_startup):
    with fluid.unique_name.guard():
        image = fluid.data(
            name='image', shape=[None] + [1, 28, 28], dtype='float32')
        predict = teacher_model.net(image, class_dim=10)
exe = fluid.Executor(fluid.CPUPlace())
exe.run(teacher_startup)

3. 选择特征图¶

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。

# get all student variables
student_vars = []
for v in student_program.list_vars():
    student_vars.append((v.name, v.shape))
#uncomment the following lines to observe student's variables for distillation
#print("="*50+"student_model_vars"+"="*50)
#print(student_vars)

# get all teacher variables
teacher_vars = []
for v in teacher_program.list_vars():
    teacher_vars.append((v.name, v.shape))
#uncomment the following lines to observe teacher's variables for distillation
#print("="*50+"teacher_model_vars"+"="*50)
#print(teacher_vars)

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。

4. 合并program (merge)并添加蒸馏loss¶

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nameprefix，其默认值是’teacher‘

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

data_name_map = {'image': 'image'}
main = slim.dist.merge(teacher_program, student_program, data_name_map, fluid.CPUPlace())
with fluid.program_guard(student_program, student_startup):
    l2_loss = slim.dist.l2_loss('teacher_bn5c_branch2b.output.1.tmp_3', 'depthwise_conv2d_11.tmp_0', student_program)
    loss = l2_loss + avg_cost
    opt = fluid.optimizer.Momentum(0.01, 0.9)
    opt.minimize(loss)
exe.run(student_startup)

5. 模型训练¶

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。代码如下：

train_reader = paddle.batch(
    paddle.dataset.mnist.train(), batch_size=128, drop_last=True)
train_feeder = fluid.DataFeeder(['image', 'label'], fluid.CPUPlace(), student_program)

for data in train_reader():
    acc1, acc5, loss_np = exe.run(student_program, feed=train_feeder.feed(data), fetch_list=[acc_top1.name, acc_top5.name, loss.name])
    print("Acc1: {:.6f}, Acc5: {:.6f}, Loss: {:.6f}".format(acc1.mean(), acc5.mean(), loss_np.mean()))