Keras预训练模型综合对比

model	convergence	acc	val_acc	loss	val_loss	size	time
VGG16	42	98.94%	98.94%	0.0333	0.0097	88.17MB	10m12s
VGG19	38	98.39%	100.00%	0.0443	0.0067	100.93MB	9m48s
ResNet50	42	99.14%	99.17%	0.0215	0.0134	188.21MB	11m31s
InceptionV3	25	98.29%	100.00%	0.0420	0.0059	211.51MB	9m12s
InceptionResNetV2	46	98.89%	99.79%	0.0284	0.0069	304.05MB	30m26s
Xception	20	97.73%	99.79%	0.0418	0.0090	279.75MB	11m36s
MobileNet	9	95.51%	98.96%	0.1255	0.9551	61.43MB	2m5s
MobileNetV2	17	96.72%	98.96%	0.0870	0.0211	70.07MB	5m0s
DenseNet121	32	98.39%	99.78%	0.0450	0.0084	76.31MB	10m22s
DenseNet169	46	98.99%	99.79%	0.0261	0.0081	128.49MB	19m41s
DenseNet201	20	98.24%	100.00%	0.0461	0.0058	162.53MB	10m13s
NASNetMobile	13	96.62%	99.17%	0.0843	0.0248	67.88MB	9m50s
NASNetLarge	4	98.28%	99.80%	0.0471	0.0045	801.81MB	9m40s

1 说明

• 参考文献：Keras 中文文档 https://keras.io/zh/
• 数据集：https://www.kaggle.com/c/dogs-vs-cats/data 共用2500张，其中80%用于训练（2000张），20%用于验证（500张）
• 优化器为SGD，学习率0.0001，动量0.9
• 批量大小32，NASNetLarge网络为26
• 回调函数ModelCheckpoint，只保存权重
• 回调函数TensorBoard，保存模型图和模型数据
• 回调函数ReduceLROnPlateau，val_loss停止优化10代后学习率降低10倍
• 回调函数EarlyStopping，val_loss停止优化30代后停止训练
• 无数据增强，输入图片大小与网络默认一致
• 配置：4×Intel Xeon E5-2650，2×Tesla P100，Ubuntu 16.04，CUDA 8.0，cuDNN 5，TensorFlow-gpu 1.4.0，Keras 2.2.4
• 测试网络：VGG16、VGG19、ResNet50、InceptionV3、InceptionResNetV2、Xception、MobileNet、MobileNetV2、DenseNet121、DenseNet169、DenseNet201、NASNetMobile、NASNetLarge

2 示例代码

2.1 notebooks及训练结果

https://download.csdn.net/download/lly1122334/11112642

2.2 目录结构

2.3 VGG16示例

In [1]:

import tensorflow as tf
from keras import optimizers
from keras import applications
from keras.utils import plot_model
from keras.utils import multi_gpu_model
from keras.models import Sequential, Model
from keras.layers import Dropout, Flatten, Dense
from keras.preprocessing.image import ImageDataGenerator
from keras.callbacks import ModelCheckpoint, EarlyStopping, TensorBoard, ReduceLROnPlateau

# Parameter
epochs = 1000
batch_size = 32

img_width, img_height = 256, 256

nb_gpus = 2
nb_train_samples = 2000
nb_validation_samples = 500

train_dir = 'data/train'
validation_dir = 'data/validation'

log_dir = './logs/vgg16'
weight_dir = './weight/vgg16.h5'
pic_model_dir = './model/vgg16.png'

Using TensorFlow backend.

In [2]:

with tf.device('/cpu:0'):

    # Pre-training network
    base_model = applications.VGG16(
        weights="imagenet",
        include_top=False,
        input_shape=(img_width, img_height, 3))

    # Custom network
    top_model = Sequential()
    top_model.add(Flatten(input_shape=base_model.output_shape[1:]))
    top_model.add(Dense(256, activation='relu'))
    top_model.add(Dropout(0.5))
    top_model.add(Dense(1, activation='sigmoid'))

    # Combined network
    model = Model(
        inputs=base_model.input, outputs=top_model(base_model.output))

    # Plot model
    plot_model(model, show_shapes=True, to_file=pic_model_dir)

model文件夹下有一张模型结构图

In [3]:

# Multiple GPU
parallel_model = multi_gpu_model(model, gpus=nb_gpus)

parallel_model.compile(
    optimizer=optimizers.SGD(lr=0.0001, momentum=0.9),
    loss='binary_crossentropy',
    metrics=['accuracy'])

In [4]:

# Data generator
train_datagen = ImageDataGenerator(rescale=1. / 255)
validation_datagen = ImageDataGenerator(rescale=1. / 255)

train_generator = train_datagen.flow_from_directory(
    train_dir,
    target_size=(img_height, img_width),
    batch_size=batch_size,
    class_mode='binary')

validation_generator = validation_datagen.flow_from_directory(
    validation_dir,
    target_size=(img_height, img_width),
    batch_size=batch_size,
    class_mode='binary',
    shuffle=False)

Found 22500 images belonging to 2 classes.
Found 2500 images belonging to 2 classes.

