TensorFlow卷积神经网络识别10-monkey-species

 

In [1]:

from tensorflow import keras
import tensorflow as tf
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt

 

In [2]:

# 文件下载地址  https://www.kaggle.com/datasets/slothkong/10-monkey-species
train_dir = './10-monkey-species/training/training'
valid_dir = './10-monkey-species/validation/validation'
label_file = './10-monkey-species/monkey_labels.txt'

 

In [3]:

df = pd.read_csv(label_file, header=0)
df

	Label	Latin Name	Common Name	Train Images	Validation Images
0	n0	alouatta_palliata\t	mantled_howler	131	26
1	n1	erythrocebus_patas\t	patas_monkey	139	28
2	n2	cacajao_calvus\t	bald_uakari	137	27
3	n3	macaca_fuscata\t	japanese_macaque	152	30
4	n4	cebuella_pygmea\t	pygmy_marmoset	131	26
5	n5	cebus_capucinus\t	white_headed_capuchin	141	28
6	n6	mico_argentatus\t	silvery_marmoset	132	26
7	n7	saimiri_sciureus\t	common_squirrel_monkey	142	28
8	n8	aotus_nigriceps\t	black_headed_night_monkey	133	27
9	n9	trachypithecus_johnii	nilgiri_langur	132	26

 

In [17]:

# 图片数据生成器
height = 128
width = 128
channels = 3
batch_size = 32
num_classes = 10
train_datagen = keras.preprocessing.image.ImageDataGenerator(
    rescale = 1. / 255,    # 归一化&浮点数
    rotation_range = 40,   # 随机旋转 0~40°之间
    width_shift_range = 0.2, # 随机水平移动
    height_shift_range = 0.2, # 随机垂直移动
    shear_range = 0.2, # 随机裁剪比例
    zoom_range = 0.2, # 随机缩放比例
    horizontal_flip = True,  # 随机水平翻转
    vertical_flip = True, # 随机垂直翻转
    fill_mode = 'nearest', # 填充模式
)
train_generator = train_datagen.flow_from_directory(train_dir, target_size=(height, width),
                                                    batch_size=batch_size, shuffle=True, class_mode='categorical')

valid_datagen = keras.preprocessing.image.ImageDataGenerator(
    rescale = 1. / 255,    # 归一化&浮点数
)
valid_generator = valid_datagen.flow_from_directory(valid_dir, target_size=(height, width),
                                                    batch_size=batch_size, shuffle=False, class_mode='categorical')

 

 

Found 1098 images belonging to 10 classes.
Found 272 images belonging to 10 classes.

 

In [5]:

x, y = train_generator.next()   # 可以看出 y是经过one hot编码的
print(x.shape, y.shape)

 

 

(32, 128, 128, 3) (32, 10)

 

In [6]:

model = tf.keras.models.Sequential()
# 卷积 # input_shape 输入参数为 (height, width, channels)
model.add(tf.keras.layers.Conv2D(filters = 32, kernel_size = 3, padding = 'same', activation = 'selu', input_shape = (height, width, 3)))
model.add(tf.keras.layers.Conv2D(filters = 32, kernel_size = 3, padding = 'same', activation = 'selu'))    # 卷积
model.add(tf.keras.layers.MaxPool2D())    # 池化
model.add(tf.keras.layers.Conv2D(filters = 64, kernel_size = 3, padding = 'same', activation = 'selu'))    # 卷积
model.add(tf.keras.layers.Conv2D(filters = 64, kernel_size = 3, padding = 'same', activation = 'selu'))    # 卷积
model.add(tf.keras.layers.MaxPool2D())    # 池化
model.add(tf.keras.layers.Conv2D(filters = 128, kernel_size = 3, padding = 'same', activation = 'selu'))   # 卷积
model.add(tf.keras.layers.Conv2D(filters = 128, kernel_size = 3, padding = 'same', activation = 'selu'))    # 卷积
model.add(tf.keras.layers.MaxPool2D())    # 池化
model.add(tf.keras.layers.Conv2D(filters = 256, kernel_size = 3, padding = 'same', activation = 'selu'))   # 卷积
model.add(tf.keras.layers.Conv2D(filters = 256, kernel_size = 3, padding = 'same', activation = 'selu'))    # 卷积
model.add(tf.keras.layers.MaxPool2D())    # 池化

