Show Me the Code之MXNet网络模型（三）

LeNet 是1986年发表的模型。

网络模型

网络层	网络类型	输出数据类型	训练参数	前一层
Input		1x28x28
1-Conv2D	卷积层	6x28x28	6x(1x5x5+1)=156	Input
1-Activation	激活层	6x28x28		1-Conv2D
1-MaxPool2D	池化层	6x14x14		1-Activation
2-Conv2D	卷积层	16x10x10	16x(6x5x5+1)=2416	1-MaxPool2D
2-Activation	激活层	16x10x10		2-Conv2D
2-MaxPool2D	池化层	16x5x5		2-Activation
3-Dense	全连接	120	120x(16x5x5+1)=48120	2-MaxPool2D
3-Activation	激活层	120		3-Dense
4-Dense	全连接	84	84x(120+1)=10164	3-Activation
4-Activation	激活层	84		4-Dense
5-Dense	全连接	10	10x(84+1)=850	4-Activation
统计			61706

由于全连接层占用了绝大部分的训练参数（95.83%），而卷积层比全连接层更加消耗计算力。因此，有人说 “ 全连接层负责参数部分，卷积层负责计算部分 ”。

Show me the code!

说明一下，这里的实现和原本LeNet并不一样。原版的激活函数选的是Sigmoid，而池化为平均值池化。

MXNet 1.5.1

# -*- coding: utf-8 -*-

import logging
import struct
import gzip
import numpy as np
import mxnet as mx

logging.getLogger().setLevel(logging.DEBUG)

# 批大小
batch_size = 32
# 学习轮数
train_epoch = 20
# 样本路径
resource_path = "fashion-mnist/"

'''
************************************************************
*                          数据准备
************************************************************
'''
# 定义读取数据的函数
def read_data( label_url, image_url ):
    with gzip.open( label_url ) as flbl:
        # 读入标签文件头
        magic, num = struct.unpack(">II", flbl.read(8))
        # 读入标签内容
        label = np.frombuffer( flbl.read(), dtype = np.uint8 )
    with gzip.open( image_url, 'rb' ) as fimg:
        # 读入图像文件头，rows和cols都是28
        magic, num, rows, cols = struct.unpack( ">IIII", fimg.read(16) )
        # 读入图像内容
        image = np.frombuffer( fimg.read(), dtype = np.uint8 )
        # 设置为正确的数组格式
        image = image.reshape( len(label), 1, rows, cols )
        # 归一化到 0~1
        image = image.astype( np.float32 ) / 255.0
    return (label, image)

# 读入数据
# 注意路径的问题
( train_lbl, train_img ) = read_data(
        resource_path + 'train-labels-idx1-ubyte.gz',
        resource_path + 'train-images-idx3-ubyte.gz'
        )
( eval_lbl , eval_img  ) = read_data(
        resource_path + 't10k-labels-idx1-ubyte.gz',
        resource_path + 't10k-images-idx3-ubyte.gz'
        )

# 迭代器
train_iter = mx.io.NDArrayIter( train_img, train_lbl, batch_size, shuffle=True )
eval_iter  = mx.io.NDArrayIter( eval_img , eval_lbl , batch_size )  # 验证集可以省略shuffle

'''
************************************************************
*                       定义神经网络模型
************************************************************
'''
# 输入层
net = mx.sym.var( 'data' )
# 第1层隐藏层
net = mx.sym.Convolution   (data=net, name='layer1_conv', num_filter=6, kernel=(5,5), pad=(2,2))
net = mx.sym.Activation    (data=net, name='layer1_act' , act_type='relu')
net = mx.sym.Pooling       (data=net, name='layer1_pool', kernel=(2,2), stride=(2,2), pool_type='max')
# 第2层隐藏层
net = mx.sym.Convolution   (data=net, name='layer2_conv', num_filter=16, kernel=(5,5))
net = mx.sym.Activation    (data=net, name='layer2_act' , act_type='relu')
net = mx.sym.Pooling       (data=net, name='layer2_pool', kernel=(2,2), stride=(2,2), pool_type='max')
# 第3层隐藏层
net = mx.sym.Flatten       (data=net, name='flatten')       # 将图像摊平
net = mx.sym.FullyConnected(data=net, name='layer3_fc'  , num_hidden=120)
net = mx.sym.Activation    (data=net, name='layer3_act' , act_type='relu')
# 第4层隐藏层
net = mx.sym.FullyConnected(data=net, name='layer4_fc'  , num_hidden=84)
net = mx.sym.Activation    (data=net, name='layer4_act' , act_type='relu')
# 输出层
net = mx.sym.FullyConnected(data=net, name='layer5_fc'  , num_hidden=10)
net = mx.sym.SoftmaxOutput (data=net, name='softmax')       # Softmax也是激活层

# 网络模型
ctx = mx.gpu() if mx.test_utils.list_gpus() else mx.cpu()   # 有GPU就用GPU
module = mx.mod.Module(symbol=net, context=ctx)

