【TensorFlow2 之014】在 TF 2.0 中实现 LeNet-5

一、说明

在这篇文章中，我们将展示如何在 TensorFlow 中实现像 \(LeNet-5\) 这样的基础卷积神经网络。LeNet-5 架构由 Yann LeCun 于 1998 年发明，是第一个卷积神经网络。

数据黑客变种rs 深度学习机器学习 TensorFlow 2020 年 2 月 29 日 | 0

1.1 教程概述：

理论重述
在 TensorFlow 中的实现

1. 理论重述

\(LeNet-5 \) 的目标是识别手写数字。因此，它作为输入 \(32\times32\times1 \) 图像。它是灰度图像，因此通道数为 \(1 \)。下面我们可以看到该网络的架构。

LeNet5架构

LeNet-5架构

I. 最初的 MNIST 现在被认为太简单了。在过去的二十年里，许多研究人员针对原始 MNIST 提出了成功的解决方案。您可以在此处查看直接比较结果。

由于该网络主要是为 MNIST 数据集设计的，因此它的性能明显更好。通过微小的改变，它就可以在 Fashion MNIST 数据集上达到这种准确性。然而，在这篇文章中，我们将坚持网络的原始架构。

关于 LeNet-5 架构，您可以在此处阅读详细的理论文章。

让我们总结一下 LeNet-5 架构的各层。

图层类型	特征图	尺寸	内核大小	跨步	激活
图像	1	32×32	–	–	–
卷积	6	28×28	5×5	1	正值
平均池化	6	14×14	2×2	2	–
卷积	16	10×10	5×5	1	正值
平均池化	16	5×5	2×2	2	–
完全连接	–	120	–	–	正值
完全连接	–	84	–	–	正值
完全连接	–	10	–	–	软最大

二、TensorFlow中的实现

交互式 Colab 笔记本可在以下链接找到
在 Google Colab 中运行

为了练习，您可以尝试通过将Fashion_mnist替换为mnist、cifar10或其他来更改数据集。

让我们从导入所有必需的库开始。导入后，我们可以使用导入的模块来加载数据。load_data () 函数将自动下载数据并将其拆分为训练集和测试集。

import datetime
import numpy as np
import tensorflow as tf
import matplotlib.pyplot as plt

from tensorflow.keras import Model
from tensorflow.keras.models import Sequential
from tensorflow.keras.losses import categorical_crossentropy
from tensorflow.keras.layers import Dense, Flatten, Conv2D, AveragePooling2D

from tensorflow.keras import datasets
from tensorflow.keras.utils import to_categorical

#from __future__ import absolute_import, division, print_function, unicode_literals

# The data, split between train and test sets:
(x_train, y_train), (x_test, y_test) = datasets.fashion_mnist.load_data()

Downloading data from https://storage.googleapis.com/tensorflow/tf-keras-datasets/train-labels-idx1-ubyte.gz
32768/29515 [=================================] - 0s 2us/step
Downloading data from https://storage.googleapis.com/tensorflow/tf-keras-datasets/train-images-idx3-ubyte.gz
26427392/26421880 [==============================] - 7s 0us/step
Downloading data from https://storage.googleapis.com/tensorflow/tf-keras-datasets/t10k-labels-idx1-ubyte.gz
8192/5148 [===============================================] - 0s 0us/step
Downloading data from https://storage.googleapis.com/tensorflow/tf-keras-datasets/t10k-images-idx3-ubyte.gz
4423680/4422102 [==============================] - 1s 0us/step

我们可以检查新数据的形状，发现我们的图像是 28×28 像素，因此我们需要添加一个新轴，它将代表多个通道。此外，对标签进行 one-hot 编码和对输入图像进行归一化也很重要。

print('x_train shape:', x_train.shape)
print(x_train.shape[0], 'train samples')
print(x_test.shape[0], 'test samples')
print(x_train[0].shape, 'image shape')

x_train shape: (60000, 28, 28)
60000 train samples
10000 test samples
(28, 28) image shape

# Add a new axis
x_train = x_train[:, :, :, np.newaxis]
x_test = x_test[:, :, :, np.newaxis]

print('x_train shape:', x_train.shape)
print(x_train.shape[0], 'train samples')
print(x_test.shape[0], 'test samples')
print(x_train[0].shape, 'image shape')

x_train shape: (60000, 28, 28, 1)
60000 train samples
10000 test samples
(28, 28, 1) image shape

# Convert class vectors to binary class matrices.

num_classes = 10
y_train = to_categorical(y_train, num_classes)
y_test = to_categorical(y_test, num_classes)

