论文中文翻译:http://blog.csdn.net/motianchi/article/details/50970136
一、文章意义
这是Google公司Inception系列的开山之作,在这篇文章中首次提出了Inception模块,后面的Inception v2&v3、Inception v4也是在这篇文章的基础上改进的。GoogLeNet是2014年ILSVRC挑战赛冠军,将Top5 的错误率降低到6.67%,是一个22层的深度网络。此外,作者团队的阵容也是吊炸天。
二、文章思路
作者在文章的第三部分“Motivation and High Consideration”部分中提到了文章的写作动机及高层次的考虑设计。作者认为获得高质量模型的最直接方法是增加网络的深度(层数)和宽度(层核或神经元数)(“The most straightforward way of improving the performance of deep neural network is by increasing theri size”),但是这样的设计思路会带来两个缺陷:
(1)“Bigger size typically means a large number of parameters,which makes the enlarged network more prone to overfitting”,随着层的增加和神经元个数的增加,参数数量肯定会增加,随之而来的就是很容易过拟合。
(2)“Another drawback of uniformly increased network size is the dramatically increased use of computational resources”,即层数和神经元个数的增加会带来更高计算资源的要求。
针对这两个问题,作者认为根本的解决办法是把全连接层改成稀疏结构,甚至卷积层也是。但是,当涉及到大量非统一的稀疏的数据结构的计算时,现在的计算设施是很低效的。作者为了将网络向稀疏结构设计,同时又兼顾计算性能,作者根据“稀疏矩阵聚类为较为密集的子矩阵来提高计算性能”的理论,据此设计了名为Inception结构。
注:文章的第三部分读的晕晕乎乎的啊,自己对“稀疏性”了解真的少,网上搜了一些,有人说是1*1卷积降维体现稀疏性,但个人感觉不仅仅是这个,希望了解的同学能给一些参考,谢谢!
三、文章结构
分为两个部分:Inception结构和整体的GoogLeNet网络结构。
1、Inception结构
图(b)在图(a)的基础上引入了1*1卷积降维(这就是Inception结构的稀疏性的体现?)。GoogLeNet并没有采用AlexNet的结构,而是用多尺度卷积(网络宽度提高)+1*1卷积降维+Pooling组成基本模块,整个网络也是Inception模块的堆叠,这是GoogLeNet网络的一大特色。
卷积尺度包括5*5,3*3,1*1,通过多个不同的尺寸的卷积核可以提取图像中的不同特征,即增加了网络的宽度。通过不同的padding使不同尺寸的卷积核最终输出的特征图大小一致。最顶层的 concatenation层是来自不同卷积核的输出特征图的叠加。
2、GoogLeNet网络结构
几点说明:
(1)每一个卷积层后面都跟着一个ReLUs层,图上省略掉了;
(2)在Inception(4a)和Inception(4d)的后面都跟着一个额外的Softmax层,用以减小“梯度消失”现象。这两个softmax的loss值最终以一个 0.3的权重连接到最后的softmax的loss值上,也有多模型融合的意思在里面,具体的链接为:Average Pooling+1*1卷积+ReLU+FC+ReLU+
Dropout+FC+SoftmaxWithLoss,Dropout的系数为0.7,FC层据说是为了方便微调用的;
(3)在Inception(4a)和Inception(4d)后都接着一个Average Pooling层;
(4)最顶端的FC层有1024个单元;
整个网络包含22层(不算池化层),是由Inception模块堆叠而成,其中的ReLU和Dropout图中省略了。除了最顶端的softmax层以外还有两个额外的softmax层,网络中穿插着Max Pooling和Average Pooling,细节网络可参考caffe里的GoogLeNet。
四、Caffe中的实现
name: "GoogleNet"
layer {
name: "data"
type: "Data"
top: "data"
top: "label"
include {
phase: TRAIN
}
transform_param {
mirror: true
crop_size: 224
mean_value: 104
mean_value: 117
mean_value: 123
}
data_param {
source: "examples/imagenet/ilsvrc12_train_lmdb"
batch_size: 32
backend: LMDB
}
}
layer {
name: "data"
type: "Data"
top: "data"
top: "label"
include {
phase: TEST
}
transform_param {
mirror: false
crop_size: 224
mean_value: 104
mean_value: 117
mean_value: 123
}
data_param {
source: "examples/imagenet/ilsvrc12_val_lmdb"
batch_size: 50
backend: LMDB
}
}
layer {
name: "conv1/7x7_s2"
type: "Convolution"
bottom: "data"
top: "conv1/7x7_s2"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
convolution_param {
num_output: 64
pad: 3
kernel_size: 7
stride: 2
weight_filler {
type: "xavier"
}
bias_filler {
type: "constant"
value: 0.2
}
}
}
layer {
name: "conv1/relu_7x7"
type: "ReLU"
bottom: "conv1/7x7_s2"
top: "conv1/7x7_s2"
}
layer {
name: "pool1/3x3_s2"
type: "Pooling"
bottom: "conv1/7x7_s2"
top: "pool1/3x3_s2"
pooling_param {
pool: MAX
kernel_size: 3
stride: 2
}
}
layer {
name: "pool1/norm1"
type: "LRN"
bottom: "pool1/3x3_s2"
top: "pool1/norm1"