# 网络模型可视化
# shape = {'data':(batch_size, 1, 28, 28)}
# mx.viz.print_summary(symbol=net, shape=shape)
# mx.viz.plot_network(symbol=net, shape=shape).view()

'''
************************************************************
*                        训练神经网络
************************************************************
'''
# 定义评价标准(Evaluation Metric)
eval_metrics = mx.metric.CompositeEvalMetric()
eval_metrics.add( mx.metric.Accuracy() );       # 准确率
eval_metrics.add( mx.metric.CrossEntropy() );   # 交叉熵

print("start train...")

module.fit(
        train_data  = train_iter,               # 训练集
        eval_data   = eval_iter,                # 验证集
        eval_metric = eval_metrics,             # 评价标准
        num_epoch   = train_epoch,              # 训练轮数
        initializer = mx.initializer.Xavier(),  # Xavier初始化策略
        optimizer   = 'sgd',                    # 随机梯度下降算法
        optimizer_params = {
            'learning_rate': 0.01,              # 学习率
            'momentum': 0.9                     # 惯性动量
            }
    )

INFO:root:Epoch[0] Train-accuracy=0.791900
INFO:root:Epoch[0] Train-cross-entropy=0.560617
INFO:root:Epoch[0] Time cost=5.205
INFO:root:Epoch[0] Validation-accuracy=0.860024
INFO:root:Epoch[0] Validation-cross-entropy=0.387932
INFO:root:Epoch[1] Train-accuracy=0.871650
INFO:root:Epoch[1] Train-cross-entropy=0.351662
INFO:root:Epoch[1] Time cost=5.167
INFO:root:Epoch[1] Validation-accuracy=0.883786
INFO:root:Epoch[1] Validation-cross-entropy=0.316452
…
INFO:root:Epoch[19] Train-accuracy=0.942633
INFO:root:Epoch[19] Train-cross-entropy=0.151282
INFO:root:Epoch[19] Time cost=5.186
INFO:root:Epoch[19] Validation-accuracy=0.907548
INFO:root:Epoch[19] Validation-cross-entropy=0.307947

MXNet 1.9.1

上面的代码，到了MXNet 1.6.0版本就用不了了。
下面是新的实现。

import time
import struct
import gzip
import numpy as np
import matplotlib.pyplot as plt
import mxnet as mx

设置批大小和CPU

batch_size = 32
device = mx.cpu(0)

定义功能函数，读取图片

def read_data( label_url, image_url ):
    with gzip.open( label_url ) as flbl:
        # 读入标签文件头
        magic, num = struct.unpack( ">II", flbl.read(8) )
        label = np.frombuffer( flbl.read(), dtype = np.uint8 )
    with gzip.open( image_url, 'rb' ) as fimg:
        # 读入图像文件头，rows和cols都是28
        magic, num, rows, cols = struct.unpack( ">IIII", fimg.read(16) )
        # 读入图像内容
        image = np.frombuffer( fimg.read(), dtype = np.uint8 )
        # 设置为正确的数组格式
        image = image.reshape( num, 1, rows, cols )
        # 归一化到 [-1,1]
        image = image.astype( np.float32 ) / 255.0
    return (label, image)

读取数据，打印维度

( train_lbl, train_img ) = read_data(
    'fashion-mnist/train-labels-idx1-ubyte.gz',
    'fashion-mnist/train-images-idx3-ubyte.gz'
)
( eval_lbl , eval_img  ) = read_data(
    'fashion-mnist/t10k-labels-idx1-ubyte.gz',
    'fashion-mnist/t10k-images-idx3-ubyte.gz'
)
print("train:", type(train_img), train_img.shape, train_img.dtype)
print("eval: ", type(eval_img),  eval_img.shape,  eval_img.dtype )

train: <class ‘numpy.ndarray’> (60000, 1, 28, 28) float32
eval: <class ‘numpy.ndarray’> (10000, 1, 28, 28) float32

查看数据图片

texts = (
    't-shirt', 'trouser', 'pullover', 'dress', 'coat',
    'sandal',  'shirt',   'sneaker',  'bag',   'ankle boot'
)
idxs = (0, 1, 2, 3, 4, 5, 7, 9, 14, 21)

for i in range(10):
    plt.subplot(2, 5, i + 1)
    idx = idxs[i]
    plt.xticks([])
    plt.yticks([])
    plt.title(texts[train_lbl[idx]])
    img = train_img[idx][0].astype( np.float32 )
    plt.imshow(img, interpolation='none', cmap='Blues')

plt.show()

创建训练迭代器。这一步设置了批大小和训练集的随机排序。

train_data = mx.gluon.data.DataLoader(
    mx.gluon.data.ArrayDataset(train_img, train_lbl),
    batch_size=batch_size,
    shuffle=True
)
eval_data = mx.gluon.data.DataLoader(
    mx.gluon.data.ArrayDataset(eval_img, eval_lbl),
    batch_size=batch_size,
    shuffle=False
)

