[Kaggle] Digit Recognizer 手写数字识别（卷积神经网络）

Digit Recognizer 练习地址

相关博文：
[Hands On ML] 3. 分类（MNIST手写数字预测）
[Kaggle] Digit Recognizer 手写数字识别
[Kaggle] Digit Recognizer 手写数字识别（简单神经网络）
04.卷积神经网络 W1.卷积神经网络

上一篇的简单神经网络，将28*28的图片展平了，每个像素在空间上的位置关系是没有考虑的，空间的信息丢失。

1. 使用 LeNet 预测

LeNet神经网络 参考博文

1.1 导入包

from keras import backend as K # 兼容不同后端的代码
from keras.models import Sequential
from keras.layers.convolutional import Conv2D
from keras.layers.convolutional import MaxPooling2D
from keras.layers.core import Activation
from keras.layers.core import Dense
from keras.layers.core import Flatten
from keras.utils import np_utils
from keras.optimizers import SGD, Adam, RMSpropimport numpy as np
%matplotlib inline
import matplotlib.pyplot as plt
import pandas as pd

1.2 建立 LeNet 模型

# 图片格式问题
# K.image_data_format() == 'channels_last' 
# 默认是last是通道  K.set_image_dim_ordering("tf")
# K.image_data_format() == 'channels_first' # K.set_image_dim_ordering("th")class LeNet:@staticmethoddef build(input_shape, classes):model = Sequential()model.add(Conv2D(20,kernel_size=5,padding='same',input_shape=input_shape,activation='relu'))model.add(MaxPooling2D(pool_size=(2,2),strides=(2,2)))model.add(Conv2D(50,kernel_size=5,padding='same',activation='relu'))model.add(MaxPooling2D(pool_size=(2,2),strides=(2,2)))model.add(Flatten())model.add(Dense(500, activation='relu'))model.add(Dense(classes,activation='softmax'))return model

1.3 读入数据

train = pd.read_csv('train.csv')
y_train_full = train['label']
X_train_full = train.drop(['label'], axis=1)
X_test_full = pd.read_csv('test.csv')

X_train_full.shape

输出：

(42000, 784)

• 数据格式转换，增加一个通道维度

X_train = np.array(X_train_full).reshape(-1,28,28) / 255.0
X_test = np.array(X_test_full).reshape(-1,28,28)/255.0
y_train = np_utils.to_categorical(y_train_full, 10) # 转成oh编码X_train = X_train[:, :, :, np.newaxis]
# m,28,28 --> m, 28, 28, 1(单通道)
X_test = X_test[:, :, :, np.newaxis]

1.4 定义模型

model = LeNet.build(input_shape=(28, 28, 1), classes=10)

• 定义优化器，配置模型

opt = Adam(learning_rate=0.001, beta_1=0.9, beta_2=0.999, decay=0.01)
model.compile(loss="categorical_crossentropy",optimizer=opt, metrics=["accuracy"])

注意：标签不采用 one-hot 编码的话，这里使用 loss="sparse_categorical_crossentropy"

1.5 训练

history = model.fit(X_train, y_train, epochs=20, batch_size=128,validation_split=0.2)

Epoch 1/20
263/263 [==============================] - 26s 98ms/step - 
loss: 0.2554 - accuracy: 0.9235 - 
val_loss: 0.0983 - val_accuracy: 0.9699
Epoch 2/20
263/263 [==============================] - 27s 103ms/step - 
loss: 0.0806 - accuracy: 0.9761 - 
val_loss: 0.0664 - val_accuracy: 0.9787
...
...
Epoch 20/20
263/263 [==============================] - 25s 97ms/step - 
loss: 0.0182 - accuracy: 0.9953 - 
val_loss: 0.0405 - val_accuracy: 0.9868

可以看见第2轮迭代结束，训练集准确率就 97.6%了，效果比之前的简单神经网络好很多

• 模型总结

model.summary()

Model: "sequential"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
conv2d (Conv2D)              (None, 28, 28, 20)        520       
_________________________________________________________________
max_pooling2d (MaxPooling2D) (None, 14, 14, 20)        0         
_________________________________________________________________
conv2d_1 (Conv2D)            (None, 14, 14, 50)        25050     
_________________________________________________________________
max_pooling2d_1 (MaxPooling2 (None, 7, 7, 50)          0         
_________________________________________________________________
flatten (Flatten)            (None, 2450)              0         
_________________________________________________________________
dense (Dense)                (None, 500)               1225500   
_________________________________________________________________
dense_1 (Dense)              (None, 10)                5010      
=================================================================
Total params: 1,256,080
Trainable params: 1,256,080
Non-trainable params: 0
_________________________________________________________________

• 绘制模型结构图

from keras.utils import plot_model
plot_model(model, './model.png', show_shapes=True)

