上个月发布了四篇文章，主要讲了深度学习中的“hello world”----mnist图像识别，以及卷积神经网络的原理详解，包括基本原理、自己手写CNN和paddlepaddle的源码解析。这篇主要跟大家讲讲如何用PaddlePaddle和Tensorflow做图像分类。所有程序都在我的github里，可以自行下载训练。
在卷积神经网络中，有五大经典模型，分别是：LeNet-5,AlexNet,GoogleNet,Vgg和ResNet。本文首先自己设计一个小型CNN网络结构来对图像进行分类，再了解一下LeNet-5网络结构对图像做分类，并用比较流行的Tensorflow框架和百度的PaddlePaddle实现LeNet-5网络结构，并对结果对比。
什么是图像分类
图像分类是根据图像的语义信息将不同类别图像区分开来，是计算机视觉中重要的基本问题，也是图像检测、图像分割、物体跟踪、行为分析等其他高层视觉任务的基础。图像分类在很多领域有广泛应用，包括安防领域的人脸识别和智能视频分析等，交通领域的交通场景识别，互联网领域基于内容的图像检索和相册自动归类，医学领域的图像识别等(引用自官网)
cifar-10数据集
CIFAR-10分类问题是机器学习领域的一个通用基准，由60000张32*32的RGB彩色图片构成，共10个分类。50000张用于训练集，10000张用于测试集。其问题是将32X32像素的RGB图像分类成10种类别:飞机，手机，鸟，猫，鹿，狗，青蛙，马，船和卡车。更多信息可以参考CIFAR-10和Alex Krizhevsky的演讲报告。常见的还有cifar-100，分类物体达到100类，以及ILSVRC比赛的100类。
自己设计CNN
了解CNN的基本网络结构后，首先自己设计一个简单的CNN网络结构对cifar-10数据进行分类。
网络结构
代码实现
1.网络结构：simple_cnn.py
#coding:utf-8
'''
Created by huxiaoman 2017.11.27
simple_cnn.py:自己设计的一个简单的cnn网络结构
''' import os
from PIL import Image
import numpy as np
import paddle.fluid as fluid
from paddle.trainer_config_helpers import * with_gpu = os.getenv('WITH_GPU', '') != '' def simple_cnn(img):
conv_pool_1 = paddle.networks.simple_img_conv_pool(
input=img,
filter_size=5,
num_filters=20,
num_channel=3,
pool_size=2,
pool_stride=2,
act=paddle.activation.Relu())
conv_pool_2 = paddle.networks.simple_img_conv_pool(
input=conv_pool_1,
filter_size=5,
num_filters=50,
num_channel=20,
pool_size=2,
pool_stride=2,
act=paddle.activation.Relu())
fc = paddle.layer.fc(
input=conv_pool_2, size=512, act=paddle.activation.Softmax())
2.训练程序：train_simple_cnn.py
#coding:utf-8
'''
Created by huxiaoman 2017.11.27
train_simple—_cnn.py:训练simple_cnn对cifar10数据集进行分类
'''
import sys, os import paddle.v2 as paddle
from simple_cnn import simple_cnn with_gpu = os.getenv('WITH_GPU', '') != '' def main():
datadim = 3 * 32 * 32
classdim = 10 # PaddlePaddle init
paddle.init(use_gpu=with_gpu, trainer_count=7) image = paddle.layer.data(
name="image", type=paddle.data_type.dense_vector(datadim)) # Add neural network config
# option 1. resnet
# net = resnet_cifar10(image, depth=32)
# option 2. vgg
net = simple_cnn(image) out = paddle.layer.fc(
input=net, size=classdim, act=paddle.activation.Softmax()) lbl = paddle.layer.data(
name="label", type=paddle.data_type.integer_value(classdim))
cost = paddle.layer.classification_cost(input=out, label=lbl) # Create parameters
parameters = paddle.parameters.create(cost) # Create optimizer
momentum_optimizer = paddle.optimizer.Momentum(
momentum=0.9,
regularization=paddle.optimizer.L2Regularization(rate=0.0002 * 128),
learning_rate=0.1 / 128.0,
learning_rate_decay_a=0.1,
learning_rate_decay_b=50000 * 100,
learning_rate_schedule='discexp') # End batch and end pass event handler
def event_handler(event):
if isinstance(event, paddle.event.EndIteration):
if event.batch_id % 100 == 0:
print "\nPass %d, Batch %d, Cost %f, %s" % (
event.pass_id, event.batch_id, event.cost, event.metrics)
else:
sys.stdout.write('.')
