TensorFlow实现VGG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]:

cfgs = {
    'vgg11': [64, 'M', 128, 'M', 256, 256, 'M', 512, 512, 'M', 512, 512, 'M'], 
    'vgg13': [64, 64, 'M', 128, 128, 'M', 256, 256, 'M', 512, 512, 'M', 512, 512, 'M'],
    'vgg16': [64, 64, 'M', 128, 128, 'M', 256, 256, 256, 'M', 512, 512, 512, 'M', 512, 512, 512, 'M'],
    'vgg19': [64, 64, 'M', 128, 128, 'M', 256, 256, 256, 256, 'M', 512, 512, 512, 512, 'M', 512, 512, 512, 512, 'M']
}

 

In [3]:

def make_feature(cfg):
    feature_layers = []
    for v in cfg:
        if v == 'M':
            feature_layers.append(keras.layers.MaxPool2D(pool_size=2, strides=2))
        else:
            feature_layers.append(keras.layers.Conv2D(v, kernel_size=3, padding='SAME', activation='selu'))
    return keras.Sequential(feature_layers, name='feature')

 

In [4]:

# 定义网络结构
def VGG(feature, height=224, width=224, classes=1000):
    input_image = keras.layers.Input(shape=(height, width, 3), dtype=tf.float32)
    x = feature(input_image)
    x = keras.layers.Flatten()(x)
    x = keras.layers.AlphaDropout(0.5)(x)
    x = keras.layers.Dense(2048, activation='selu')(x)
    x = keras.layers.AlphaDropout(0.5)(x)
    x = keras.layers.Dense(2048, activation='selu')(x)
    x = keras.layers.Dense(classes)(x)
    output = keras.layers.Softmax()(x)
    return keras.models.Model(inputs=input_image, outputs=output)

# 获取模型
def get_vgg_model(model_name='vgg16', height=224, width=224, classes=10):
    cfg = cfgs[model_name]
    return VGG(make_feature(cfg), height=height, width=width, classes=classes)

 

In [5]:

# 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'
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 [6]:

# 图片数据生成器
height = 224
width = 224
channels = 3
batch_size = 32
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 [7]:

# 配置网络
model = get_vgg_model(model_name='vgg16', height=height, width=width, classes=classes)
model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['acc'])
# 训练
history = model.fit(train_generator, steps_per_epoch=train_generator.samples//batch_size, epochs=10,
                    validation_data=valid_generator, validation_steps=valid_generator.samples//batch_size)

 

 

Epoch 1/10
34/34 [==============================] - 267s 8s/step - loss: 499.8480 - acc: 0.1144 - val_loss: 42.7053 - val_acc: 0.1055
Epoch 2/10
34/34 [==============================] - 286s 8s/step - loss: 25.4051 - acc: 0.1041 - val_loss: 22.3587 - val_acc: 0.1094
Epoch 3/10
34/34 [==============================] - 282s 8s/step - loss: 5.8247 - acc: 0.0976 - val_loss: 12.3608 - val_acc: 0.1055
Epoch 4/10
34/34 [==============================] - 293s 9s/step - loss: 3.8310 - acc: 0.1257 - val_loss: 11.0524 - val_acc: 0.1094
Epoch 5/10
34/34 [==============================] - 283s 8s/step - loss: 3.3107 - acc: 0.1135 - val_loss: 11.6788 - val_acc: 0.1094
Epoch 6/10
34/34 [==============================] - 284s 8s/step - loss: 3.2179 - acc: 0.1004 - val_loss: 9.5298 - val_acc: 0.1055
Epoch 7/10
34/34 [==============================] - 278s 8s/step - loss: 3.1169 - acc: 0.0891 - val_loss: 7.9141 - val_acc: 0.1055
Epoch 8/10
34/34 [==============================] - 276s 8s/step - loss: 3.0021 - acc: 0.1069 - val_loss: 7.8924 - val_acc: 0.1055
Epoch 9/10
34/34 [==============================] - 271s 8s/step - loss: 3.0272 - acc: 0.0844 - val_loss: 6.5881 - val_acc: 0.0977
Epoch 10/10
34/34 [==============================] - 268s 8s/step - loss: 2.9080 - acc: 0.0910 - val_loss: 6.9141 - val_acc: 0.1055

 

In [8]:

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

 

 

 

In [10]:

model.evaluate(valid_generator)   # 效果很差，用官方的 keras.applications.VGG19 进行训练，也是一样的结果

 

 

9/9 [==============================] - 16s 2s/step - loss: 6.7439 - acc: 0.0993

Out[10]:

[6.7438836097717285, 0.0992647036910057]