lrn_param {
local_size: 5
alpha: 0.0001
beta: 0.75
}
}
layer {
name: "conv2/3x3_reduce"
type: "Convolution"
bottom: "pool1/norm1"
top: "conv2/3x3_reduce"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
convolution_param {
num_output: 64
kernel_size: 1
weight_filler {
type: "xavier"
}
bias_filler {
type: "constant"
value: 0.2
}
}
}
layer {
name: "conv2/relu_3x3_reduce"
type: "ReLU"
bottom: "conv2/3x3_reduce"
top: "conv2/3x3_reduce"
}
layer {
name: "conv2/3x3"
type: "Convolution"
bottom: "conv2/3x3_reduce"
top: "conv2/3x3"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
convolution_param {
num_output: 192
pad: 1
kernel_size: 3
weight_filler {
type: "xavier"
}
bias_filler {
type: "constant"
value: 0.2
}
}
}
layer {
name: "conv2/relu_3x3"
type: "ReLU"
bottom: "conv2/3x3"
top: "conv2/3x3"
}
layer {
name: "conv2/norm2"
type: "LRN"
bottom: "conv2/3x3"
top: "conv2/norm2"
lrn_param {
local_size: 5
alpha: 0.0001
beta: 0.75
}
}
layer {
name: "pool2/3x3_s2"
type: "Pooling"
bottom: "conv2/norm2"
top: "pool2/3x3_s2"
pooling_param {
pool: MAX
kernel_size: 3
stride: 2
}
}
layer {
name: "inception_3a/1x1"
type: "Convolution"
bottom: "pool2/3x3_s2"
top: "inception_3a/1x1"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
convolution_param {
num_output: 64
kernel_size: 1
weight_filler {
type: "xavier"
}
bias_filler {
type: "constant"
value: 0.2
}
}
}
layer {
name: "inception_3a/relu_1x1"
type: "ReLU"
bottom: "inception_3a/1x1"
top: "inception_3a/1x1"
}
layer {
name: "inception_3a/3x3_reduce"
type: "Convolution"
bottom: "pool2/3x3_s2"
top: "inception_3a/3x3_reduce"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
convolution_param {
num_output: 96
kernel_size: 1
weight_filler {
type: "xavier"
}
bias_filler {
type: "constant"
value: 0.2
}
}
}
layer {
name: "inception_3a/relu_3x3_reduce"
type: "ReLU"
bottom: "inception_3a/3x3_reduce"
top: "inception_3a/3x3_reduce"
}
layer {
name: "inception_3a/3x3"
type: "Convolution"
bottom: "inception_3a/3x3_reduce"
top: "inception_3a/3x3"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
convolution_param {
num_output: 128
pad: 1
kernel_size: 3
weight_filler {
type: "xavier"
}
bias_filler {
type: "constant"
value: 0.2
}
}
}
layer {
name: "inception_3a/relu_3x3"
type: "ReLU"
bottom: "inception_3a/3x3"
top: "inception_3a/3x3"
}
layer {
name: "inception_3a/5x5_reduce"
type: "Convolution"
bottom: "pool2/3x3_s2"
top: "inception_3a/5x5_reduce"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
convolution_param {
num_output: 16
kernel_size: 1
weight_filler {
type: "xavier"
}
bias_filler {
type: "constant"
value: 0.2
}
}
}
layer {
name: "inception_3a/relu_5x5_reduce"
type: "ReLU"
bottom: "inception_3a/5x5_reduce"
top: "inception_3a/5x5_reduce"
}
layer {
name: "inception_3a/5x5"
type: "Convolution"
bottom: "inception_3a/5x5_reduce"
top: "inception_3a/5x5"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
convolution_param {
num_output: 32
pad: 2
kernel_size: 5
weight_filler {
type: "xavier"
}
bias_filler {
type: "constant"
value: 0.2
}
}
}
layer {
name: "inception_3a/relu_5x5"