定义一个评价类。用MXNet自带的也行，但是我想封装show()

class UserMetrics(mx.metric.CompositeEvalMetric):
    def __init__(self, name='user', output_names=None, label_names=None):
        # 初始化PyTorch父类
        super().__init__(name=name, output_names=output_names, label_names=label_names)
        super().add( mx.metric.Accuracy() )
        super().add( mx.metric.CrossEntropy() )
    def reset(self):
        super().reset()
        self.tic = time.time()
    def show(self, epoch=0, tag='[ ]'):
        cost = time.time() - self.tic
        name, val = super().get()
        print("Epoch %2d: %s cost:%.1fs %s:%.3f, %s:%.3f"
              % ( epoch, tag, cost, name[0], val[0], name[1], val[1] ) 
        )

定义网络

# 网络类型
net = mx.gluon.nn.HybridSequential()

# 中间层
net.add(
    # 第一层
    mx.gluon.nn.Conv2D( channels=6,  kernel_size=(5,5), strides=(1,1), padding=(2,2) ),
    mx.gluon.nn.Activation( 'relu' ),
    mx.gluon.nn.MaxPool2D( pool_size=(2,2) ),
    # 第二层
    mx.gluon.nn.Conv2D( channels=16, kernel_size=(5,5), strides=(1,1), padding=(0,0) ),
    mx.gluon.nn.Activation( 'relu' ),
    mx.gluon.nn.MaxPool2D( pool_size=(2,2) ),
    # 第三层
    mx.gluon.nn.Dense( 120 ),
    mx.gluon.nn.Activation( 'relu' ),
    # 第四层
    mx.gluon.nn.Dense( 84 ),
    mx.gluon.nn.Activation( 'relu' )
)

# 输出层
net.output = mx.gluon.nn.Dense( 10 )

# 初始化
net.initialize( init=mx.init.Xavier(), ctx=device )

# 展示网络
net.summary(mx.ndarray.zeros(shape=(1, 1, 28, 28), dtype=np.float32, ctx=device))

# 符号式加速
net.hybridize()

损失函数，训练器，评价器

# 损失函数
loss_function = mx.gluon.loss.SoftmaxCrossEntropyLoss()
# 求解器
optimizer = mx.optimizer.SGD( learning_rate=0.01, momentum=0.0, multi_precision=False )
# 训练器
trainer = mx.gluon.Trainer( params=net.collect_params(), optimizer=optimizer )
# 评价器
metrics = UserMetrics()

开始训练 20 epoch

for epoch in range(20):
    # train
    metrics.reset()
    for datas, labels in train_data:
        
        actual_batch_size = datas.shape[0]
        
        # split batch and load into corresponding devices
        datas  = mx.gluon.utils.split_and_load( datas,  [device] )
        labels = mx.gluon.utils.split_and_load( labels, [device] )

        # The forward pass and the loss computation
        with mx.autograd.record():
            outputs = [ net(data) for data in datas ]
            losses = [ loss_function(output, label) for output, label in zip(outputs, labels) ]
            
        # compute gradients
        for loss in losses:
            loss.backward()
        
        # update parameters
        trainer.step( batch_size=actual_batch_size )
        
        # update metric
        for output, label in zip(outputs, labels):
            metrics.update( preds=mx.ndarray.softmax(output, axis=1), labels=label )
            
    metrics.show( epoch=epoch, tag='[ train ]' )
    
    # eval
    metrics.reset()
    for datas, labels in eval_data:
        # split batch and load into corresponding devices
        datas  = mx.gluon.utils.split_and_load(datas,  [device])
        labels = mx.gluon.utils.split_and_load(labels, [device])
        # The forward pass
        outputs = [ net(data) for data in datas ]
        # update metric
        for output, label in zip(outputs, labels):
            metrics.update( preds=mx.ndarray.softmax(output, axis=1), labels=label )
    metrics.show( epoch=epoch, tag='[  eval ]' )

Epoch 0: [ train ] cost:8.0s accuracy:0.693, cross-entropy:0.847
Epoch 0: [ eval ] cost:0.6s accuracy:0.784, cross-entropy:0.588
Epoch 1: [ train ] cost:7.9s accuracy:0.805, cross-entropy:0.531
Epoch 1: [ eval ] cost:0.6s accuracy:0.837, cross-entropy:0.453
Epoch 2: [ train ] cost:7.9s accuracy:0.835, cross-entropy:0.452
Epoch 2: [ eval ] cost:0.6s accuracy:0.851, cross-entropy:0.420
…
Epoch 18: [ train ] cost:8.0s accuracy:0.908, cross-entropy:0.250
Epoch 18: [ eval ] cost:0.6s accuracy:0.896, cross-entropy:0.290
Epoch 19: [ train ] cost:8.0s accuracy:0.910, cross-entropy:0.244
Epoch 19: [ eval ] cost:0.6s accuracy:0.899, cross-entropy:0.273