In [5]:

# Callbacks
checkpointer = ModelCheckpoint(
    filepath=weight_dir,
    verbose=1,
    save_best_only=True,
    save_weights_only=True)

tensorBoard = TensorBoard(
    log_dir=log_dir,
    write_graph=True,
    write_images=True)

reduceLROnPlateau = ReduceLROnPlateau(factor=0.1, patience=10, verbose=1)

earlystopping = EarlyStopping(patience=30, verbose=1)

In [6]:

# Fit
history = parallel_model.fit_generator(
    train_generator,
    steps_per_epoch=nb_train_samples // batch_size,
    epochs=epochs,
    verbose=1,
    validation_data=validation_generator,
    validation_steps=nb_validation_samples // batch_size,
    callbacks=[checkpointer, tensorBoard, reduceLROnPlateau, earlystopping])

Epoch 1/1000
62/62 [==============================] - 19s 303ms/step - loss: 0.6373 - acc: 0.6346 - val_loss: 0.5899 - val_acc: 0.6958

Epoch 00001: val_loss improved from inf to 0.58985, saving model to ./weight/vgg16.h5
......
Epoch 42/1000
62/62 [==============================] - 14s 232ms/step - loss: 0.0333 - acc: 0.9894 - val_loss: 0.0097 - val_acc: 0.9979


Epoch 00042: val_loss improved from 0.01013 to 0.00973, saving model to ./weight/vgg16.h5
......
Epoch 52/1000
62/62 [==============================] - 14s 225ms/step - loss: 0.0364 - acc: 0.9849 - val_loss: 0.0197 - val_acc: 0.9917

Epoch 00052: val_loss did not improve from 0.00973

Epoch 00052: ReduceLROnPlateau reducing learning rate to 9.999999747378752e-06.
......
Epoch 62/1000
62/62 [==============================] - 14s 230ms/step - loss: 0.0227 - acc: 0.9919 - val_loss: 0.0248 - val_acc: 0.9896

Epoch 00062: val_loss did not improve from 0.00973

Epoch 00062: ReduceLROnPlateau reducing learning rate to 9.999999747378752e-07.
......
Epoch 72/1000
62/62 [==============================] - 14s 229ms/step - loss: 0.0194 - acc: 0.9929 - val_loss: 0.0348 - val_acc: 0.9833

Epoch 00072: val_loss did not improve from 0.00973

Epoch 00072: ReduceLROnPlateau reducing learning rate to 9.999999974752428e-08.
Epoch 00072: early stopping

训练过程中启动tensorboard，监视训练过程

tensorboard --logdir=logs/vgg16

Tensorboard SCALARS

Tensorboard GRAPHS

In [7]:

import sys
import matplotlib.pyplot as plt

In [8]:

acc = history.history['acc']
val_acc = history.history['val_acc']
loss = history.history['loss']
val_loss = history.history['val_loss']
lr = history.history['lr']

plt.figure(1)
plt.figure(figsize=(8, 18), dpi=100)

#Plot accuracy
plt.subplot(311)
plt.plot(acc)
plt.plot(val_acc)
plt.title('Accuracy')
plt.xlabel('Epoch')
plt.ylabel('Accuracy')
plt.legend(['Train', 'Validation'], loc='best')
x = 0
maximum = -sys.maxsize
for y in val_acc:
    if y > maximum:
        maximum = y
        plt.text(x, y, round(y, 4), ha='center', va='bottom')
    x += 1

# Plot loss
plt.subplot(312)
plt.plot(loss)
plt.plot(val_loss)
plt.title('Loss')
plt.xlabel('Epoch')
plt.ylabel('Loss')
plt.legend(['Train', 'Validation'], loc='best')
x = 0
minimum = sys.maxsize
for y in val_loss:
    if y < minimum:
        minimum = y
        plt.text(x, y, round(y, 4), ha='center', va='bottom')
    x += 1

# Plot learning rate
plt.subplot(313)
plt.plot(lr)
plt.title('Learning Rate')
plt.xlabel('Epoch')
plt.ylabel('Learning Rate')
x = 0
minimum = sys.maxsize
for y in lr:
    if y < minimum:
        minimum = y
        plt.text(x, y, y, ha='center', va='bottom')
    x += 1

plt.show()