sys.stdout.flush()
if isinstance(event, paddle.event.EndPass):
# save parameters
with open('params_pass_%d.tar' % event.pass_id, 'w') as f:
parameters.to_tar(f) result = trainer.test(
reader=paddle.batch(
paddle.dataset.cifar.test10(), batch_size=128),
feeding={'image': 0,
'label': 1})
print "\nTest with Pass %d, %s" % (event.pass_id, result.metrics) # Create trainer
trainer = paddle.trainer.SGD(
cost=cost, parameters=parameters, update_equation=momentum_optimizer) # Save the inference topology to protobuf.
inference_topology = paddle.topology.Topology(layers=out)
with open("inference_topology.pkl", 'wb') as f:
inference_topology.serialize_for_inference(f) trainer.train(
reader=paddle.batch(
paddle.reader.shuffle(
paddle.dataset.cifar.train10(), buf_size=50000),
batch_size=128),
num_passes=200,
event_handler=event_handler,
feeding={'image': 0,
'label': 1}) # inference
from PIL import Image
import numpy as np
import os def load_image(file):
im = Image.open(file)
im = im.resize((32, 32), Image.ANTIALIAS)
im = np.array(im).astype(np.float32)
# The storage order of the loaded image is W(widht),
# H(height), C(channel). PaddlePaddle requires
# the CHW order, so transpose them.
im = im.transpose((2, 0, 1)) # CHW
# In the training phase, the channel order of CIFAR
# image is B(Blue), G(green), R(Red). But PIL open
# image in RGB mode. It must swap the channel order.
im = im[(2, 1, 0), :, :] # BGR
im = im.flatten()
im = im / 255.0
return im test_data = []
cur_dir = os.path.dirname(os.path.realpath(__file__))
test_data.append((load_image(cur_dir + '/image/dog.png'), )) # users can remove the comments and change the model name
# with open('params_pass_50.tar', 'r') as f:
# parameters = paddle.parameters.Parameters.from_tar(f) probs = paddle.infer(
output_layer=out, parameters=parameters, input=test_data)
lab = np.argsort(-probs) # probs and lab are the results of one batch data
print "Label of image/dog.png is: %d" % lab[0][0] if __name__ == '__main__':
main()
3.结果输出
I1128 21:44:30.218085 14733 Util.cpp:166] commandline: --use_gpu=True --trainer_count=7
[INFO 2017-11-28 21:44:35,874 layers.py:2539] output for __conv_pool_0___conv: c = 20, h = 28, w = 28, size = 15680
[INFO 2017-11-28 21:44:35,874 layers.py:2667] output for __conv_pool_0___pool: c = 20, h = 14, w = 14, size = 3920
[INFO 2017-11-28 21:44:35,875 layers.py:2539] output for __conv_pool_1___conv: c = 50, h = 10, w = 10, size = 5000
[INFO 2017-11-28 21:44:35,876 layers.py:2667] output for __conv_pool_1___pool: c = 50, h = 5, w = 5, size = 1250
I1128 21:44:35.881502 14733 MultiGradientMachine.cpp:99] numLogicalDevices=1 numThreads=7 numDevices=8
I1128 21:44:35.928449 14733 GradientMachine.cpp:85] Initing parameters..
I1128 21:44:36.056259 14733 GradientMachine.cpp:92] Init parameters done. Pass 0, Batch 0, Cost 2.302628, {'classification_error_evaluator': 0.9296875}
................................................................................
```
Pass 199, Batch 200, Cost 0.869726, {'classification_error_evaluator': 0.3671875}
...................................................................................................
