一、準備訓練數據

下載數據集

validation驗證集
train訓練集

數據集結構如下：在這里插入圖片描述
將數據集解壓到自己選擇的目錄下就行
最后的結構效果如下：

二、構建模型

ImageDataGenerator

真實數據中，往往圖片尺寸大小不一，需要裁剪成一樣大小，一般為正方形
數據量比較大，不能一下子全部裝入內存中
經常需要進行修改參數，比如輸出的尺寸，增補圖像拉伸等

from tensorflow import keras
import tensorflow as tf
import matplotlib.pyplot as plt
import numpy as np
from tensorflow.keras.optimizers import RMSprop
from tensorflow.keras.preprocessing.image import ImageDataGenerator

#創建兩個數據生成器，指定scaling范圍為0-1
train_datagen = ImageDataGenerator(rescale=1/255)
validation_datagen = ImageDataGenerator(rescale=1/255)#將train_datagen數據生成器指向數據集所在文件夾
train_generator = train_datagen.flow_from_directory(r'.\images\training',#訓練集所在文件夾target_size=(300,300),#指定輸出尺寸batch_size=32,class_mode='binary'#指定二分類
)
#Found 1027 images belonging to 2 classes.#將validation_datagen數據生成器指向數據集所在文件夾
validation_generator = validation_datagen.flow_from_directory(r'.\images\validation',#驗證集所在文件夾target_size=(300,300),#指定輸出尺寸batch_size=32,class_mode='binary'#指定二分類
)
#Found 256 images belonging to 2 classes.

