行動技術與應用
Lesson 9: Convolutional Neural Network
Outline
- Keras R安裝與運作回顧
- 什麼是「卷積神經網路」(CNN)
- 卷積神經網路應用於MNIST資料集
Keras R安裝回顧
NN圖解
輸入層
隱藏層
卷積核
TensorFlow/Keras(R版本)安裝
R-Keras開發環境: 總整理
r-tensorflow
conda create --name r-tensorflow python=3.7
activate r-tensorflowpip install -I -U tensorflow
conda install -c conda-forge keras
(R Rtools)
TensorFlow R
keras R
devtools::install_github("rstudio/tensorflow")
devtools::install_github("rstudio/keras")
library(keras) ## Keras for R
❹
❶
❷
❸
❺
TensorFlow/Keras(R版本) 測試1
在RStudio內測試rstudio/tensorflow
```{r eval=FALSE}
# 安裝'devtools' package：方便從github安裝套件
if(!"devtools" %in% installed.packages())
install.packages('devtools')
require(devtools)
# install tensorflow(如果你要tensorflow的話)
devtools::install_github("rstudio/tensorflow")
# installing keras(如果你要keras的話)
devtools::install_github("rstudio/keras")
```
```{r eval=FALSE}
# 測試tensorflow: Hello
require(tensorflow)
session = tf$Session()
hello <- tf$constant('Hi, TensorFlow!')
session$run(hello)
```hello.rmd
TensorFlow/Keras(R版本) 測試2
在RStudio內測試rstudio/keras
```{r eval=FALSE}
library(keras)
model <- keras_model_sequential()
model %>%
layer_dense(units = 256, activation = 'relu', input_shape= c(784)) %>%
layer_dropout(rate = 0.4) %>%
layer_dense(units = 128, activation = 'relu') %>%
layer_dropout(rate = 0.3) %>%
layer_dense(units = 10, activation = 'softmax')
summary(model)
```keras.rmd
Keras 基本運作方式回顧
- 安裝keras R
- 準備資料集: 訓練集/測試集- 以FashionMinst資料集為例
- 建立深度學習模型: Sequential Model
- 建立模型: 隱藏層
- 編譯模型: 損失函數、優化演算法
- 訓練與評估(模型正確機率)
- 目的: 找出最適函數(最佳預測函數)
- 應用(預測)
Keras R安裝Keras R套件
- 安裝 Keras R套件
```{r eval=FALSE}
if(!"devtools" %in% installed.packages())
install.packages("devtools");
if(!"keras" %in% installed.packages()) {
require(devtools);
install_github("rstudio/keras");
}
library(keras)
install_keras() # 如需GPU設定，參考文件網址
```install_keras() 文件網址：https://keras.rstudio.com/reference/install_keras.html
Keras R準備資料: Fashion-MNIST資料集
單張圖: 28*28圖點, 256灰階
0 T-shirt/top
1 Trouser
2 Pullover
3 Dress
4 Coat
5 Sandal
6 Shirt
7 Sneaker
8 Bag
9 Ankle boot
10種分類: 標記0~9
Keras R準備資料: Fashion-MNIST資料集
```{r eval=FALSE}
fashion_mnist <- dataset_fashion_mnist()
# 訓練圖片/訓練標籤, 測試圖片/測試標籤
c(train_images, train_labels) %<-% fashion_mnist$train
c(test_images, test_labels) %<-% fashion_mnist$test
class_names = c('T-shirt/top',
'Trouser',
'Pullover',
'Dress',
'Coat',
'Sandal',
'Shirt',
'Sneaker',
'Bag',
'Ankle boot')
```2. 準備資料：載入
Keras R準備資料: Fashion-MNIST資料集
```{r eval=FALSE}
dim(train_images)
dim(train_labels)
train_labels[1:20]
dim(test_images)
dim(test_labels)
train_images <- train_images / 255
test_images <- test_images / 255
```2. 準備資料：檢視資料/資料前處理
訓練集: 60000萬筆，測試集: 10000筆
Keras R建立模型: Fashion-MNIST資料集
```{r eval=FALSE}
# 建立模型
model <- keras_model_sequential()
model %>%
layer_flatten(input_shape = c(28, 28)) %>%
layer_dense(units = 128, activation = 'relu') %>%
layer_dense(units = 10, activation = 'softmax')
# 編譯模型
model %>% compile(
optimizer = 'adam',
loss = 'sparse_categorical_crossentropy',