Pass 199, Batch 300, Cost 0.801396, {'classification_error_evaluator': 0.3046875}
..........................................................................................I1128 23:21:39.443141 14733 MultiGradientMachine.cpp:99] numLogicalDevices=1 numThreads=7 numDevices=8 Test with Pass 199, {'classification_error_evaluator': 0.5248000025749207}
Label of image/dog.png is: 9
我开了7个线程，用了8个Tesla K80 GPU训练，batch_size = 128,迭代次数200次，耗时1h37min，错误分类率为0.5248，这个结果，emm，不算很高，我们可以把它作为一个baseline，后面对其进行调优。
LeNet-5网络结构
Lenet-5网络结构来源于Yan LeCun提出的,原文为《Gradient-based learning applied to document recognition》，论文里使用的是mnist手写数字作为输入数据（32 * 32）进行验证。我们来看一下网络结构。
LeNet-5一共有8层: 1个输入层+3个卷积层(C1、C3、C5)+2个下采样层(S2、S4)+1个全连接层(F6)+1个输出层，每层有多个feature map(自动提取的多组特征)。
Input输入层
cifar10 数据集，每一张图片尺寸：32 * 32
C1 卷积层
- 6个feature_map，卷积核大小 5 * 5 ，feature_map尺寸：28 * 28
- 每个卷积神经元的参数数目：5 * 5 = 25个和一个bias参数
- 连接数目：（5*5+1）* 6 *（28*28） = 122,304
- 参数共享：每个feature_map内共享参数，$\therefore$共(5*5+1)*6 = 156个参数
S2 下采样层（池化层）
- 6个14*14的feature_map，pooling大小 2* 2
- 每个单元与上一层的feature_map中的一个2*2的滑动窗口连接，不重叠，因此S2每个feature_map大小是C1中feature_map大小的1/4
- 连接数：（2*2+1)*1*14*14*6 = 5880个
- 参数共享：每个feature_map内共享参数，有2 * 6 = 12个训练参数
C3 卷积层
这层略微复杂，S2神经元与C3是多对多的关系，比如最简单方式：用S2的所有feature map与C3的所有feature map做全连接(也可以对S2抽样几个feature map出来与C3某个feature map连接)，这种全连接方式下：6个S2的feature map使用6个独立的5×5卷积核得到C3中1个feature map(生成每个feature map时对应一个bias)，C3*有16个feature map，所以该层需要学习的参数个数为：(5×5×6+1)×16=2416个，神经元连接数为：2416×8×8=154624个。
S4 下采样层
同S2，如果采用Max Pooling/Mean Pooling，则该层需要学习的参数个数为0个，神经元连接数为：(2×2+1)×16×4×4=1280个。
C5卷积层
类似C3，用S4的所有feature map与C5的所有feature map做全连接，这种全连接方式下：16个S4的feature map使用16个独立的1×1卷积核得到C5中1个feature map(生成每个feature map时对应一个bias)，C5*有120个feature map，所以该层需要学习的参数个数为：(1×1×16+1)×120=2040个，神经元连接数为：2040个。
F6 全连接层
将C5层展开得到4×4×120=1920个节点，并接一个全连接层，考虑bias，该层需要学习的参数和连接个数为：(1920+1)*84=161364个。
输出层
该问题是个10分类问题，所以有10个输出单元，通过softmax做概率归一化，每个分类的输出单元对应84个输入。
LeNet-5的PaddlePaddle实现
1.网络结构 lenet.py
#coding:utf-8
'''
Created by huxiaoman 2017.11.27
lenet.py:LeNet-5
''' import os
from PIL import Image
import numpy as np
import paddle.v2 as paddle
from paddle.trainer_config_helpers import * with_gpu = os.getenv('WITH_GPU', '') != '' def lenet(img):
conv_pool_1 = paddle.networks.simple_img_conv_pool(
input=img,
filter_size=5,
num_filters=6,
num_channel=3,
pool_size=2,
pool_stride=2,
act=paddle.activation.Relu())
conv_pool_2 = paddle.networks.simple_img_conv_pool(
input=conv_pool_1,
filter_size=5,
num_filters=16,
pool_size=2,
pool_stride=2,
act=paddle.activation.Relu())
conv_3 = img_conv_layer(
input = conv_pool_2,
filter_size = 1,
