การจำแนกดอกไม้ (Flower) โดยใช้ CNN อย่างง่ายบน Tensorflow + Keras

สรุปขั้นตอน

Prepare data
Build a model
Pre-process image
Train data with model
Show train result

Save model
Save history
Show time used
Plot train accuracy and loss

Prepare test data

Read test csv
[x] Load image
[x] Convert image to array.
[x] Put image array in list
[x] Convert list to numpy array
[y] Transform label in form of numpy array 0 to n-1 (LabelEncoder)
[y] Convert list to one-hot array (to_categorical)

Predict test data
Show test result

Evaluate model
Convert one-hot to list
Save prediction
Plot confusion matrix
Print and save classification report

1. Prepare data

ในตัวอย่างนี้ เราจะใช้ Flowers Recognition dataset เป็นรูปดอกไม้ 5 ชนิดแยกตามโฟลเดอร์

จัดโฟลเดอร์โปรเจ็คให้อยู่ในโครงสร้างต่อไปนี้

train

daisy
dandelion
rose
sunflower
tulip

test

หมายเหตุ ในโฟลเดอร์ที่โหลดมา จะมีไฟล์ชื่อ flickr.py และ flickr.pyc ติดมาด้วย ให้แน่ใจว่าลบไฟล์นั้นออกจากโฟลเดอร์ด้วย เพราะต้องการดึงแค่เฉพาะไฟล์รูปภาพเท่านั้น

เราจะสร้าง test set โดยสุ่มดอกไม้แต่ละชนิดมาจำนวน 100 ดอก มาใส่โฟลเดอร์ test และสร้างไฟล์ test.csv กำกับภาพและชนิด

import os
import random
import pandas as pd

path = 'train'
clsList = os.listdir(path)
testList = []

print('Original number of images per class:')
for i in clsList:
    curDir = os.listdir(path+'/'+i)
    print( '- ' + str(i) + ' = ' + str(len(curDir)) )

print('\nPick 50 images per class for test...')
for i in clsList:
    curDir = os.listdir(path+'/'+i)
    rd = random.sample(curDir, k=100)
    for j in range(len(rd)):
        os.rename(path+'/'+i+'/'+rd[j],'test/'+rd[j])
        testList.append([rd[j],i])
        
print('\nRemained number of images per class:')
for i in clsList:
    curDir = os.listdir(path+'/'+i)
    print( '- ' + str(i) + ' = ' + str(len(curDir)) )
    
testDf = pd.DataFrame(testList, columns = ['image', 'category']) 
shuffleDf = testDf.sample(frac=1).reset_index(drop=True)
shuffleDf.to_csv(r'test.csv', index = None, header=True)

ผลลัพธ์

Original number of images per class:
- daisy = 769
- dandelion = 1055
- rose = 784
- sunflower = 734
- tulip = 984

Pick 100 images per class for test...

Remained number of images per class:
- daisy = 669
- dandelion = 955
- rose = 684
- sunflower = 634
- tulip = 884

2. Build a model

from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Conv2D, MaxPooling2D
from tensorflow.keras.layers import Activation, Dropout, Flatten, Dense

model = Sequential()
model.add(Conv2D(32, (3, 3), input_shape=(150, 150, 3)))
model.add(Activation('relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))

model.add(Conv2D(32, (3, 3)))
model.add(Activation('relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))

model.add(Conv2D(64, (3, 3)))
model.add(Activation('relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))

model.add(Flatten()) 
model.add(Dense(64))
model.add(Activation('relu'))
model.add(Dropout(0.5))
model.add(Dense(5))
model.add(Activation('softmax'))

model.summary()

3. Pre-process image

from keras.preprocessing.image import ImageDataGenerator

train_datagen = ImageDataGenerator(rescale=1./255) 

train_generator = train_datagen.flow_from_directory('train/', 
                                                 target_size=(150,150),
                                                 color_mode='rgb',
                                                 batch_size=32,
                                                 class_mode='categorical',
                                                 shuffle=True)