metrics = c('accuracy')
)
```3. 建立深度學習模型: 建立/編譯
隱藏層: 1層
Keras R訓練模型: Fashion-MNIST資料集
```{r eval=FALSE}
# 訓練模型
model %>% fit(train_images, train_labels, epochs = 5)
```4. 訓練模型
Epoch 1/5
60000/60000 [==============================] - 5s 79us/step - loss: 0.4923 - acc: 0.8273
Epoch 2/5
60000/60000 [==============================] - 4s 70us/step - loss: 0.3740 - acc: 0.8641
Epoch 3/5
60000/60000 [==============================] - 4s 70us/step - loss: 0.3346 - acc: 0.8775
Epoch 4/5
60000/60000 [==============================] - 4s 71us/step - loss: 0.3117 - acc: 0.8861
Epoch 5/5
60000/60000 [==============================] - 4s 75us/step - loss: 0.2931 - acc: 0.8916
Keras R訓練模型: Fashion-MNIST資料集
4. 訓練模型
0.8916
0.2931
Keras R評估模型: Fashion-MNIST資料集
```{r eval=FALSE}
# 評估正確性
score <- model %>% evaluate(test_images, test_labels)
cat('Test loss:', score$loss, "\n")
cat('Test accuracy:', score$acc, "\n")
```5. 評估模型(測試集)
Test loss: 0.3526644
Test accuracy: 0.8737
Keras R運用: Fashion-MNIST資料集
```{r eval=FALSE}
# 檢視一筆預測結果
predictions <- model %>% predict(test_images) # 10000筆測試集套用模型預測
predictions[1, ] # 印出第一筆
answer <- which.max(predictions[1, ]) - 1 # 預測機率最高者 - 1
test_labels[1] # 陣列index:1-10,對應test_labels: 0-9
# 印出25筆預測結果
par(mfcol=c(5,5))
par(mar=c(0, 0, 1.5, 0), xaxs='i', yaxs='i')
for (i in 1:25) {
img <- test_images[i, , ]
img <- t(apply(img, 2, rev))
# subtract 1 as labels go from 0 to 9
predicted_label <- which.max(predictions[i, ]) - 1
true_label <- test_labels[i]
if (predicted_label == true_label) {
color <- '#008800'
} else {
color <- '#bb0000'
}
image(1:28, 1:28, img, col = gray((0:255)/255), xaxt = 'n', yaxt = 'n',
main = paste0(class_names[predicted_label + 1], " (",
class_names[true_label + 1], ")"),
col.main = color)
}
```Keras R運用: Fashion-MNIST資料集
紅色為預測錯誤
Keras R應用: Fashion-MNIST資料集
```{r eval=FALSE}
# 預測新影像
img <- test_images[1, , , drop = FALSE] # 此處以測試集第一張替代
dim(img) # 28*28矩陣 0~1數值:灰階
predictions <- model %>% predict(img) # 以模型進行預測
predictions # 分類機率串列(1~10)
# 減1 因為標籤為0～9
which.max(predictions)-1 # 預測答案
class_pred <- model %>% predict_classes(img)
class_pred # 標準答案
```---
title: "Fashion MINST"
output: html_document
---
## 安裝 Keras R (rstudio/keras)
```{r eval=FALSE}
if(!"devtools" %in% installed.packages())
install.packages("devtools");
if(!"keras" %in% installed.packages()) {
require(devtools);
install_github("rstudio/keras");
}
if(!"ggplot2" %in% installed.packages()) {
install.packages("ggplot2");
}
```
```{r eval=FALSE}
## 1.載入Keras
library(keras)
# install_keras()
```
```{r eval=FALSE}
## 2.準備資料：載入fashion mnist資料庫: 含訓練集/測試
fashion_mnist <- dataset_fashion_mnist()
# 訓練圖片/訓練標籤, 測試圖片/測試標籤
c(train_images, train_labels) %<-% fashion_mnist$train
c(test_images, test_labels) %<-% fashion_mnist$test
class_names = c('T-shirt/top',
'Trouser',
'Pullover',
'Dress',
'Coat',
'Sandal',
'Shirt',
'Sneaker',
'Bag',
'Ankle boot')
```
```{r eval=FALSE}
## 3.準備資料：檢視資料/資料前處理
dim(train_images)
dim(train_labels)
train_labels[1:20]
dim(test_images)