3 测试结果

3.1 VGG16

权重大小：88.17MB

Epoch 42/1000
62/62 [==============================] - 14s 232ms/step - loss: 0.0333 - acc: 0.9894 - val_loss: 0.0097 - val_acc: 0.9979


Epoch 00042: val_loss improved from 0.01013 to 0.00973, saving model to ./weight/vgg16.h5

3.2 VGG19

权重大小：100.93MB

Epoch 38/1000
62/62 [==============================] - 13s 216ms/step - loss: 0.0443 - acc: 0.9839 - val_loss: 0.0067 - val_acc: 1.0000

Epoch 00038: val_loss improved from 0.01237 to 0.00675, saving model to ./weight/vgg19.h5

3.3 ResNet50

权重大小：188.21MB

Epoch 42/1000

62/62 [==============================] - 16s 258ms/step - loss: 0.0215 - acc: 0.9914 - val_loss: 0.0134 - val_acc: 0.9917

Epoch 00042: val_loss improved from 0.01615 to 0.01336, saving model to ./weight/ResNet50.h5

3.4 InceptionV3

权重大小：211.51MB

Epoch 25/1000
62/62 [==============================] - 22s 350ms/step - loss: 0.0420 - acc: 0.9829 - val_loss: 0.0059 - val_acc: 1.0000

Epoch 00025: val_loss improved from 0.01020 to 0.00590, saving model to ./weight/InceptionV3.h5

3.5 InceptionResNetV2

权重大小：304.05MB

Epoch 46/1000
62/62 [==============================] - 39s 636ms/step - loss: 0.0284 - acc: 0.9889 - val_loss: 0.0069 - val_acc: 0.9979

Epoch 00046: val_loss improved from 0.00855 to 0.00692, saving model to ./weight/InceptionResNetV2.h5

3.6 Xception

权重大小：279.75MB

Epoch 20/1000
62/62 [==============================] - 34s 555ms/step - loss: 0.0474 - acc: 0.9773 - val_loss: 0.0090 - val_acc: 0.9979

Epoch 00020: val_loss improved from 0.01220 to 0.00898, saving model to ./weight/Xception.h5

3.7 MobileNet

权重大小：61.43MB

Epoch 9/1000
62/62 [==============================] - 14s 221ms/step - loss: 0.1255 - acc: 0.9551 - val_loss: 0.0217 - val_acc: 0.9896

Epoch 00009: val_loss improved from 0.03307 to 0.02166, saving model to ./weight/MobileNet.h5

3.8 MobileNetV2

权重大小：70.07MB

Epoch 17/1000
62/62 [==============================] - 17s 279ms/step - loss: 0.0870 - acc: 0.9672 - val_loss: 0.0211 - val_acc: 0.9896

Epoch 00017: val_loss improved from 0.03336 to 0.02112, saving model to ./weight/MobileNetV2.h5

3.9 DenseNet121

权重大小：76.31MB

Epoch 32/1000
62/62 [==============================] - 20s 317ms/step - loss: 0.0450 - acc: 0.9839 - val_loss: 0.0084 - val_acc: 0.9978

Epoch 00032: val_loss improved from 0.01282 to 0.00844, saving model to ./weight/DenseNet121.h5

3.10 DenseNet169

权重大小：128.49MB

Epoch 46/1000
62/62 [==============================] - 26s 415ms/step - loss: 0.0261 - acc: 0.9899 - val_loss: 0.0081 - val_acc: 0.9979

Epoch 00046: val_loss improved from 0.00868 to 0.00806, saving model to ./weight/DenseNet169.h5

3.11 DenseNet201

权重大小：162.53MB

Epoch 20/1000
62/62 [==============================] - 30s 485ms/step - loss: 0.0461 - acc: 0.9824 - val_loss: 0.0058 - val_acc: 1.0000

Epoch 00020: val_loss improved from 0.01274 to 0.00580, saving model to ./weight/DenseNet201.h5

3.12 NASNetMobile

权重大小：67.88MB

Epoch 13/1000
62/62 [==============================] - 45s 730ms/step - loss: 0.0843 - acc: 0.9662 - val_loss: 0.0248 - val_acc: 0.9917

Epoch 00013: val_loss improved from 0.03313 to 0.02478, saving model to ./weight/NASNetMobile.h5

3.13 NASNetLarge

权重大小：801.81MB

Epoch 4/1000
76/76 [==============================] - 144s 2s/step - loss: 0.0471 - acc: 0.9828 - val_loss: 0.0045 - val_acc: 0.9980

Epoch 00004: val_loss improved from 0.02069 to 0.00451, saving model to ./weight/NASNetLarge.h5