num_filters = 120,
stride = 1)
fc = paddle.layer.fc(
input=conv_3, size=84, act=paddle.activation.Sigmoid())
return fc
2.训练代码 train_lenet.py
#coding:utf-8
'''
Created by huxiaoman 2017.11.27
train_lenet.py:训练LeNet-5对cifar10数据集进行分类
''' import sys, os import paddle.v2 as paddle
from lenet import lenet with_gpu = os.getenv('WITH_GPU', '') != '' def main():
datadim = 3 * 32 * 32
classdim = 10 # PaddlePaddle init
paddle.init(use_gpu=with_gpu, trainer_count=7) image = paddle.layer.data(
name="image", type=paddle.data_type.dense_vector(datadim)) # Add neural network config
# option 1. resnet
# net = resnet_cifar10(image, depth=32)
# option 2. vgg
net = lenet(image) out = paddle.layer.fc(
input=net, size=classdim, act=paddle.activation.Softmax()) lbl = paddle.layer.data(
name="label", type=paddle.data_type.integer_value(classdim))
cost = paddle.layer.classification_cost(input=out, label=lbl) # Create parameters
parameters = paddle.parameters.create(cost) # Create optimizer
momentum_optimizer = paddle.optimizer.Momentum(
momentum=0.9,
regularization=paddle.optimizer.L2Regularization(rate=0.0002 * 128),
learning_rate=0.1 / 128.0,
learning_rate_decay_a=0.1,
learning_rate_decay_b=50000 * 100,
learning_rate_schedule='discexp') # End batch and end pass event handler
def event_handler(event):
if isinstance(event, paddle.event.EndIteration):
if event.batch_id % 100 == 0:
print "\nPass %d, Batch %d, Cost %f, %s" % (
event.pass_id, event.batch_id, event.cost, event.metrics)
else:
sys.stdout.write('.')
sys.stdout.flush()
if isinstance(event, paddle.event.EndPass):
# save parameters
with open('params_pass_%d.tar' % event.pass_id, 'w') as f:
parameters.to_tar(f) result = trainer.test(
reader=paddle.batch(
paddle.dataset.cifar.test10(), batch_size=128),
feeding={'image': 0,
'label': 1})
print "\nTest with Pass %d, %s" % (event.pass_id, result.metrics) # Create trainer
trainer = paddle.trainer.SGD(
cost=cost, parameters=parameters, update_equation=momentum_optimizer) # Save the inference topology to protobuf.
inference_topology = paddle.topology.Topology(layers=out)
with open("inference_topology.pkl", 'wb') as f:
inference_topology.serialize_for_inference(f) trainer.train(
reader=paddle.batch(
paddle.reader.shuffle(
paddle.dataset.cifar.train10(), buf_size=50000),
batch_size=128),
num_passes=200,
event_handler=event_handler,
feeding={'image': 0,
'label': 1}) # inference
from PIL import Image
import numpy as np
import os def load_image(file):
im = Image.open(file)
im = im.resize((32, 32), Image.ANTIALIAS)
im = np.array(im).astype(np.float32)