dim(test_labels)
train_images <- train_images / 255
test_images <- test_images / 255
```
```{r eval=FALSE}
# 建立模型
model <- keras_model_sequential()
model %>%
layer_flatten(input_shape = c(28, 28)) %>%
layer_dense(units = 128, activation = 'relu') %>%
layer_dense(units = 10, activation = 'softmax')
# 編譯模型
model %>% compile(
optimizer = 'adam',
loss = 'sparse_categorical_crossentropy',
metrics = c('accuracy')
)
```
```{r eval=FALSE}
# 訓練模型
model %>% fit(train_images, train_labels, epochs = 5)
# 評估正確性
score <- model %>% evaluate(test_images, test_labels)
cat('Test loss:', score$loss, "\n")
cat('Test accuracy:', score$acc, "\n")
```
```{r eval=FALSE}
# 檢視一筆預測結果
predictions <- model %>% predict(test_images) # 10000筆測試集套用模型預測
predictions[1, ] # 印出第一筆
answer <- which.max(predictions[1, ]) - 1 # 預測機率最高者 - 1
test_labels[1] # 陣列index:1-10,對應test_labels: 0-9
# 印出25筆預測結果
par(mfcol=c(5,5))
par(mar=c(0, 0, 1.5, 0), xaxs='i', yaxs='i')
for (i in 1:25) {
img <- test_images[i, , ]
img <- t(apply(img, 2, rev))
# subtract 1 as labels go from 0 to 9
predicted_label <- which.max(predictions[i, ]) - 1
true_label <- test_labels[i]
if (predicted_label == true_label) {
color <- '#008800'
} else {
color <- '#bb0000'
}
image(1:28, 1:28, img, col = gray((0:255)/255), xaxt = 'n', yaxt = 'n',
main = paste0(class_names[predicted_label + 1], " (",
class_names[true_label + 1], ")"),
col.main = color)
}
```
```{r eval=FALSE}
# 預測新影像
img <- test_images[1, , , drop = FALSE] # 此處以測試集第一張替代
dim(img) # 28*28矩陣 0~1數值:灰階
predictions <- model %>% predict(img) # 以模型進行預測
predictions # 分類機率串列(1~10)
# 減1 因為標籤為0～9
which.max(predictions)-1 # 預測答案
class_pred <- model %>% predict_classes(img)
class_pred # 標準答案
```完整範例
Keras R修改(1/2)
```{r eval=FALSE}
# 建立模型
model <- keras_model_sequential()
model %>%
layer_flatten(input_shape = c(28, 28)) %>%
layer_dense(units = 128, activation = 'relu') %>%
layer_dense(units = 10, activation = 'softmax')
```隱藏層: 1層
增加 layer_dense() 全連接層, layer_dropout() 降低模型複雜度
```{r eval=FALSE}
# 訓練模型
model %>% fit(train_images, train_labels, epochs = 5)
```❶修改模型
❷調整訓練參數
調整fit()參數：epochs(回合數), batch_size預設值32
Keras R修改(2/2)
```{r eval=FALSE}
# 編譯模型
model %>% compile(
optimizer = 'adam',
loss = 'sparse_categorical_crossentropy',
metrics = c('accuracy')
)
```❸修改優化參數
optimizer: 優化器(優化演算法)
loss: 損失函數
Keras R 優化函式(optimization)
優化過程: 迭代進行
- Epochs參數
- 優化演算法
梯度搜尋
找出最適函數
優化目的: 找出最適函數
- 誤差最小
- 函數: 直線/曲線
誤差? 損失函數
Keras R 損失函數(loss function)
損失函數：
- 計算目前的「最適函數」與資料點的誤差
均方根誤差
絕對誤差
平方和誤差
Keras R 優化函式(optimization)
Keras R 優化器
model %>% compile(
optimizer = 'adam',
loss = 'sparse_categorical_crossentropy',
metrics = c('accuracy')
)
Keras R 損失函數(loss function)
只能用於分類問題
卷積神經網路CNN理論
Convolutional Neural Network (CNN)
模仿人腦視覺神經運作
1962, Hubel & Weisel的實驗
神經元可辨識不同形狀(低階特徵)
低階組合辨識高階特徵(例：人臉)
組合
特徵階層
卷積神經網路CNN理論
Convolutional Neural Network (CNN)
(1) 底層逐層學習低階特徵 (2)由低階特徵組合出高階特徵
像素明暗
形狀邊緣
更多形狀物體
哪些形狀物件可辨識人臉
低階
高階
低階
高階
卷積神經網路CNN適用範圍
Convolutional Neural Network (CNN)
(1) 適用於classification tasks
