MXNET深度学习框架-22-使用gluon的GoogleNet
GoogleNet和VGG是2014年ImageNet挑战赛(ILSVRC14)的双雄,其中,GoogLeNet获得了冠军、VGGNet获得了亚军。对于上述两种结构来说,它们的共同点都是比较深的CNN模型,不同的是,GoogleNet更深,且模型大小比VGGNet还要小一写。
下面我们来探讨一下GoogleNet,说到GoogleNet,先来说说Inception结构,它基于NiN思想做了很大的改进:
通过上图可以看到,一个输入总共有4条路径到达“Concat”层。下面通过相关代码来实现一下:
class inception(gn.nn.Block):
def __init__(self,n1_1,n2_1,n2_3,n3_1,n3_5,n4_1,**kwargs):
'''
:param n1_1: 第1条路径的卷积输出数(1X1卷积)
:param n2_1: 第2条路径的卷积输出数(1X1卷积)
:param n2_3: 第2条路径的卷积输出数(3X3卷积)
:param n3_1: 第3条路径的卷积输出数(1X1卷积)
:param n3_5: 第3条路径的卷积输出数(5X5卷积)
:param n4_1: 第3条路径的卷积输出数(1X1卷积)
'''
super(inception,self).__init__(**kwargs)
with self.name_scope():
# path 1
self.p1_conv_1=gn.nn.Conv2D(n1_1,kernel_size=1,activation="relu")
# path 2
self.p2_conv_1=gn.nn.Conv2D(n2_1,kernel_size=1,activation="relu")
self.p2_conv_3=gn.nn.Conv2D(n2_3,kernel_size=3,padding=1,activation="relu") # padding=1说明输出和输入的h,w不变,如果变化了concat就不行了
# path 3
self.p3_conv_1 = gn.nn.Conv2D(n3_1, kernel_size=1, activation="relu")
self.p3_conv_5 = gn.nn.Conv2D(n3_5, kernel_size=5, padding=2, activation="relu")
# path 4
self.p4_pool_3 = gn.nn.MaxPool2D(pool_size=3,padding=1,strides=1)
self.p4_conv_5 = gn.nn.Conv2D(n4_1, kernel_size=1, activation="relu")
def forward(self, x):
p1=self.p1_conv_1(x)
p2=self.p2_conv_3(self.p2_conv_1(x))
p3=self.p3_conv_5(self.p3_conv_1(x))
p4=self.p4_conv_5(self.p4_pool_3(x))
return nd.concat(p1,p2,p3,p4,dim=1) # 融合
为什么Inception结构里要使用1×1卷积?试想一下,1条大河里面有4条支流汇聚在一起,那么大河的水量是不是就会陡然增大?前一章NiN也讲过了1×1卷积可以改变通道数,也就是使支流的水量减少,这样大河的水量才不会增多。
接下来我们运行一个实例看看:
# 运行一个实例
test=inception(64,96,128,16,32,32) #
test.initialize()
x=nd.random_normal(shape=(32,3,64,64)) # NCHW
print(test(x).shape)
结果:
我们可以看到,,除了通道数变了以外,其余的都保持不变。那么256是怎么来的:
Inception结构共有4条路经,所以把所有路径上的最后一个块的通道数加起来就好了:64+128+32+32=256。
下面我们按照Googlenet的结构来构建一下模型(模型结构参照这篇博文:GoogLeNet网络结构)
'''---模型定义---'''
def Googlenet():
# GoogleNet 可以分为6块
# block1
b1=gn.nn.Sequential()
b1.add(gn.nn.Conv2D(64,7,2,padding=3,activation="relu"),
gn.nn.MaxPool2D(pool_size=3,strides=2,padding=1)) # 这里不使用LRN
# block2
b2=gn.nn.Sequential()
b2.add(gn.nn.Conv2D(64,1,1,activation="relu"),
gn.nn.Conv2D(192,3,1,padding=1,activation="relu"),
gn.nn.MaxPool2D(pool_size=3,strides=2,padding=1))
# block3 --inception
b3=gn.nn.Sequential()
b3.add(inception(64,96,128,16,32,32),
inception(128,128,192,32,96,64),
gn.nn.MaxPool2D(pool_size=3,strides=2,padding=1))
# block4 --inception
b4 = gn.nn.Sequential()
b4.add(inception(192, 96, 208, 16, 48, 64),
inception(160, 112, 224, 24, 64, 64),
inception(128, 128, 256, 24, 64, 64),
inception(112, 144, 288, 32, 64, 64),
inception(256, 160, 320, 32, 128, 128),
gn.nn.MaxPool2D(pool_size=3, strides=2, padding=1))
# block5 --inception
b5 = gn.nn.Sequential()
b5.add(inception(256, 160, 320, 32, 128, 128),
inception(384, 192, 384, 48, 128, 128),
gn.nn.AvgPool2D())
