本篇試著以CNN實作手寫辨識，將每個步驟分割實作，以便釐清所有過程
前置作業
使用tensorflow搭建keras
%env KERAS_BACKEND=tensorflow
導入必要的package
其中 %matplotlib inline 這行用來內嵌繪圖，並且可以在最後省略 plt.show() 的撰寫
%matplotlib inline
import numpy as np
import matplotlib.pyplot as plt
將手寫辨識的model引入，也就是 MNIST
from keras.datasets import mnist # Using TensorFlow backend.
將model中的資料讀出並且分為 訓練資料(train) 和 測試資料(test)
(x_train, y_train), (x_test, y_test) = mnist.load_data()
input資料格式整理
模型中每筆資料都是 28 * 28 的陣列，CNN其實吃的資料型態是「圖」，所以不用像一般NN訓練將二維圖拉平。 but! 平常的圖都有R、G、B三個channel，而每個channel都是一個矩陣，也就是一張圖可能是三個矩陣! 但是我們資料的圖是灰階，也就是只有一個channel，所以要明確告知keras。 換句話說，我們的資料格式要從 (28, 28) 變成 (28, 28, 1)，這個 1 表示的就是 1 個 channel
x_train = x_train.reshape(60000, 28,28,1)
x_test = x_test.reshape(10000, 28,28,1)
查看第9487筆資料其格式是否為 28 * 28 一個 channel的型式
x_train[9487].shape #(28, 28, 1)
打印出原始的28*28資料查看，上述看出只有一個channel，故 index 為 0，將 channel從頭到尾打印
x = x_train[9487][:,:,0]
print(x)
# array([[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
# 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
# 0, 0],
# [ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
# 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
# 0, 0],
# [ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
# 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
# 0, 0],
# [ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
# 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
# 0, 0],
# [ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 240,
# 253, 165, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
# 0, 0],
# [ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 32, 241,
# 252, 252, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
# 0, 0],
# [ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 141, 252,
# 252, 252, 98, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
# 0, 0],
# [ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 30, 252, 252,
# 252, 252, 149, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
# 0, 0],
# [ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 21, 217, 252,
# 252, 252, 149, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
# 0, 0],
# [ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 134, 252,
# 252, 252, 223, 31, 0, 0, 0, 0, 0, 0, 0, 0, 0,
# 0, 0],
# [ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 134, 252,
# 252, 252, 253, 44, 0, 0, 0, 0, 0, 0, 0, 0, 0,
# 0, 0],
# [ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 8, 239,
# 252, 252, 253, 44, 0, 0, 0, 0, 0, 0, 0, 0, 0,
# 0, 0],
# [ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 238,
# 252, 252, 253, 44, 0, 0, 0, 0, 0, 0, 0, 0, 0,
# 0, 0],
# [ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 238,
# 252, 252, 253, 44, 0, 0, 0, 0, 0, 0, 0, 0, 0,
# 0, 0],
# [ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 178,
# 253, 253, 255, 106, 0, 0, 0, 0, 0, 0, 0, 0, 0,
# 0, 0],
# [ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 90,
# 252, 252, 253, 206, 21, 0, 0, 0, 0, 0, 0, 0, 0,
# 0, 0],
# [ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 90,
# 252, 252, 253, 252, 88, 0, 0, 0, 0, 0, 0, 0, 0,
# 0, 0],
# [ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 90,
# 252, 252, 253, 252, 88, 0, 0, 0, 0, 0, 0, 0, 0,
# 0, 0],
# [ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 63,
# 234, 252, 253, 252, 220, 0, 0, 0, 0, 0, 0, 0, 0,
# 0, 0],
# [ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
# 176, 252, 253, 252, 237, 0, 0, 0, 0, 0, 0, 0, 0,
# 0, 0],
# [ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