(2) 輸入為: images, video, text, or sound
特徵學習
卷積神經網路CNN架構
Convolutional Neural Network (CNN)
(1) 卷積層 (convolution layer): 學習特徵
(2) 池化層 (pooling Layer): 取樣 subsampling
(3) 全連接層: 原multi-layer perceptron
卷積層
池化層
特徵學習
分類模型
全連接層
卷積層
池化層
原來之multilayer perceptron網路
卷積神經網路CNN架構-卷積層
Convolutional Neural Network (CNN)
(1) 卷積層 (convolution layer): 學習特徵
2.決定使用3*3矩陣(濾鏡)
1. 輸入圖片5*5
3.產生3*3 特徵圖
(5-3+1)
Kernel 或稱 Filter: 就是濾鏡的功能
濾鏡的效果: 用於邊緣偵測的情形
卷積神經網路CNN架構-池化層
Convolutional Neural Network (CNN)
(2) 池化層 (pooling Layer): 取樣 subsampling
1.N*N特徵圖
3.產生2*2 池化特徵圖
2.使用N/2*N/2取樣
最大池化: 取樣該區最大值➞代表值
卷積神經網路CNN架構
Convolutional Neural Network (CNN)
特徵擷取: 兩層卷積層+兩層池化層 (c1)
2.決定使用5*5矩陣
1. 輸入圖片32*32
3. C1產生28*28 特徵圖
(32-5+1)
4. 2*2取樣
5. S1產生14*14特徵圖
6. C2產生10*10特徵圖
(14-5+1)
7. 2*2取樣
8. S1產生5*5特徵圖
(28/2)
(10/2)
9. 輸入層:5*5特徵圖
卷積神經網路CNN: MNIST
Convolutional Neural Network (CNN)
卷積神經網路CNN: MNIST
Convolutional Neural Network (CNN)
平坦層
隱藏層
輸出層
10個濾鏡
20個濾鏡
MNIST資料集
訓練集：60000筆
測試集：10000筆
3維陣列（images筆數, 圖寬, 圖高）
圖：28*28 共784圖點
60000 rows
28 columns
.
.
.
28 slices
MNIST資料集CNN:模型設定
model <- keras_model_sequential() %>%
# 卷積層C1
layer_conv_2d(filters = 32, kernel_size = c(3,3), activation = 'relu',
input_shape = input_shape) %>%
# 卷積層C2
layer_conv_2d(filters = 64, kernel_size = c(3,3), activation = 'relu') %>%
# 池化層S2
layer_max_pooling_2d(pool_size = c(2, 2)) %>%
layer_dropout(rate = 0.25) %>%
# 平坦層 12＊12﹡64 節點
layer_flatten() %>%
# 隱藏層
layer_dense(units = 128, activation = 'relu') %>%
layer_dropout(rate = 0.5) %>%
# 輸出層
layer_dense(units = num_classes, activation = 'softmax')卷積層C1: 輸出32圖層, 3*3矩陣, 輸入層 (28, 28, 1圖層)
卷積層C2: 輸出64圖層, 3*3矩陣
池化層S2: 縮減圖點成12*12, 仍是64圖層
MNIST資料集CNN:模型設定
12*12*24個節點
(9216+1) * 128
(128+1) * 10
（3*3+1)*32個
32*（24*24)+64
MNIST資料集CNN:評估設定
# Compile model
model %>% compile(
loss = loss_categorical_crossentropy,
optimizer = optimizer_adadelta(),
metrics = c('accuracy')
)
MNIST資料集CNN:訓練階段
# Train model
model %>% fit(
x_train, y_train,
batch_size = batch_size,
epochs = epochs,
validation_split = 0.2
)# Data Preparation ------------------------------------
batch_size <- 128 # 分批大小
num_classes <- 10 # 輸出層節點數0~9機率
epochs <- 12 # 訓練回合數
卷積神經網路CNN:完整範例
```{r eval=FALSE}
library(keras)
# Data Preparation -----------------------------------------------------
batch_size <- 128 # 分批大小
num_classes <- 10 # 輸出層節點數0~9機率
epochs <- 12 # 訓練回合數
# 圖片圖點數 28*28
img_rows <- 28
img_cols <- 28
# 讀取資料集, 分成訓練集,測試集
mnist <- dataset_mnist()
x_train <- mnist$train$x
y_train <- mnist$train$y
x_test <- mnist$test$x
y_test <- mnist$test$y
# 將原本資料集的三維陣列,轉為矩陣
x_train <- array_reshape(x_train, c(nrow(x_train), img_rows, img_cols, 1))
x_test <- array_reshape(x_test, c(nrow(x_test), img_rows, img_cols, 1))
input_shape <- c(img_rows, img_cols, 1) ## 輸入圖片 28*28* 1圖層
# 將圖點灰階0~255 轉成 [0,1] range
x_train <- x_train / 255
x_test <- x_test / 255
cat('x_train_shape:', dim(x_train), '\n')
cat(nrow(x_train), 'train samples\n')
cat(nrow(x_test), 'test samples\n')
# 將答案轉成矩陣
y_train <- to_categorical(y_train, num_classes)