4. Train data with model

import time 

startTime = time.time()

step_size_train = train_generator.n//train_generator.batch_size

history = model.fit_generator(generator=train_generator,
                   steps_per_epoch=step_size_train,
                   epochs=10)

endTime = time.time()

5. Show train result

5.1 Save model

model.save_weights('my-model_save-wieghts.h5')
model.save('my-model_save.h5')

5.2 Save history

import pandas as pd

histDf = pd.DataFrame(history.history) 
histCsv = 'my-model_history.csv'
with open(histCsv, mode='w') as f:
    histDf.to_csv(f)

5.3 Show time used

print("Strat time = "+str(startTime))
print("End time = "+str(endTime))
print("Use time = "+str(endTime-startTime))

5.4 Plot train accuracy and loss

import numpy as np
import matplotlib.pyplot as plt
%matplotlib inline 

acc = history.history['accuracy']
loss = history.history['loss']
epochs = range(1, len(acc) + 1)

#Train and validation accuracy
plt.style.use('default')
plt.grid(True)
plt.xticks(np.arange(0, len(acc)+1, 1.0))
plt.plot(epochs, acc, 'b', label='Training Accurarcy')
plt.title('Training Accuracy (My Model)')
plt.xlabel("Epoch #")
plt.ylabel("Accuracy")
plt.legend()
plt.grid(True)
plt.figure()
plt.plot(epochs, loss, 'r', label='Training Loss')
plt.grid(True)
plt.xticks(np.arange(0, len(acc)+1, 1.0))
plt.title('Training Loss (My Model)')
plt.xlabel("Epoch #")
plt.ylabel("Loss")
plt.legend()
plt.show()

6. Prepare test data

6.1 Read test csv

6.2 [x] Load image

6.3 [x] Convert image to array

6.4 [x] Put image array in list

6.5 [x] Convert list to numpy array

import pandas as pd
import numpy as np
from keras.preprocessing.image import load_img
from keras.preprocessing.image import img_to_array

testCsv = pd.read_csv('test.csv')
testArrListX = [img_to_array(load_img('test/'+imgName, target_size=(150,150))) for imgName in testCsv.image]
x_test = np.array(testArrListX)/255.0

6.6 [y] Transform label to numpy array 0 to n-1 (LabelEncoder)

6.7 [y] Convert list to one-hot array

testY = testCsv['category'].values

from sklearn.preprocessing import LabelEncoder
le = LabelEncoder()
y_test = le.fit_transform(testY)

from keras.utils import to_categorical
y_test_onehot = to_categorical(y_test)

7. Predict test data

y_pred_onehot = model.predict(x_test)

8. Show test result

8.1 Evaluate model

model.evaluate(x_test, y_test_onehot)

8.2 Convert one-hot to list

y_pred = []
for i in range(len(y_pred_onehot)):
    y_pred.append(np.argmax(y_pred_onehot[i]))

หรือ

y_pred = [np.argmax(y_pred_onehot[i]) for i in range(len(y_pred_onehot))]

8.3 Save prediction

testDict = {'image': testCsv.image, 'y_label': testY, 'y_true': testY_le, 'y_pred': y_pred}
dfPred = pd.DataFrame(testDict)
exportPred = dfPred.to_csv(r'prediction.csv', index = None, header=True)

8.4 Plot confusion matrix

from sklearn.metrics import confusion_matrix
cfArr = confusion_matrix(y_test, y_pred)

import seaborn as sn
dfCm = pd.DataFrame(cfArr, index = ['daisy', 'dandelion', 'rose', 'sunflower', 'tulip'], columns = ['daisy', 'dandelion', 'rose', 'sunflower', 'tulip'])
plt.figure()
sn.heatmap(dfCm, annot=True, cmap="Blues")
plt.title('Confusion Matrix (My Model)',fontsize = 15)
plt.show()

8.5 Print and save classification report

from sklearn.metrics import classification_report
print(classification_report(y_test, y_pred))

report = classification_report(y_test, y_pred, output_dict=True)
dfRp = pd.DataFrame(report).transpose()
exportRp = dfRp.to_csv(r'classification-report.csv', index = None, header=True)