# 45, 252, 253, 252, 245, 71, 0, 0, 0, 0, 0, 0, 0,
# 0, 0],
# [ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
# 29, 215, 253, 252, 237, 8, 0, 0, 0, 0, 0, 0, 0,
# 0, 0],
# [ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
# 0, 149, 253, 252, 237, 0, 0, 0, 0, 0, 0, 0, 0,
# 0, 0],
# [ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
# 0, 61, 165, 252, 237, 0, 0, 0, 0, 0, 0, 0, 0,
# 0, 0],
# [ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
# 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
# 0, 0],
# [ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
# 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
# 0, 0],
# [ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
# 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
# 0, 0],
# [ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
# 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
# 0, 0]], dtype=uint8)
畫出來
plt.imshow(x, cmap='Greys') # <matplotlib.image. AxesImage at 0x275d255cd68>
輸出格式整理
from keras.utils import np_utils
y_train=np_utils.to_categorical(y_train, 10)
y_test = np_utils.to_categorical(y_test,10)
打造CNN
CNN有兩種Hidden Layer
- Convolution Layer
如下圖，透過kernel(filter)縮小圖片尺寸方便CNN訓練，減少訓練時間
由於Convolution透過kernel提取特徵的作法，使圖片中的訊息易漏失，如下圖，同樣圖片左右兩邊的提取會不同
- Pooling Layer
透過池化(pooling)，例如Max-pooling可以去除雜訊(noise)保留圖片的重要訊息，如下圖，
假設我們要做的CNN如下
- 做3次convolution(意味著要做三次filter) 每次都接一個max-pooling
- filter 大小我們設為 3 * 3
- max-pooling 大小為2*2
CNN 的一個小技巧是每層的filter數目是越來越多。 做完convolution之後，我們會將其拉平成一個一為向量，再送入一個標準的神經網路。 這個神經網路設計是
- 只有一個Hidden-Layer
* 200個神經元
下面導入的包中，layer分別代表的是
- Sequential: 空白神經網路
- Dense: 標準神經網路
- Activation: 激活函數
- Flatten: CNN做完之後在進入Fully Connected Layer前會透過這個平坦化(Flatten)攤平合併到一個大向量內，
如下圖，Flatten攤平合併成一個大向量後，再接到ANN(NN)，最後output結果
from keras.models import Sequential
from keras.layers import Dense, Activation, Flatten, Conv2D, MaxPooling2D
from keras.optimizers import SGD
先建立一個空的神經網路
model=Sequential()
建立convolution
首先，在神經網路中，先建立第1層，參數分別代表
Conv2D( filter數量 , (filter大小) , 輸出大小一樣 , input大小 )
model.add(Conv2D(32, (3,3), padding='same', input_shape=(28,28,1)))
model.add(Activation('relu'))
加入池化層(pooling layer)
接著加入第一層的pooling層，並設定pool size 為 2 * 2 矩陣
model.add(MaxPooling2D(pool_size=(2,2)))
依照第一層做法，接著建構第二層，並將kernel(filter隨之放大)
model.add(Conv2D(64, (3,3), padding='same'))
model.add(Activation('relu'))
model.add(MaxPooling2D(pool_size=(2,2)))
model.add(Conv2D(128, (3,3), padding='same'))
model.add(Activation('relu'))
model.add(MaxPooling2D(pool_size=(2,2)))
接著用Flatten拉成一向量，並使用200個(可選)神經元的，且使用relu(可選)激活
model.add(Flatten())
model.add(Dense(200))
model.add(Activation('relu'))
最後，輸出使用10個神經元的標準神經網路，並用softmax激活(因為希望數字和為1)
model.add(Dense(10))
model.add(Activation('softmax'))
組裝
損失函數(Loss Function)使用MSE 優化器(optimizer)使用SGD 指標(metrics)使用準確率(accuracy)
model.compile(loss='mse', optimizer=SGD(lr=0.05), metrics=['accuracy'])
檢視CNN網路架構
model.summary()
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
conv2d_1 (Conv2D) (None, 28, 28, 32) 320
_________________________________________________________________
activation_1 (Activation) (None, 28, 28, 32) 0
_________________________________________________________________
max_pooling2d_1 (MaxPooling2 (None, 14, 14, 32) 0
_________________________________________________________________
conv2d_2 (Conv2D) (None, 14, 14, 64) 18496
_________________________________________________________________
activation_2 (Activation) (None, 14, 14, 64) 0