三、訓練模型

model = tf.keras.models.Sequential()model.add(tf.keras.layers.Conv2D(64,(3,3),activation='relu',input_shape=(300,300,3)))
model.add(tf.keras.layers.MaxPooling2D(2,2))
model.add(tf.keras.layers.Conv2D(32,(3,3),activation='relu',input_shape=(28,28,1)))
model.add(tf.keras.layers.MaxPooling2D(2,2))
model.add(tf.keras.layers.Flatten())
model.add(tf.keras.layers.Dense(128,activation='relu'))
model.add(tf.keras.layers.Dense(1,activation='sigmoid'))model.compile(optimizer=RMSprop(lr=0.001),loss="binary_crossentropy",metrics=['accuracy'])model.fit(train_generator,epochs=20,validation_data = validation_generator)"""
Epoch 1/20
8/8 [==============================] - 8s 996ms/step - loss: 7.9712 - acc: 0.5000
33/33 [==============================] - 136s 4s/step - loss: 7.4488 - acc: 0.5151 - val_loss: 7.9712 - val_acc: 0.5000
Epoch 2/20
8/8 [==============================] - 8s 968ms/step - loss: 7.9712 - acc: 0.5000
33/33 [==============================] - 136s 4s/step - loss: 7.8380 - acc: 0.5131 - val_loss: 7.9712 - val_acc: 0.5000
Epoch 3/20
8/8 [==============================] - 8s 1s/step - loss: 7.9712 - acc: 0.5000
33/33 [==============================] - 141s 4s/step - loss: 7.6964 - acc: 0.5131 - val_loss: 7.9712 - val_acc: 0.5000
Epoch 4/20
8/8 [==============================] - 8s 1s/step - loss: 7.9712 - acc: 0.5000
33/33 [==============================] - 143s 4s/step - loss: 7.8380 - acc: 0.5131 - val_loss: 7.9712 - val_acc: 0.5000
Epoch 5/20
8/8 [==============================] - 8s 1s/step - loss: 7.9712 - acc: 0.5000
33/33 [==============================] - 145s 4s/step - loss: 7.5547 - acc: 0.5131 - val_loss: 7.9712 - val_acc: 0.5000
Epoch 6/20
8/8 [==============================] - 8s 1s/step - loss: 7.9712 - acc: 0.5000
33/33 [==============================] - 144s 4s/step - loss: 7.8380 - acc: 0.5131 - val_loss: 7.9712 - val_acc: 0.5000
Epoch 7/20
8/8 [==============================] - 8s 1s/step - loss: 7.9712 - acc: 0.5000
33/33 [==============================] - 146s 4s/step - loss: 7.5547 - acc: 0.5131 - val_loss: 7.9712 - val_acc: 0.5000
Epoch 8/20
8/8 [==============================] - 8s 1s/step - loss: 7.9712 - acc: 0.5000
33/33 [==============================] - 144s 4s/step - loss: 7.8380 - acc: 0.5131 - val_loss: 7.9712 - val_acc: 0.5000
Epoch 9/20
8/8 [==============================] - 8s 1s/step - loss: 7.9712 - acc: 0.5000
33/33 [==============================] - 144s 4s/step - loss: 7.6964 - acc: 0.5131 - val_loss: 7.9712 - val_acc: 0.5000
Epoch 10/20
8/8 [==============================] - 8s 1s/step - loss: 7.9712 - acc: 0.5000
33/33 [==============================] - 144s 4s/step - loss: 7.8380 - acc: 0.5131 - val_loss: 7.9712 - val_acc: 0.5000
Epoch 11/20
8/8 [==============================] - 8s 1s/step - loss: 7.9712 - acc: 0.5000
33/33 [==============================] - 143s 4s/step - loss: 7.6964 - acc: 0.5131 - val_loss: 7.9712 - val_acc: 0.5000
Epoch 12/20
8/8 [==============================] - 8s 982ms/step - loss: 7.9712 - acc: 0.5000
33/33 [==============================] - 138s 4s/step - loss: 7.6964 - acc: 0.5131 - val_loss: 7.9712 - val_acc: 0.5000
Epoch 13/20
8/8 [==============================] - 8s 968ms/step - loss: 7.9712 - acc: 0.5000
33/33 [==============================] - 135s 4s/step - loss: 7.6964 - acc: 0.5131 - val_loss: 7.9712 - val_acc: 0.5000
Epoch 14/20
8/8 [==============================] - 8s 974ms/step - loss: 7.9712 - acc: 0.5000
33/33 [==============================] - 135s 4s/step - loss: 7.8380 - acc: 0.5131 - val_loss: 7.9712 - val_acc: 0.5000
Epoch 15/20
8/8 [==============================] - 8s 971ms/step - loss: 7.9712 - acc: 0.5000
33/33 [==============================] - 135s 4s/step - loss: 7.6964 - acc: 0.5131 - val_loss: 7.9712 - val_acc: 0.5000
Epoch 16/20
8/8 [==============================] - 8s 972ms/step - loss: 7.9712 - acc: 0.5000
33/33 [==============================] - 135s 4s/step - loss: 7.8380 - acc: 0.5131 - val_loss: 7.9712 - val_acc: 0.5000
Epoch 17/20
8/8 [==============================] - 8s 983ms/step - loss: 7.9712 - acc: 0.5000
33/33 [==============================] - 135s 4s/step - loss: 7.8380 - acc: 0.5131 - val_loss: 7.9712 - val_acc: 0.5000
Epoch 18/20
8/8 [==============================] - 8s 976ms/step - loss: 7.9712 - acc: 0.5000
33/33 [==============================] - 136s 4s/step - loss: 7.8380 - acc: 0.5131 - val_loss: 7.9712 - val_acc: 0.5000
Epoch 19/20
8/8 [==============================] - 8s 969ms/step - loss: 7.9712 - acc: 0.5000
33/33 [==============================] - 134s 4s/step - loss: 7.8380 - acc: 0.5131 - val_loss: 7.9712 - val_acc: 0.5000
Epoch 20/20
8/8 [==============================] - 9s 1s/step - loss: 7.9712 - acc: 0.5000
33/33 [==============================] - 137s 4s/step - loss: 7.5547 - acc: 0.5131 - val_loss: 7.9712 - val_acc: 0.5000
"""