# The storage order of the loaded image is W(widht),
# H(height), C(channel). PaddlePaddle requires
# the CHW order, so transpose them.
im = im.transpose((2, 0, 1)) # CHW
# In the training phase, the channel order of CIFAR
# image is B(Blue), G(green), R(Red). But PIL open
# image in RGB mode. It must swap the channel order.
im = im[(2, 1, 0), :, :] # BGR
im = im.flatten()
im = im / 255.0
return im test_data = []
cur_dir = os.path.dirname(os.path.realpath(__file__))
test_data.append((load_image(cur_dir + '/image/dog.png'), )) # users can remove the comments and change the model name
# with open('params_pass_50.tar', 'r') as f:
# parameters = paddle.parameters.Parameters.from_tar(f) probs = paddle.infer(
output_layer=out, parameters=parameters, input=test_data)
lab = np.argsort(-probs) # probs and lab are the results of one batch data
print "Label of image/dog.png is: %d" % lab[0][0] if __name__ == '__main__':
main()
3.结果输出
I1129 14:52:44.314946 15153 Util.cpp:166] commandline: --use_gpu=True --trainer_count=7
[INFO 2017-11-29 14:52:50,490 layers.py:2539] output for __conv_pool_0___conv: c = 6, h = 28, w = 28, size = 4704
[INFO 2017-11-29 14:52:50,491 layers.py:2667] output for __conv_pool_0___pool: c = 6, h = 14, w = 14, size = 1176
[INFO 2017-11-29 14:52:50,491 layers.py:2539] output for __conv_pool_1___conv: c = 16, h = 10, w = 10, size = 1600
[INFO 2017-11-29 14:52:50,492 layers.py:2667] output for __conv_pool_1___pool: c = 16, h = 5, w = 5, size = 400
[INFO 2017-11-29 14:52:50,493 layers.py:2539] output for __conv_0__: c = 120, h = 5, w = 5, size = 3000
I1129 14:52:50.498749 15153 MultiGradientMachine.cpp:99] numLogicalDevices=1 numThreads=7 numDevices=8
I1129 14:52:50.545882 15153 GradientMachine.cpp:85] Initing parameters..
I1129 14:52:50.651103 15153 GradientMachine.cpp:92] Init parameters done. Pass 0, Batch 0, Cost 2.331898, {'classification_error_evaluator': 0.9609375}
```
......
Pass 199, Batch 300, Cost 0.004373, {'classification_error_evaluator': 0.0}
..........................................................................................I1129 16:17:08.678097 15153 MultiGradientMachine.cpp:99] numLogicalDevices=1 numThreads=7 numDevices=8 Test with Pass 199, {'classification_error_evaluator': 0.39579999446868896}
Label of image/dog.png is: 7
同样是7个线程，8个Tesla K80 GPU，batch_size = 128,迭代次数200次，耗时1h25min，错误分类率为0.3957，相比与simple_cnn的0.5248提高了12.91%。当然，这个结果也并不是很好，如果输出详细的日志，可以看到在训练的过程中loss先降后升，说明有一定程度的过拟合，对于如何防止过拟合，我们在后面会详细讲解。
有一个可视化CNN的网站可以对mnist和cifar10分类的网络结构进行可视化，这是cifar-10 BaseCNN的网络结构：
LeNet-5的Tensorflow实现
tensorflow版本的LeNet-5版本的可以参照models/tutorials/image/cifar10/(https://github.com/tensorflow/models/tree/master/tutorials/image/cifar10)的步骤来训练，不过这里面的代码包含了很多数据处理、权重衰减以及正则化的一些方法防止过拟合。按照官方写的,batch_size=128时在Tesla K40上迭代10w次需要4小时，准确率能达到86%。不过如果不对数据做处理，直接跑的话，效果应该没有这么好。不过可以仔细借鉴cifar10_inputs.py里的distorted_inouts函数对数据预处理增大数据集的思想，以及cifar10.py里对于权重和偏置的衰减设置等。目前迭代到1w次左右，cost是0.98，acc是78.4%
对于未进行数据处理的cifar10我准备也跑一次，看看效果如何，与paddle的结果对比一下。不过得等到周末再补上了 = =
总结
本节用常规的cifar-10数据集做图像分类，用了三种实现方式，第一种是自己设计的一个简单的cnn，第二种是LeNet-5，第三种是Tensorflow实现的LeNet-5，对比速度可以见一下表格：
可以看到LeNet-5相比于原始的simple_cnn在准确率和速度方面都有一定的的提升，等tensorflow版本跑完后可以把结果加上去再对比一下。不过用Lenet-5网络结构后，结果虽然有一定的提升，但是还是不够理想，在日志里看到loss的信息基本可以推断出是过拟合，对于神经网络训练过程中出现的过拟合情况我们应该如何避免，下期我们讲着重讲解。此外在下一节将介绍AlexNet，并对分类做一个实验，对比其效果。
参考文献
1.LeNet-5论文：《Gradient-based learning applied to document recognition》
2.可视化CNN：http://shixialiu.com/publications/cnnvis/demo/