โค้ดเต็ม

# ---------- 1. Prepare data ---------- 
import os
import random
import pandas as pd

path = 'train'
clsList = os.listdir(path)
testList = []

print('Original number of images per class:')
for i in clsList:
    curDir = os.listdir(path+'/'+i)
    print( '- ' + str(i) + ' = ' + str(len(curDir)) )

print('\nPick 50 images per class for test...')
for i in clsList:
    curDir = os.listdir(path+'/'+i)
    rd = random.sample(curDir, k=100)
    for j in range(len(rd)):
        os.rename(path+'/'+i+'/'+rd[j],'test/'+rd[j])
        testList.append([rd[j],i])
        
print('\nRemained number of images per class:')
for i in clsList:
    curDir = os.listdir(path+'/'+i)
    print( '- ' + str(i) + ' = ' + str(len(curDir)) )
    
testDf = pd.DataFrame(testList, columns = ['image', 'category']) 
shuffleDf = testDf.sample(frac=1).reset_index(drop=True)
shuffleDf.to_csv(r'test.csv', index = None, header=True)


# ---------- 2. Build a model ---------- 
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Conv2D, MaxPooling2D
from tensorflow.keras.layers import Activation, Dropout, Flatten, Dense

model = Sequential()
model.add(Conv2D(32, (3, 3), input_shape=(150, 150, 3)))
model.add(Activation('relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))

model.add(Conv2D(32, (3, 3)))
model.add(Activation('relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))

model.add(Conv2D(64, (3, 3)))
model.add(Activation('relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))

model.add(Flatten()) 
model.add(Dense(64))
model.add(Activation('relu'))
model.add(Dropout(0.5))
model.add(Dense(5))
model.add(Activation('softmax'))

model.summary()


# ---------- 3. Pre-process image ---------- 
from keras.preprocessing.image import ImageDataGenerator

train_datagen = ImageDataGenerator(rescale=1./255) 
train_generator = train_datagen.flow_from_directory('train/', 
                                                 target_size=(150,150),
                                                 color_mode='rgb',
                                                 batch_size=32,
                                                 class_mode='categorical',
                                                 shuffle=True)
             

# ---------- 4. Train data with model ---------- 
import time 

startTime = time.time()
step_size_train = train_generator.n//train_generator.batch_size
history = model.fit_generator(generator=train_generator,
                   steps_per_epoch=step_size_train,
                   epochs=10)
endTime = time.time()


# ---------- 5. Show train result ---------- 
# ---- Save model & history, Show time used ----
model.save_weights('my-model_save-wieghts.h5')
model.save('my-model_save.h5')

import pandas as pd
histDf = pd.DataFrame(history.history) 
histCsv = 'my-model_history.csv'
with open(histCsv, mode='w') as f:
    histDf.to_csv(f)

print("Strat time = "+str(startTime))
print("End time = "+str(endTime))
print("Use time = "+str(endTime-startTime))

# ---- Plot train accuracy and loss ----
import numpy as np
import matplotlib.pyplot as plt
%matplotlib inline 

acc = history.history['accuracy']
loss = history.history['loss']
epochs = range(1, len(acc) + 1)

#Train and validation accuracy
plt.style.use('default')
plt.grid(True)
plt.xticks(np.arange(0, len(acc)+1, 1.0))
plt.plot(epochs, acc, 'b', label='Training Accurarcy')
plt.title('Training Accuracy (My Model)')
plt.xlabel("Epoch #")
plt.ylabel("Accuracy")
plt.legend()
plt.grid(True)
plt.figure()
plt.plot(epochs, loss, 'r', label='Training Loss')
plt.grid(True)
plt.xticks(np.arange(0, len(acc)+1, 1.0))
plt.title('Training Loss (My Model)')
plt.xlabel("Epoch #")
plt.ylabel("Loss")
plt.legend()
plt.show()