type: "ReLU"
bottom: "inception_3a/5x5"
top: "inception_3a/5x5"
}
layer {
name: "inception_3a/pool"
type: "Pooling"
bottom: "pool2/3x3_s2"
top: "inception_3a/pool"
pooling_param {
pool: MAX
kernel_size: 3
stride: 1
pad: 1
}
}
layer {
name: "inception_3a/pool_proj"
type: "Convolution"
bottom: "inception_3a/pool"
top: "inception_3a/pool_proj"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
convolution_param {
num_output: 32
kernel_size: 1
weight_filler {
type: "xavier"
}
bias_filler {
type: "constant"
value: 0.2
}
}
}
layer {
name: "inception_3a/relu_pool_proj"
type: "ReLU"
bottom: "inception_3a/pool_proj"
top: "inception_3a/pool_proj"
}
layer {
name: "inception_3a/output"
type: "Concat"
bottom: "inception_3a/1x1"
bottom: "inception_3a/3x3"
bottom: "inception_3a/5x5"
bottom: "inception_3a/pool_proj"
top: "inception_3a/output"
}
layer {
name: "inception_3b/1x1"
type: "Convolution"
bottom: "inception_3a/output"
top: "inception_3b/1x1"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
convolution_param {
num_output: 128
kernel_size: 1
weight_filler {
type: "xavier"
}
bias_filler {
type: "constant"
value: 0.2
}
}
}
layer {
name: "inception_3b/relu_1x1"
type: "ReLU"
bottom: "inception_3b/1x1"
top: "inception_3b/1x1"
}
layer {
name: "inception_3b/3x3_reduce"
type: "Convolution"
bottom: "inception_3a/output"
top: "inception_3b/3x3_reduce"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
convolution_param {
num_output: 128
kernel_size: 1
weight_filler {
type: "xavier"
}
bias_filler {
type: "constant"
value: 0.2
}
}
}
layer {
name: "inception_3b/relu_3x3_reduce"
type: "ReLU"
bottom: "inception_3b/3x3_reduce"
top: "inception_3b/3x3_reduce"
}
layer {
name: "inception_3b/3x3"
type: "Convolution"
bottom: "inception_3b/3x3_reduce"
top: "inception_3b/3x3"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
convolution_param {
num_output: 192
pad: 1
kernel_size: 3
weight_filler {
type: "xavier"
}
bias_filler {
type: "constant"
value: 0.2
}
}
}
layer {
name: "inception_3b/relu_3x3"
type: "ReLU"
bottom: "inception_3b/3x3"
top: "inception_3b/3x3"
}
layer {
name: "inception_3b/5x5_reduce"
type: "Convolution"
bottom: "inception_3a/output"
top: "inception_3b/5x5_reduce"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
convolution_param {
num_output: 32
kernel_size: 1
weight_filler {
type: "xavier"
}
bias_filler {
type: "constant"
value: 0.2
}
}
}
layer {
name: "inception_3b/relu_5x5_reduce"
type: "ReLU"
bottom: "inception_3b/5x5_reduce"
top: "inception_3b/5x5_reduce"
}
layer {
name: "inception_3b/5x5"
type: "Convolution"
bottom: "inception_3b/5x5_reduce"
top: "inception_3b/5x5"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
convolution_param {
num_output: 96
pad: 2
kernel_size: 5
weight_filler {
type: "xavier"
}
bias_filler {
type: "constant"
value: 0.2
}
}
}
layer {
name: "inception_3b/relu_5x5"
type: "ReLU"
bottom: "inception_3b/5x5"
top: "inception_3b/5x5"
}
layer {
name: "inception_3b/pool"
type: "Pooling"
bottom: "inception_3a/output"
top: "inception_3b/pool"
pooling_param {