# Data normalization
x_train = x_train.astype('float32')
x_test = x_test.astype('float32')
x_train /= 255
x_test /= 255

现在，是时候开始使用 TensorFlow 2.0 来构建我们的卷积神经网络了。最简单的方法是使用 Sequential API。我们将其包装在一个名为LeNet的类中。输入是图像，输出是类概率向量。

在 tf.keras 中，顺序模型表示层的线性堆栈，在本例中它遵循网络架构。

# LeNet-5 model
class LeNet(Sequential):
    def __init__(self, input_shape, nb_classes):
        super().__init__()

        self.add(Conv2D(6, kernel_size=(5, 5), strides=(1, 1), activation='tanh', input_shape=input_shape, padding="same"))
        self.add(AveragePooling2D(pool_size=(2, 2), strides=(2, 2), padding='valid'))
        self.add(Conv2D(16, kernel_size=(5, 5), strides=(1, 1), activation='tanh', padding='valid'))
        self.add(AveragePooling2D(pool_size=(2, 2), strides=(2, 2), padding='valid'))
        self.add(Flatten())
        self.add(Dense(120, activation='tanh'))
        self.add(Dense(84, activation='tanh'))
        self.add(Dense(nb_classes, activation='softmax'))

        self.compile(optimizer='adam',
                    loss=categorical_crossentropy,
                    metrics=['accuracy'])

model = LeNet(x_train[0].shape, num_classes)
model.summary()

Model: "le_net"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
conv2d (Conv2D)              (None, 28, 28, 6)         156       
_________________________________________________________________
average_pooling2d (AveragePo (None, 14, 14, 6)         0         
_________________________________________________________________
conv2d_1 (Conv2D)            (None, 10, 10, 16)        2416      
_________________________________________________________________
average_pooling2d_1 (Average (None, 5, 5, 16)          0         
_________________________________________________________________
flatten (Flatten)            (None, 400)               0         
_________________________________________________________________
dense (Dense)                (None, 120)               48120     
_________________________________________________________________
dense_1 (Dense)              (None, 84)                10164     
_________________________________________________________________
dense_2 (Dense)              (None, 10)                850       
=================================================================
Total params: 61,706
Trainable params: 61,706
Non-trainable params: 0

创建模型后，我们需要训练它的参数以使其强大。让我们训练给定数量的 epoch 模型。我们可以在TensorBoard中看到训练的进度。

# Place the logs in a timestamped subdirectory
# This allows to easy select different training runs
# In order not to overwrite some data, it is useful to have a name with a timestamp
log_dir="logs/fit/" + datetime.datetime.now().strftime("%Y%m%d-%H%M%S")
# Specify the callback object
tensorboard_callback = tf.keras.callbacks.TensorBoard(log_dir=log_dir, histogram_freq=1)

# tf.keras.callback.TensorBoard ensures that logs are created and stored
# We need to pass callback object to the fit method
# The way to do this is by passing the list of callback objects, which is in our case just one

model.fit(x_train, y=y_train, 
          epochs=20, 
          validation_data=(x_test, y_test), 
          callbacks=[tensorboard_callback],
          verbose=0)

%tensorboard --logdir logs/fit

仅 20 个 epoch 就已经不错了。

之后，我们可以做出一些预测并将其可视化。使用predict_classes()函数，我们可以从网络中获取准确的类别而不是概率值。它与使用numpy.argmax()相同。我们将用红色显示错误的预测，用蓝色表示正确的预测。

class_names = ['T-shirt/top', 'Trouser', 'Pullover', 'Dress', 'Coat',
               'Sandal', 'Shirt', 'Sneaker', 'Bag', 'Ankle boot']

prediction_values = model.predict_classes(x_test)

# set up the figure
fig = plt.figure(figsize=(15, 7))
fig.subplots_adjust(left=0, right=1, bottom=0, top=1, hspace=0.05, wspace=0.05)

# plot the images: each image is 28x28 pixels
for i in range(50):
    ax = fig.add_subplot(5, 10, i + 1, xticks=[], yticks=[])
    ax.imshow(x_test[i,:].reshape((28,28)),cmap=plt.cm.gray_r, interpolation='nearest')
  
    if prediction_values[i] == np.argmax(y_test[i]):
        # label the image with the blue text
        ax.text(0, 7, class_names[prediction_values[i]], color='blue')
    else:
        # label the image with the red text
        ax.text(0, 7, class_names[prediction_values[i]], color='red')