1.6 绘制训练曲线

pd.DataFrame(history.history).plot(figsize=(8, 5))
plt.grid(True)
plt.gca().set_ylim(0, 1) # set the vertical range to [0-1]
plt.show()

1.7 预测提交

y_pred = model.predict(X_test)
pred = y_pred.argmax(axis=1).reshape(-1)
print(pred.shape)image_id = pd.Series(range(1,len(pred)+1))
output = pd.DataFrame({
    'ImageId':image_id, 'Label':pred})
output.to_csv("submission_NN.csv",  index=False)

LeNet 模型得分 0.98607，比上一篇的简单NN模型（得分 0.97546），好了 1.061%

2. 使用 VGG16 迁移学习

VGG16 help 文档：

Help on function VGG16 in module tensorflow.python.keras.applications.vgg16:VGG16(include_top=True, weights='imagenet', input_tensor=None, input_shape=None, pooling=None, classes=1000, classifier_activation='softmax')Instantiates the VGG16 model.Reference paper:- [Very Deep Convolutional Networks for Large-Scale Image Recognition](https://arxiv.org/abs/1409.1556) (ICLR 2015)By default, it loads weights pre-trained on ImageNet. Check 'weights' forother options.This model can be built both with 'channels_first' data format(channels, height, width) or 'channels_last' data format(height, width, channels).The default input size for this model is 224x224.Caution: Be sure to properly pre-process your inputs to the application.Please see `applications.vgg16.preprocess_input` for an example.Arguments:include_top: whether to include the 3 fully-connectedlayers at the top of the network.weights: one of `None` (random initialization),'imagenet' (pre-training on ImageNet),or the path to the weights file to be loaded.input_tensor: optional Keras tensor(i.e. output of `layers.Input()`)to use as image input for the model.input_shape: optional shape tuple, only to be specifiedif `include_top` is False (otherwise the input shapehas to be `(224, 224, 3)`(with `channels_last` data format)or `(3, 224, 224)` (with `channels_first` data format).It should have exactly 3 input channels,and width and height should be no smaller than 32.E.g. `(200, 200, 3)` would be one valid value.pooling: Optional pooling mode for feature extractionwhen `include_top` is `False`.- `None` means that the output of the model will bethe 4D tensor output of thelast convolutional block.- `avg` means that global average poolingwill be applied to the output of thelast convolutional block, and thusthe output of the model will be a 2D tensor.- `max` means that global max pooling willbe applied.classes: optional number of classes to classify imagesinto, only to be specified if `include_top` is True, andif no `weights` argument is specified.classifier_activation: A `str` or callable. The activation function to useon the "top" layer. Ignored unless `include_top=True`. Set`classifier_activation=None` to return the logits of the "top" layer.Returns:A `keras.Model` instance.Raises:ValueError: in case of invalid argument for `weights`,or invalid input shape.ValueError: if `classifier_activation` is not `softmax` or `None` whenusing a pretrained top layer.

2.1 导入包

import numpy as np
%matplotlib inline
import matplotlib.pyplot as plt
import pandas as pd
import cv2
from keras.optimizers import Adam
from keras.models import Model
from keras.utils import np_utils
from keras.models import Sequential
from keras.layers import Flatten
from keras.layers import Dense
from keras.layers import Input
from keras.layers import Dropout
from keras.applications.vgg16 import VGG16

2.2 定义模型

vgg16 = VGG16(weights='imagenet',include_top=False,input_shape=(32, 32, 3))
# VGG16 模型在include_top=False时，可以自定义输入大小，至少32x32，通道必须是3mylayer = vgg16.output
mylayer = Flatten()(mylayer)
mylayer = Dense(128, activation='relu')(mylayer)
mylayer = Dropout(0.3)(mylayer)
mylayer = Dense(10, activation='softmax')(mylayer)model = Model(inputs=vgg16.inputs, outputs=mylayer)for layer in vgg16.layers:layer.trainable = False # vgg16的各个层不训练

2.3 数据处理

train = pd.read_csv('train.csv')
y_train_full = train['label']
X_train_full = train.drop(['label'], axis=1)
X_test_full = pd.read_csv('test.csv')

• 将单通道的数据，复制成3通道的（vgg16要求3通道的），再resize成 32*32的，vgg16 要求图片最低分辨率是 32*32

def process(data):data = np.array(data).reshape(-1,28,28)output = np.zeros((data.shape[0], 32, 32, 3))for i in range(data.shape[0]):img = data[i]rgb_array = np.zeros((img.shape[0], img.shape[1], 3), "uint8")rgb_array[:, :, 0], rgb_array[:, :, 1], rgb_array[:, :, 2] = img, img, imgpic = cv2.resize(rgb_array, (32, 32), interpolation=cv2.INTER_LINEAR)output[i] = picoutput = output.astype('float32')/255.0return output

y_train = np_utils.to_categorical(y_train_full, 10)
X_train = process(X_train_full)
X_test = process(X_test_full)print(X_train.shape)
print(X_test.shape)