pool: MAX
kernel_size: 3
stride: 1
pad: 1
}
}
layer {
name: "inception_3b/pool_proj"
type: "Convolution"
bottom: "inception_3b/pool"
top: "inception_3b/pool_proj"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
convolution_param {
num_output: 64
kernel_size: 1
weight_filler {
type: "xavier"
}
bias_filler {
type: "constant"
value: 0.2
}
}
}
layer {
name: "inception_3b/relu_pool_proj"
type: "ReLU"
bottom: "inception_3b/pool_proj"
top: "inception_3b/pool_proj"
}
layer {
name: "inception_3b/output"
type: "Concat"
bottom: "inception_3b/1x1"
bottom: "inception_3b/3x3"
bottom: "inception_3b/5x5"
bottom: "inception_3b/pool_proj"
top: "inception_3b/output"
}
layer {
name: "pool3/3x3_s2"
type: "Pooling"
bottom: "inception_3b/output"
top: "pool3/3x3_s2"
pooling_param {
pool: MAX
kernel_size: 3
stride: 2
}
}
layer {
name: "inception_4a/1x1"
type: "Convolution"
bottom: "pool3/3x3_s2"
top: "inception_4a/1x1"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
convolution_param {
num_output: 192
kernel_size: 1
weight_filler {
type: "xavier"
}
bias_filler {
type: "constant"
value: 0.2
}
}
}
layer {
name: "inception_4a/relu_1x1"
type: "ReLU"
bottom: "inception_4a/1x1"
top: "inception_4a/1x1"
}
layer {
name: "inception_4a/3x3_reduce"
type: "Convolution"
bottom: "pool3/3x3_s2"
top: "inception_4a/3x3_reduce"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
convolution_param {
num_output: 96
kernel_size: 1
weight_filler {
type: "xavier"
}
bias_filler {
type: "constant"
value: 0.2
}
}
}
layer {
name: "inception_4a/relu_3x3_reduce"
type: "ReLU"
bottom: "inception_4a/3x3_reduce"
top: "inception_4a/3x3_reduce"
}
layer {
name: "inception_4a/3x3"
type: "Convolution"
bottom: "inception_4a/3x3_reduce"
top: "inception_4a/3x3"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
convolution_param {
num_output: 208
pad: 1
kernel_size: 3
weight_filler {
type: "xavier"
}
bias_filler {
type: "constant"
value: 0.2
}
}
}
layer {
name: "inception_4a/relu_3x3"
type: "ReLU"
bottom: "inception_4a/3x3"
top: "inception_4a/3x3"
}
layer {
name: "inception_4a/5x5_reduce"
type: "Convolution"
bottom: "pool3/3x3_s2"
top: "inception_4a/5x5_reduce"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
convolution_param {
num_output: 16
kernel_size: 1
weight_filler {
type: "xavier"
}
bias_filler {
type: "constant"
value: 0.2
}
}
}
layer {
name: "inception_4a/relu_5x5_reduce"
type: "ReLU"
bottom: "inception_4a/5x5_reduce"
top: "inception_4a/5x5_reduce"
}
layer {
name: "inception_4a/5x5"
type: "Convolution"
bottom: "inception_4a/5x5_reduce"
top: "inception_4a/5x5"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
convolution_param {
num_output: 48
pad: 2
kernel_size: 5
weight_filler {
type: "xavier"
}
bias_filler {
type: "constant"
value: 0.2
}
}
}
layer {
name: "inception_4a/relu_5x5"
type: "ReLU"
bottom: "inception_4a/5x5"
top: "inception_4a/5x5"
}
layer {
name: "inception_4a/pool"
type: "Pooling"
bottom: "pool3/3x3_s2"
top: "inception_4a/pool"
pooling_param {
pool: MAX
kernel_size: 3
stride: 1
pad: 1
}
}
layer {
name: "inception_4a/pool_proj"