# 输入输出shape: 具体而言，是将一个维度大于或等于3的高维矩阵，“压扁”为一个二维矩阵。即保留第一个维度（如：batch的个数），
# 然后将剩下维度的值相乘为“压扁”矩阵的第二个维度。如输入是(None, 32，32，3)，则输出是(None, 3072)
model.add(tf.keras.layers.Flatten())   # Flatten层用来将输入“压平”，常用在从卷积层到全连接层的过渡。Flatten不影响batch的大小。
model.add(tf.keras.layers.Dense(2048, activation='selu'))
model.add(tf.keras.layers.AlphaDropout(0.3))
model.add(tf.keras.layers.Dense(1024, activation='selu'))
model.add(tf.keras.layers.AlphaDropout(0.25))
model.add(tf.keras.layers.Dense(512, activation='selu'))
model.add(tf.keras.layers.Dense(128, activation='selu'))
model.add(tf.keras.layers.Dense(10, activation='softmax'))

# 配置网络
model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['acc'])

 

In [7]:

# 训练
train_num = train_generator.samples
valid_num = valid_generator.samples
history = model.fit(train_generator, steps_per_epoch=train_num//batch_size, epochs=10,
                    validation_data=valid_generator, validation_steps=valid_num//batch_size)

 

 

Epoch 1/10
34/34 [==============================] - 42s 1s/step - loss: 18.0979 - acc: 0.1285 - val_loss: 4.8664 - val_acc: 0.1562
Epoch 2/10
34/34 [==============================] - 33s 957ms/step - loss: 2.6509 - acc: 0.1754 - val_loss: 2.7836 - val_acc: 0.2188
Epoch 3/10
34/34 [==============================] - 31s 921ms/step - loss: 2.2543 - acc: 0.2073 - val_loss: 3.4844 - val_acc: 0.1172
Epoch 4/10
34/34 [==============================] - 31s 900ms/step - loss: 2.0929 - acc: 0.2514 - val_loss: 2.6293 - val_acc: 0.2617
Epoch 5/10
34/34 [==============================] - 31s 917ms/step - loss: 2.0205 - acc: 0.2711 - val_loss: 2.4111 - val_acc: 0.2461
Epoch 6/10
34/34 [==============================] - 31s 919ms/step - loss: 1.9451 - acc: 0.3002 - val_loss: 2.5277 - val_acc: 0.2852
Epoch 7/10
34/34 [==============================] - 32s 924ms/step - loss: 1.8503 - acc: 0.3283 - val_loss: 2.5161 - val_acc: 0.3125
Epoch 8/10
34/34 [==============================] - 30s 882ms/step - loss: 1.8364 - acc: 0.3386 - val_loss: 2.5492 - val_acc: 0.2656
Epoch 9/10
34/34 [==============================] - 30s 867ms/step - loss: 1.7762 - acc: 0.3940 - val_loss: 1.8829 - val_acc: 0.3750
Epoch 10/10
34/34 [==============================] - 30s 873ms/step - loss: 1.7863 - acc: 0.3687 - val_loss: 2.1468 - val_acc: 0.3359

 

In [16]:

pd.DataFrame(history.history).plot(figsize=(8, 5))
plt.grid()
plt.gca().set_ylim(0, 10)
plt.show()

 

 

 

In [13]:

model.evaluate(valid_generator)   # 可以看出效果很差，还是需要用ResNet50这样的复杂网络才可以

 

 

9/9 [==============================] - 3s 297ms/step - loss: 2.0539 - acc: 0.3640

Out[13]:

[2.053945541381836, 0.3639705777168274]