# block6 --Dense
b6 = gn.nn.Sequential()
b6.add(gn.nn.Flatten(),
gn.nn.Dropout(0.5),
gn.nn.Dense(10))
net=gn.nn.Sequential()
net.add(b1,b2,b3,b4,b5,b6)
return net
虽然有点长,但理解了其实也很简单,下面运行一个实例看看大致结构:
# 例子
net=Googlenet()
X = nd.random.uniform(shape=(1, 1, 96, 96))
net.initialize()
for layer in net:
X = layer(X)
print(layer.name, 'output shape:\t', X.shape)
结果:
接下来的训练部分与数据准备部分与前几章几乎一样,下面附上所有代码:
import mxnet.ndarray as nd
import mxnet.autograd as ag
import mxnet.gluon as gn
import mxnet as mx
import matplotlib.pyplot as plt
from mxnet import init
class inception(gn.nn.Block):
def __init__(self,n1_1,n2_1,n2_3,n3_1,n3_5,n4_1,**kwargs):
'''
:param n1_1: 第1条路径的卷积输出数(1X1卷积)
:param n2_1: 第2条路径的卷积输出数(1X1卷积)
:param n2_3: 第2条路径的卷积输出数(3X3卷积)
:param n3_1: 第3条路径的卷积输出数(1X1卷积)
:param n3_5: 第3条路径的卷积输出数(5X5卷积)
:param n4_1: 第3条路径的卷积输出数(1X1卷积)
'''
super(inception,self).__init__(**kwargs)
with self.name_scope():
# path 1
self.p1_conv_1=gn.nn.Conv2D(n1_1,kernel_size=1,activation="relu")
# path 2
self.p2_conv_1=gn.nn.Conv2D(n2_1,kernel_size=1,activation="relu")
self.p2_conv_3=gn.nn.Conv2D(n2_3,kernel_size=3,padding=1,activation="relu") # padding=1说明输出和输入的h,w不变,如果变化了concat就不行了
# path 3
self.p3_conv_1 = gn.nn.Conv2D(n3_1, kernel_size=1, activation="relu")
self.p3_conv_5 = gn.nn.Conv2D(n3_5, kernel_size=5, padding=2, activation="relu")
# path 4
self.p4_pool_3 = gn.nn.MaxPool2D(pool_size=3,padding=1,strides=1)
self.p4_conv_5 = gn.nn.Conv2D(n4_1, kernel_size=1, activation="relu")
def forward(self, x):
p1=self.p1_conv_1(x)
p2=self.p2_conv_3(self.p2_conv_1(x))
p3=self.p3_conv_5(self.p3_conv_1(x))
p4=self.p4_conv_5(self.p4_pool_3(x))
return nd.concat(p1,p2,p3,p4,dim=1) # 融合
# 运行一个实例
# test=inception(64,96,128,16,32,32) #
# test.initialize()
# x=nd.random_normal(shape=(32,3,64,64)) # NCHW
# print(test(x).shape)
'''---模型定义---'''
def Googlenet():
# GoogleNet 可以分为6块
# block1
b1=gn.nn.Sequential()
b1.add(gn.nn.Conv2D(64,7,2,padding=3,activation="relu"),
gn.nn.MaxPool2D(pool_size=3,strides=2,padding=1)) # 这里不使用LRN
# block2
b2=gn.nn.Sequential()
b2.add(gn.nn.Conv2D(64,1,1,activation="relu"),
gn.nn.Conv2D(192,3,1,padding=1,activation="relu"),
gn.nn.MaxPool2D(pool_size=3,strides=2,padding=1))
# block3 --inception
b3=gn.nn.Sequential()
b3.add(inception(64,96,128,16,32,32),
inception(128,128,192,32,96,64),
gn.nn.MaxPool2D(pool_size=3,strides=2,padding=1))
# block4 --inception
b4 = gn.nn.Sequential()
b4.add(inception(192, 96, 208, 16, 48, 64),
inception(160, 112, 224, 24, 64, 64),
inception(128, 128, 256, 24, 64, 64),
inception(112, 144, 288, 32, 64, 64),
inception(256, 160, 320, 32, 128, 128),
gn.nn.MaxPool2D(pool_size=3, strides=2, padding=1))
# block5 --inception
b5 = gn.nn.Sequential()
b5.add(inception(256, 160, 320, 32, 128, 128),
inception(384, 192, 384, 48, 128, 128),
gn.nn.AvgPool2D())