type: "Convolution"
bottom: "inception_4a/pool"
top: "inception_4a/pool_proj"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
convolution_param {
num_output: 64
kernel_size: 1
weight_filler {
type: "xavier"
}
bias_filler {
type: "constant"
value: 0.2
}
}
}
layer {
name: "inception_4a/relu_pool_proj"
type: "ReLU"
bottom: "inception_4a/pool_proj"
top: "inception_4a/pool_proj"
}
layer {
name: "inception_4a/output"
type: "Concat"
bottom: "inception_4a/1x1"
bottom: "inception_4a/3x3"
bottom: "inception_4a/5x5"
bottom: "inception_4a/pool_proj"
top: "inception_4a/output"
}
layer {
name: "loss1/ave_pool"
type: "Pooling"
bottom: "inception_4a/output"
top: "loss1/ave_pool"
pooling_param {
pool: AVE
kernel_size: 5
stride: 3
}
}
layer {
name: "loss1/conv"
type: "Convolution"
bottom: "loss1/ave_pool"
top: "loss1/conv"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
convolution_param {
num_output: 128
kernel_size: 1
weight_filler {
type: "xavier"
}
bias_filler {
type: "constant"
value: 0.2
}
}
}
layer {
name: "loss1/relu_conv"
type: "ReLU"
bottom: "loss1/conv"
top: "loss1/conv"
}
layer {
name: "loss1/fc"
type: "InnerProduct"
bottom: "loss1/conv"
top: "loss1/fc"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
inner_product_param {
num_output: 1024
weight_filler {
type: "xavier"
}
bias_filler {
type: "constant"
value: 0.2
}
}
}
layer {
name: "loss1/relu_fc"
type: "ReLU"
bottom: "loss1/fc"
top: "loss1/fc"
}
layer {
name: "loss1/drop_fc"
type: "Dropout"
bottom: "loss1/fc"
top: "loss1/fc"
dropout_param {
dropout_ratio: 0.7
}
}
layer {
name: "loss1/classifier"
type: "InnerProduct"
bottom: "loss1/fc"
top: "loss1/classifier"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
inner_product_param {
num_output: 1000
weight_filler {
type: "xavier"
}
bias_filler {
type: "constant"
value: 0
}
}
}
layer {
name: "loss1/loss"
type: "SoftmaxWithLoss"
bottom: "loss1/classifier"
bottom: "label"
top: "loss1/loss1"
loss_weight: 0.3
}
layer {
name: "loss1/top-1"
type: "Accuracy"
bottom: "loss1/classifier"
bottom: "label"
top: "loss1/top-1"
include {
phase: TEST
}
}
layer {
name: "loss1/top-5"
type: "Accuracy"
bottom: "loss1/classifier"
bottom: "label"
top: "loss1/top-5"
include {
phase: TEST
}
accuracy_param {
top_k: 5
}
}
layer {
name: "inception_4b/1x1"
type: "Convolution"
bottom: "inception_4a/output"
top: "inception_4b/1x1"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
convolution_param {
num_output: 160
kernel_size: 1
weight_filler {
type: "xavier"
}
bias_filler {
type: "constant"
value: 0.2
}
}
}
layer {
name: "inception_4b/relu_1x1"
type: "ReLU"
bottom: "inception_4b/1x1"
top: "inception_4b/1x1"
}
layer {
name: "inception_4b/3x3_reduce"
type: "Convolution"
bottom: "inception_4a/output"
top: "inception_4b/3x3_reduce"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
convolution_param {
num_output: 112
kernel_size: 1
weight_filler {
type: "xavier"
}
bias_filler {
type: "constant"
value: 0.2
}
}
}
layer {
name: "inception_4b/relu_3x3_reduce"
type: "ReLU"
bottom: "inception_4b/3x3_reduce"