输出：

(42000, 32, 32, 3)
(28000, 32, 32, 3)

• 看一看处理后的图片

img = X_train[0]
plt.imshow(img)
np.set_printoptions(threshold=np.inf)# 全部显示矩阵
# print(X_train[0])

2.4 配置模型、训练

opt = Adam(learning_rate=0.001, beta_1=0.9, beta_2=0.999, decay=0.01)
model.compile(loss="categorical_crossentropy",optimizer=opt, metrics=["accuracy"])

history = model.fit(X_train, y_train, epochs=50, batch_size=128,validation_split=0.2)

输出：

Epoch 1/50
263/263 [==============================] - 101s 384ms/step - 
loss: 0.9543 - accuracy: 0.7212 - 
val_loss: 0.5429 - val_accuracy: 0.8601
...
Epoch 10/50
263/263 [==============================] - 110s 417ms/step - 
loss: 0.3284 - accuracy: 0.9063 - 
val_loss: 0.2698 - val_accuracy: 0.9263
...
Epoch 40/50
263/263 [==============================] - 114s 433ms/step - 
loss: 0.2556 - accuracy: 0.9254 - 
val_loss: 0.2121 - val_accuracy: 0.9389
...
Epoch 50/50
263/263 [==============================] - 110s 420ms/step - 
loss: 0.2466 - accuracy: 0.9272 - 
val_loss: 0.2058 - val_accuracy: 0.9406

model.summary()

输出：

Model: "functional_15"
_________________________________________________________________
Layer (type)                 Output Shape              Param # 
=================================================================
input_23 (InputLayer)        [(None, 32, 32, 3)]       0         
_________________________________________________________________
block1_conv1 (Conv2D)        (None, 32, 32, 64)        1792      
_________________________________________________________________
block1_conv2 (Conv2D)        (None, 32, 32, 64)        36928     
_________________________________________________________________
block1_pool (MaxPooling2D)   (None, 16, 16, 64)        0         
_________________________________________________________________
block2_conv1 (Conv2D)        (None, 16, 16, 128)       73856     
_________________________________________________________________
block2_conv2 (Conv2D)        (None, 16, 16, 128)       147584    
_________________________________________________________________
block2_pool (MaxPooling2D)   (None, 8, 8, 128)         0         
_________________________________________________________________
block3_conv1 (Conv2D)        (None, 8, 8, 256)         295168    
_________________________________________________________________
block3_conv2 (Conv2D)        (None, 8, 8, 256)         590080    
_________________________________________________________________
block3_conv3 (Conv2D)        (None, 8, 8, 256)         590080    
_________________________________________________________________
block3_pool (MaxPooling2D)   (None, 4, 4, 256)         0         
_________________________________________________________________
block4_conv1 (Conv2D)        (None, 4, 4, 512)         1180160   
_________________________________________________________________
block4_conv2 (Conv2D)        (None, 4, 4, 512)         2359808   
_________________________________________________________________
block4_conv3 (Conv2D)        (None, 4, 4, 512)         2359808   
_________________________________________________________________
block4_pool (MaxPooling2D)   (None, 2, 2, 512)         0         
_________________________________________________________________
block5_conv1 (Conv2D)        (None, 2, 2, 512)         2359808   
_________________________________________________________________
block5_conv2 (Conv2D)        (None, 2, 2, 512)         2359808   
_________________________________________________________________
block5_conv3 (Conv2D)        (None, 2, 2, 512)         2359808   
_________________________________________________________________
block5_pool (MaxPooling2D)   (None, 1, 1, 512)         0         
_________________________________________________________________
flatten_19 (Flatten)         (None, 512)               0         
_________________________________________________________________
dense_28 (Dense)             (None, 128)               65664     
_________________________________________________________________
dropout_9 (Dropout)          (None, 128)               0         
_________________________________________________________________
dense_29 (Dense)             (None, 10)                1290      
=================================================================
Total params: 14,781,642
Trainable params: 66,954
Non-trainable params: 14,714,688
_________________________________________________________________

• 绘制模型结构

from keras.utils import plot_model
plot_model(model, './model.png', show_shapes=True)

2.5 预测提交

y_pred = model.predict(X_test)
pred = y_pred.argmax(axis=1).reshape(-1)
print(pred.shape)
print(pred)

image_id = pd.Series(range(1,len(pred)+1))
output = pd.DataFrame({
    'ImageId':image_id, 'Label':pred})
output.to_csv("submission_NN.csv",  index=False)

预测得分：0.93696

可能是由于 VGG16模型是用 224*224 的图片训练的权重，我们使用的是 28*28 的图片，可能不能很好的使用VGG16已经训练好的权重

---

我的CSDN博客地址 https://michael.blog.csdn.net/

长按或扫码关注我的公众号（Michael阿明），一起加油、一起学习进步！