# block6 --Dense
b6 = gn.nn.Sequential()
b6.add(gn.nn.Flatten(),
gn.nn.Dropout(0.5),
gn.nn.Dense(10))
net=gn.nn.Sequential()
net.add(b1,b2,b3,b4,b5,b6)
return net
# # 例子
# net=Googlenet()
# X = nd.random.uniform(shape=(1, 1, 96, 96))
# net.initialize()
# for layer in net:
# X = layer(X)
# print(layer.name, 'output shape:\t', X.shape)
ctx=mx.gpu()
net=Googlenet()
net.initialize(init=init.Xavier(),ctx=ctx)
'''---读取数据和预处理---'''
def load_data_fashion_mnist(batch_size, resize=None):
transformer = []
if resize:
transformer += [gn.data.vision.transforms.Resize(resize)]
transformer += [gn.data.vision.transforms.ToTensor()]
transformer = gn.data.vision.transforms.Compose(transformer)
mnist_train = gn.data.vision.FashionMNIST(train=True)
mnist_test = gn.data.vision.FashionMNIST(train=False)
train_iter = gn.data.DataLoader(
mnist_train.transform_first(transformer), batch_size, shuffle=True)
test_iter = gn.data.DataLoader(
mnist_test.transform_first(transformer), batch_size, shuffle=False)
return train_iter, test_iter
batch_size=128
train_iter,test_iter=load_data_fashion_mnist(batch_size,resize=96) # 96,因为图片加大的话训练很慢,而且显存会吃不消
# softmax和交叉熵损失函数
# 由于将它们分开会导致数值不稳定(前两章博文的结果可以对比),所以直接使用gluon提供的API
cross_loss=gn.loss.SoftmaxCrossEntropyLoss()
# 定义准确率
def accuracy(output,label):
return nd.mean(output.argmax(axis=1)==label).asscalar()
def evaluate_accuracy(data_iter,net):# 定义测试集准确率
acc=0
for data,label in data_iter:
data, label = data.as_in_context(ctx), label.as_in_context(ctx)
label = label.astype('float32')
output=net(data)
acc+=accuracy(output,label)
return acc/len(data_iter)
# softmax和交叉熵分开的话数值可能会不稳定
cross_loss=gn.loss.SoftmaxCrossEntropyLoss()
# 优化
train_step=gn.Trainer(net.collect_params(),'sgd',{"learning_rate":0.01})
# 训练
lr=0.1
epochs=20
for epoch in range(epochs):
n=0
train_loss=0
train_acc=0
for image,y in train_iter:
image, y = image.as_in_context(ctx), y.as_in_context(ctx)
y = y.astype('float32')
with ag.record():
output = net(image)
loss = cross_loss(output, y)
loss.backward()
train_step.step(batch_size)
train_loss += nd.mean(loss).asscalar()
train_acc += accuracy(output, y)
test_acc = evaluate_accuracy(test_iter, net)
print("Epoch %d, Loss:%f, Train acc:%f, Test acc:%f"
%(epoch,train_loss/len(train_iter),train_acc/len(train_iter),test_acc))
训练结果:
当然,Goolenet有很多版本,本章实现的是最初的版本,而Inception块的通道数分配之比也是在ImageNet数据集上通过大量的实验得来的。
结论:将原先1个大卷积变成4个小卷积,信息流动稍微能够大一些,特征更容易抓取,且计算量能有所降低。
推荐阅读
-
深度学习零基础使用 PyTorch 框架跑 MNIST 数据集的第四天:单例测试
-
TensorFlow和Caffe、MXNet、Keras等深度学习框架的对比
-
MXNET深度学习框架-12-使用gluon实现LeNet-5
-
MXNET深度学习框架-21-使用gluon的NiN(Network in Network)
-
MXNET深度学习框架-06-使用gluon实现逻辑回归
-
MXNET深度学习框架-08-使用gluon实现多层感知机(MLP)
-
MXNET深度学习框架-04-使用gluon实现线性回归
-
MXNET深度学习框架-20-使用Gluon的Batch Normalization(批量归一化)
-
TensorFlow和Caffe、MXNet、Keras等深度学习框架的对比
-
MXNET深度学习框架-16-使用gluon实现dropout