四、優化參數

import os
import tensorflow as tf
from tensorflow.keras.preprocessing.image import ImageDataGenerator
from tensorflow.keras.optimizers import RMSprop
from kerastuner.tuners import Hyperband
from kerastuner.engine.hyperparameters import HyperParameters#創建兩個數據生成器，指定scaling范圍為0-1
train_datagen = ImageDataGenerator(rescale=1/255)
validation_datagen = ImageDataGenerator(rescale=1/255)#將train_datagen數據生成器指向數據集所在文件夾
train_generator = train_datagen.flow_from_directory(r'.\images\training',#訓練集所在文件夾target_size=(150,150),#指定輸出尺寸batch_size=32,class_mode='binary'#指定二分類
)#將validation_datagen數據生成器指向數據集所在文件夾
validation_generator = validation_datagen.flow_from_directory(r'.\images\validation',#驗證集所在文件夾target_size=(150,150),#指定輸出尺寸batch_size=32,class_mode='binary'#指定二分類
)hp = HyperParameters()def build_model(hp):model = tf.keras.models.Sequential()model.add(tf.keras.layers.Conv2D(hp.Choice('num_filters_layer0',values=[16,64],default=16),(3,3),activation='relu',input_shape=(150,150,3))),model.add(tf.keras.layers.MaxPooling2D(2,2)),for i in range(hp.Int("num_conv_layers",1,3)):model.add(tf.keras.layers.Conv2D(hp.Choice(f'num_filters_layer{i}',values=[16,64],default=16),(3,3),activation='relu')),model.add(tf.keras.layers.MaxPooling2D(2,2)),model.add(tf.keras.layers.Conv2D(64,(3,3),activation='relu')),model.add(tf.keras.layers.MaxPooling2D(2,2)),model.add(tf.keras.layers.Flatten()),model.add(tf.keras.layers.Dense(hp.Int("hidden_units",128,512,step=32),activation='relu')),model.add(tf.keras.layers.Dense(1,activation='sigmoid'))#是否，一個神經元就行model.compile(optimizer=RMSprop(lr=0.001),loss="binary_crossentropy",metrics=['accuracy'])return modeltuner = Hyperband(#將訓練好的參數存放起來build_model,objective='val_acc',max_epochs=15,directory='horse_human_params',hyperparameters=hp,project_name='my_horse_human_project'
)tuner.search(train_generator,epochs=10,validation_data=validation_generator)
"""
Trial 13 Complete [00h 02m 13s]
val_acc: 0.87109375Best val_acc So Far: 0.890625
Total elapsed time: 00h 14m 22sSearch: Running Trial #14Value             |Best Value So Far |Hyperparameter
16                |64                |num_filters_layer0
5                 |2                 |tuner/epochs
2                 |0                 |tuner/initial_epoch
2                 |2                 |tuner/bracket
1                 |0                 |tuner/round
0000              |None              |tuner/trial_id
"""

在這里插入圖片描述

best_hps = tuner.get_best_hyperparameters(1)[0]#根據最優，把模型構建出來
print(best_hps.values)
"""
{'num_filters_layer0': 64, 'num_conv_layers': 2, 'hidden_units': 256, 'num_filters_layer1': 16, 'num_filters_layer2': 64, 'tuner/epochs': 2, 'tuner/initial_epoch': 0, 'tuner/bracket': 2, 'tuner/round': 0}
"""

model = tuner.hypermodel.build(best_hps)
model.summary()
"""
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
conv2d_1 (Conv2D)            (None, 148, 148, 64)      1792      
_________________________________________________________________
max_pooling2d_1 (MaxPooling2 (None, 74, 74, 64)        0         
_________________________________________________________________
conv2d_2 (Conv2D)            (None, 72, 72, 64)        36928     
_________________________________________________________________
max_pooling2d_2 (MaxPooling2 (None, 36, 36, 64)        0         
_________________________________________________________________
conv2d_3 (Conv2D)            (None, 34, 34, 16)        9232      
_________________________________________________________________
max_pooling2d_3 (MaxPooling2 (None, 17, 17, 16)        0         
_________________________________________________________________
conv2d_4 (Conv2D)            (None, 15, 15, 64)        9280      
_________________________________________________________________
max_pooling2d_4 (MaxPooling2 (None, 7, 7, 64)          0         
_________________________________________________________________
flatten (Flatten)            (None, 3136)              0         
_________________________________________________________________
dense (Dense)                (None, 256)               803072    
_________________________________________________________________
dense_1 (Dense)              (None, 1)                 257       
=================================================================
Total params: 860,561
Trainable params: 860,561
Non-trainable params: 0
_________________________________________________________________
"""