top: "inception_4b/3x3_reduce"
}
layer {
name: "inception_4b/3x3"
type: "Convolution"
bottom: "inception_4b/3x3_reduce"
top: "inception_4b/3x3"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
convolution_param {
num_output: 224
pad: 1
kernel_size: 3
weight_filler {
type: "xavier"
}
bias_filler {
type: "constant"
value: 0.2
}
}
}
layer {
name: "inception_4b/relu_3x3"
type: "ReLU"
bottom: "inception_4b/3x3"
top: "inception_4b/3x3"
}
layer {
name: "inception_4b/5x5_reduce"
type: "Convolution"
bottom: "inception_4a/output"
top: "inception_4b/5x5_reduce"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
convolution_param {
num_output: 24
kernel_size: 1
weight_filler {
type: "xavier"
}
bias_filler {
type: "constant"
value: 0.2
}
}
}
layer {
name: "inception_4b/relu_5x5_reduce"
type: "ReLU"
bottom: "inception_4b/5x5_reduce"
top: "inception_4b/5x5_reduce"
}
layer {
name: "inception_4b/5x5"
type: "Convolution"
bottom: "inception_4b/5x5_reduce"
top: "inception_4b/5x5"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
convolution_param {
num_output: 64
pad: 2
kernel_size: 5
weight_filler {
type: "xavier"
}
bias_filler {
type: "constant"
value: 0.2
}
}
}
layer {
name: "inception_4b/relu_5x5"
type: "ReLU"
bottom: "inception_4b/5x5"
top: "inception_4b/5x5"
}
layer {
name: "inception_4b/pool"
type: "Pooling"
bottom: "inception_4a/output"
top: "inception_4b/pool"
pooling_param {
pool: MAX
kernel_size: 3
stride: 1
pad: 1
}
}
layer {
name: "inception_4b/pool_proj"
type: "Convolution"
bottom: "inception_4b/pool"
top: "inception_4b/pool_proj"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
convolution_param {
num_output: 64
kernel_size: 1
weight_filler {
type: "xavier"
}
bias_filler {
type: "constant"
value: 0.2
}
}
}
layer {
name: "inception_4b/relu_pool_proj"
type: "ReLU"
bottom: "inception_4b/pool_proj"
top: "inception_4b/pool_proj"
}
layer {
name: "inception_4b/output"
type: "Concat"
bottom: "inception_4b/1x1"
bottom: "inception_4b/3x3"
bottom: "inception_4b/5x5"
bottom: "inception_4b/pool_proj"
top: "inception_4b/output"
}
layer {
name: "inception_4c/1x1"
type: "Convolution"
bottom: "inception_4b/output"
top: "inception_4c/1x1"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
convolution_param {
num_output: 128
kernel_size: 1
weight_filler {
type: "xavier"
}
bias_filler {
type: "constant"
value: 0.2
}
}
}
layer {
name: "inception_4c/relu_1x1"
type: "ReLU"
bottom: "inception_4c/1x1"
top: "inception_4c/1x1"
}
layer {
name: "inception_4c/3x3_reduce"
type: "Convolution"
bottom: "inception_4b/output"
top: "inception_4c/3x3_reduce"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
convolution_param {
num_output: 128
kernel_size: 1
weight_filler {
type: "xavier"
}
bias_filler {
type: "constant"
value: 0.2
}
}
}
layer {
name: "inception_4c/relu_3x3_reduce"
type: "ReLU"
bottom: "inception_4c/3x3_reduce"
top: "inception_4c/3x3_reduce"
}
layer {
name: "inception_4c/3x3"
type: "Convolution"
bottom: "inception_4c/3x3_reduce"
top: "inception_4c/3x3"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