_________________________________________________________________
max_pooling2d_2 (MaxPooling2 (None, 7, 7, 64) 0
_________________________________________________________________
conv2d_3 (Conv2D) (None, 7, 7, 128) 73856
_________________________________________________________________
activation_3 (Activation) (None, 7, 7, 128) 0
_________________________________________________________________
max_pooling2d_3 (MaxPooling2 (None, 3, 3, 128) 0
_________________________________________________________________
flatten_1 (Flatten) (None, 1152) 0
_________________________________________________________________
dense_1 (Dense) (None, 200) 230600
_________________________________________________________________
activation_4 (Activation) (None, 200) 0
_________________________________________________________________
dense_2 (Dense) (None, 10) 2010
_________________________________________________________________
activation_5 (Activation) (None, 10) 0
=================================================================
Total params: 325,282
Trainable params: 325,282
Non-trainable params: 0
_________________________________________________________________
參數計算
每一筆input的通道都會透過filter做特徵提取(feture ectraction)，其中filter大小代表每個filter中擁有的神經元數量，且每個filter會有一個偏差值(bias)，每個filter中的神經元所用參數量表示為 (input資料的通道數量 * (filter大小) + bias) * filter數量
接著，每個filter產出的Feture Map經過filter過濾後，通道數會因filter而改變，也就是說conv_1產出的Feture Map的通道數量會因為conv_1中的32個filter增加為32個channel，使得 conv_2要input的資料變成32個通道，所以conv_2的 input資料的通道數量 等於32
conv_1: (1 * 3 * 3 + 1) * 32 = 320 conv_2: (32 * 3 * 3 + 1) * 64 = 18496
model.fit(x_train, y_train, batch_size=100, epochs=12)
# 60000/60000 [==============================] - 8s 134us/step - loss: 0.120- acc: 0.3652
# Epoch 2/12
# 60000/60000 [==============================] - 5s 81us/step - loss: 0.007- acc: 0.9534
# Epoch 3/12
# 60000/60000 [==============================] - 5s 80us/step - loss: 0.004- acc: 0.9732
# Epoch 4/12
# 60000/60000 [==============================] - 5s 80us/step - loss: 0.002- acc: 0.9810
# Epoch 5/12
# 60000/60000 [==============================] - 5s 82us/step - loss: 0.002- acc: 0.9852
# Epoch 6/12
# 60000/60000 [==============================] - 5s 81us/step - loss: 0.001- acc: 0.9881
# Epoch 7/12
# 60000/60000 [==============================] - 5s 80us/step - loss: 0.001- acc: 0.9906
# Epoch 8/12
# 60000/60000 [==============================] - 5s 79us/step - loss: 0.001- acc: 0.9921
# Epoch 9/12
# 60000/60000 [==============================] - 5s 80us/step - loss: 0.001- acc: 0.9933
# Epoch 10/12
# 60000/60000 [==============================] - 5s 81us/step - loss9.7978e-04 - acc: 0.9943
# Epoch 11/12
# 60000/60000 [==============================] - 5s 82us/step - loss8.5365e-04 - acc: 0.9952
# Epoch 12/12
# 60000/60000 [==============================] - 5s 82us/step - loss7.6462e-04 - acc: 0.9958
<keras.callbacks.History at 0x275d28d7ac8>
結果測試
分數
score = model.evaluate(x_test, y_test)
# 10000/10000 [==============================] - 1s 81us/step
查看loss function和準確率
print('loss:', score[0])
print('acc:', score[1])
# loss: 0.002139582320782871
# acc: 0.9862
儲存model和weight
model_json= model.to_json()
open('handwriteing_model_cnn.json', 'w').write(model_json)
model.save_weights('handwriting_weights_cnn.h5')
丟入test資料做驗證
predict = model.predict_classes(x_test)
使用隨機資料做抽取驗證
pick=np.random.randint(1,9999,5)
for i in range(5):
plt.subplot(1,5,i+1)
plt.imshow(x_test[pick[i]].reshape(28,28), cmap='Greys')
plt.title(predict[pick[i]])
plt.axis("off")