# ---------- 6. Prepare test data ----------
# ---- Load image, convert to array, put in list, convert to numpy array, normalized ----
import pandas as pd
import numpy as np
from keras.preprocessing.image import load_img
from keras.preprocessing.image import img_to_array

testCsv = pd.read_csv('test.csv')
testArrListX = [img_to_array(load_img('test/'+imgName, target_size=(150,150))) for imgName in testCsv.image]
x_test = np.array(testArrListX)/255.0

# ---- Transform label to mumpy array, convert to one-hot ----
y_pred = [np.argmax(y_pred_onehot[i]) for i in range(len(y_pred_onehot))]

# ---------- 7. Predict test data ----------
y_pred_onehot = model.predict(x_test)

# ---------- 8. Show test result ----------
# ---- Evaluate model ----
model.evaluate(x_test, y_test_onehot)

# ---- Convert one-hot to list ----
y_pred = [np.argmax(y_pred_onehot[i]) for i in range(len(y_pred_onehot))]

# ---- Save prediction ----
testDict = {'image': testCsv.image, 'y_label': testY, 'y_true': testY_le, 'y_pred': y_pred}
dfPred = pd.DataFrame(testDict)
exportPred = dfPred.to_csv(r'prediction.csv', index = None, header=True)

# ---- Plot confusion matrix ----
from sklearn.metrics import confusion_matrix
cfArr = confusion_matrix(y_test, y_pred)

import seaborn as sn
dfCm = pd.DataFrame(cfArr, index = ['daisy', 'dandelion', 'rose', 'sunflower', 'tulip'], columns = ['daisy', 'dandelion', 'rose', 'sunflower', 'tulip'])
plt.figure()
sn.heatmap(dfCm, annot=True, cmap="Blues")
plt.title('Confusion Matrix (My Model)',fontsize = 15)
plt.show()

# ---- Print and save classification report ----
from sklearn.metrics import classification_report
print(classification_report(y_test, y_pred))

report = classification_report(y_test, y_pred, output_dict=True)
dfRp = pd.DataFrame(report).transpose()
exportRp = dfRp.to_csv(r'classification-report.csv', index = None, header=True)

โค้ดย่อ

import pandas as pd
import numpy as np
import seaborn as sn
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Conv2D, MaxPooling2D
from tensorflow.keras.layers import Activation, Dropout, Flatten, Dense
from keras.preprocessing.image import load_img, img_to_array, ImageDataGenerator
from sklearn.preprocessing import LabelEncoder
from keras.utils import to_categorical
from sklearn.metrics import confusion_matrix
from sklearn.metrics import classification_report

model = Sequential()
model.add(Conv2D(32, (3, 3), input_shape=(150, 150, 3)))
model.add(Activation('relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Conv2D(32, (3, 3)))
model.add(Activation('relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Conv2D(64, (3, 3)))
model.add(Activation('relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Flatten()) 
model.add(Dense(64))
model.add(Activation('relu'))
model.add(Dropout(0.5))
model.add(Dense(5))
model.add(Activation('softmax'))
model.summary()

train_datagen = ImageDataGenerator(rescale=1./255) 
train_generator = train_datagen.flow_from_directory('train/', 
                                                 target_size=(150,150),
                                                 color_mode='rgb',
                                                 batch_size=32,
                                                 class_mode='categorical',
                                                 shuffle=True)
             
step_size_train = train_generator.n//train_generator.batch_size
history = model.fit_generator(generator=train_generator,
                   steps_per_epoch=step_size_train,
                   epochs=10)

testCsv = pd.read_csv('test.csv')
testArrListX = [img_to_array(load_img('test/'+imgName, target_size=(150,150))) for imgName in testCsv.image]
x_test = np.array(testArrListX)/255.0
testY = testCsv['category'].values
le = LabelEncoder()
y_test = le.fit_transform(testY)
y_test_onehot = to_categorical(y_test)

model.evaluate(x_test, y_test_onehot)
y_pred_onehot = model.predict(x_test)

y_pred = [np.argmax(y_pred_onehot[i]) for i in range(len(y_pred_onehot))]
confusion_matrix(y_test, y_pred)
print(classification_report(y_test, y_pred))

Verzaru's Notes