convolution_param {
num_output: 256
pad: 1
kernel_size: 3
weight_filler {
type: "xavier"
}
bias_filler {
type: "constant"
value: 0.2
}
}
}
layer {
name: "inception_4c/relu_3x3"
type: "ReLU"
bottom: "inception_4c/3x3"
top: "inception_4c/3x3"
}
layer {
name: "inception_4c/5x5_reduce"
type: "Convolution"
bottom: "inception_4b/output"
top: "inception_4c/5x5_reduce"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
convolution_param {
num_output: 24
kernel_size: 1
weight_filler {
type: "xavier"
}
bias_filler {
type: "constant"
value: 0.2
}
}
}
layer {
name: "inception_4c/relu_5x5_reduce"
type: "ReLU"
bottom: "inception_4c/5x5_reduce"
top: "inception_4c/5x5_reduce"
}
layer {
name: "inception_4c/5x5"
type: "Convolution"
bottom: "inception_4c/5x5_reduce"
top: "inception_4c/5x5"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
convolution_param {
num_output: 64
pad: 2
kernel_size: 5
weight_filler {
type: "xavier"
}
bias_filler {
type: "constant"
value: 0.2
}
}
}
layer {
name: "inception_4c/relu_5x5"
type: "ReLU"
bottom: "inception_4c/5x5"
top: "inception_4c/5x5"
}
layer {
name: "inception_4c/pool"
type: "Pooling"
bottom: "inception_4b/output"
top: "inception_4c/pool"
pooling_param {
pool: MAX
kernel_size: 3
stride: 1
pad: 1
}
}
layer {
name: "inception_4c/pool_proj"
type: "Convolution"
bottom: "inception_4c/pool"
top: "inception_4c/pool_proj"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
convolution_param {
num_output: 64
kernel_size: 1
weight_filler {
type: "xavier"
}
bias_filler {
type: "constant"
value: 0.2
}
}
}
layer {
name: "inception_4c/relu_pool_proj"
type: "ReLU"
bottom: "inception_4c/pool_proj"
top: "inception_4c/pool_proj"
}
layer {
name: "inception_4c/output"
type: "Concat"
bottom: "inception_4c/1x1"
bottom: "inception_4c/3x3"
bottom: "inception_4c/5x5"
bottom: "inception_4c/pool_proj"
top: "inception_4c/output"
}
layer {
name: "inception_4d/1x1"
type: "Convolution"
bottom: "inception_4c/output"
top: "inception_4d/1x1"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
convolution_param {
num_output: 112
kernel_size: 1
weight_filler {
type: "xavier"
}
bias_filler {
type: "constant"
value: 0.2
}
}
}
layer {
name: "inception_4d/relu_1x1"
type: "ReLU"
bottom: "inception_4d/1x1"
top: "inception_4d/1x1"
}
layer {
name: "inception_4d/3x3_reduce"
type: "Convolution"
bottom: "inception_4c/output"
top: "inception_4d/3x3_reduce"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
convolution_param {
num_output: 144
kernel_size: 1
weight_filler {
type: "xavier"
}
bias_filler {
type: "constant"
value: 0.2
}
}
}
layer {
name: "inception_4d/relu_3x3_reduce"
type: "ReLU"
bottom: "inception_4d/3x3_reduce"
top: "inception_4d/3x3_reduce"
}
layer {
name: "inception_4d/3x3"
type: "Convolution"
bottom: "inception_4d/3x3_reduce"
top: "inception_4d/3x3"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
convolution_param {
num_output: 288
pad: 1
kernel_size: 3
weight_filler {
type: "xavier"
}
bias_filler {
type: "constant"
value: 0.2
}
}
}
layer {
name: "inception_4d/relu_3x3"
type: "ReLU"