y_test <- to_categorical(y_test, num_classes)
# Define Model -----------------------------------------------------------
# Define model
model <- keras_model_sequential() %>%
layer_conv_2d(filters = 32, kernel_size = c(3,3), activation = 'relu',
input_shape = input_shape) %>%
layer_conv_2d(filters = 64, kernel_size = c(3,3), activation = 'relu') %>%
layer_max_pooling_2d(pool_size = c(2, 2)) %>%
layer_dropout(rate = 0.25) %>%
layer_flatten() %>%
layer_dense(units = 128, activation = 'relu') %>%
layer_dropout(rate = 0.5) %>%
layer_dense(units = num_classes, activation = 'softmax')
# Compile model
model %>% compile(
loss = loss_categorical_crossentropy,
optimizer = optimizer_adadelta(),
metrics = c('accuracy')
)
# Train model
model %>% fit(
x_train, y_train,
batch_size = batch_size,
epochs = epochs,
validation_split = 0.2
)
scores <- model %>% evaluate(
x_test, y_test, verbose = 0
)
# Output metrics
cat('Test loss:', scores[[1]], '\n')
cat('Test accuracy:', scores[[2]], '\n')
```
卷積神經網路CNN:訓練結果
Train on 48000 samples, validate on 12000 samples
Epoch 1/12
48000/48000 [==============================] - 141s 3ms/step - loss: 0.3055 - acc: 0.9050 - val_loss: 0.0725 - val_acc: 0.9803
Epoch 2/12
48000/48000 [==============================] - 148s 3ms/step - loss: 0.0989 - acc: 0.9709 - val_loss: 0.0539 - val_acc: 0.9845
Epoch 3/12
48000/48000 [==============================] - 145s 3ms/step - loss: 0.0737 - acc: 0.9787 - val_loss: 0.0455 - val_acc: 0.9868
Epoch 4/12
48000/48000 [==============================] - 146s 3ms/step - loss: 0.0586 - acc: 0.9823 - val_loss: 0.0427 - val_acc: 0.9881
Epoch 5/12
48000/48000 [==============================] - 148s 3ms/step - loss: 0.0519 - acc: 0.9846 - val_loss: 0.0393 - val_acc: 0.9887
Epoch 6/12
48000/48000 [==============================] - 149s 3ms/step - loss: 0.0458 - acc: 0.9863 - val_loss: 0.0398 - val_acc: 0.9891
Epoch 7/12
48000/48000 [==============================] - 114s 2ms/step - loss: 0.0375 - acc: 0.9879 - val_loss: 0.0396 - val_acc: 0.9880
Epoch 8/12
48000/48000 [==============================] - 113s 2ms/step - loss: 0.0363 - acc: 0.9886 - val_loss: 0.0390 - val_acc: 0.9892
Epoch 9/12
48000/48000 [==============================] - 113s 2ms/step - loss: 0.0336 - acc: 0.9895 - val_loss: 0.0372 - val_acc: 0.9905
Epoch 10/12
48000/48000 [==============================] - 122s 3ms/step - loss: 0.0306 - acc: 0.9907 - val_loss: 0.0375 - val_acc: 0.9899
Epoch 11/12
48000/48000 [==============================] - 138s 3ms/step - loss: 0.0255 - acc: 0.9919 - val_loss: 0.0381 - val_acc: 0.9899
Epoch 12/12
48000/48000 [==============================] - 139s 3ms/step - loss: 0.0264 - acc: 0.9917 - val_loss: 0.0350 - val_acc: 0.9903卷積神經網路CNN:訓練結果
卷積神經網路CNN:測試結果
> # Output metrics
> cat('Test loss:', scores[[1]], '\n')
Test loss: 0.02830164
> cat('Test accuracy:', scores[[2]], '\n')
Test accuracy: 0.9909 MLP 測試正確率為 98.18%
CNN為99.09%
行動技術與應用
By Leuo-Hong Wang
行動技術與應用
Lesson 9: Convolutional Neural Network