bottom: "inception_4d/3x3"
top: "inception_4d/3x3"
}
layer {
name: "inception_4d/5x5_reduce"
type: "Convolution"
bottom: "inception_4c/output"
top: "inception_4d/5x5_reduce"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
convolution_param {
num_output: 32
kernel_size: 1
weight_filler {
type: "xavier"
}
bias_filler {
type: "constant"
value: 0.2
}
}
}
layer {
name: "inception_4d/relu_5x5_reduce"
type: "ReLU"
bottom: "inception_4d/5x5_reduce"
top: "inception_4d/5x5_reduce"
}
layer {
name: "inception_4d/5x5"
type: "Convolution"
bottom: "inception_4d/5x5_reduce"
top: "inception_4d/5x5"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
convolution_param {
num_output: 64
pad: 2
kernel_size: 5
weight_filler {
type: "xavier"
}
bias_filler {
type: "constant"
value: 0.2
}
}
}
layer {
name: "inception_4d/relu_5x5"
type: "ReLU"
bottom: "inception_4d/5x5"
top: "inception_4d/5x5"
}
layer {
name: "inception_4d/pool"
type: "Pooling"
bottom: "inception_4c/output"
top: "inception_4d/pool"
pooling_param {
pool: MAX
kernel_size: 3
stride: 1
pad: 1
}
}
layer {
name: "inception_4d/pool_proj"
type: "Convolution"
bottom: "inception_4d/pool"
top: "inception_4d/pool_proj"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
convolution_param {
num_output: 64
kernel_size: 1
weight_filler {
type: "xavier"
}
bias_filler {
type: "constant"
value: 0.2
}
}
}
layer {
name: "inception_4d/relu_pool_proj"
type: "ReLU"
bottom: "inception_4d/pool_proj"
top: "inception_4d/pool_proj"
}
layer {
name: "inception_4d/output"
type: "Concat"
bottom: "inception_4d/1x1"
bottom: "inception_4d/3x3"
bottom: "inception_4d/5x5"
bottom: "inception_4d/pool_proj"
top: "inception_4d/output"
}
layer {
name: "loss2/ave_pool"
type: "Pooling"
bottom: "inception_4d/output"
top: "loss2/ave_pool"
pooling_param {
pool: AVE
kernel_size: 5
stride: 3
}
}
layer {
name: "loss2/conv"
type: "Convolution"
bottom: "loss2/ave_pool"
top: "loss2/conv"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
convolution_param {
num_output: 128
kernel_size: 1
weight_filler {
type: "xavier"
}
bias_filler {
type: "constant"
value: 0.2
}
}
}
layer {
name: "loss2/relu_conv"
type: "ReLU"
bottom: "loss2/conv"
top: "loss2/conv"
}
layer {
name: "loss2/fc"
type: "InnerProduct"
bottom: "loss2/conv"
top: "loss2/fc"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
inner_product_param {
num_output: 1024
weight_filler {
type: "xavier"
}
bias_filler {
type: "constant"
value: 0.2
}
}
}
layer {
name: "loss2/relu_fc"
type: "ReLU"
bottom: "loss2/fc"
top: "loss2/fc"
}
layer {
name: "loss2/drop_fc"
type: "Dropout"
bottom: "loss2/fc"
top: "loss2/fc"
dropout_param {
dropout_ratio: 0.7
}
}
layer {
name: "loss2/classifier"
type: "InnerProduct"
bottom: "loss2/fc"
top: "loss2/classifier"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
inner_product_param {
num_output: 1000
weight_filler {
type: "xavier"
}
bias_filler {
type: "constant"
value: 0
}
}
}
layer {
name: "loss2/loss"
type: "SoftmaxWithLoss"
bottom: "loss2/classifier"
bottom: "label"
top: "loss2/loss1"
loss_weight: 0.3
}
layer {
name: "loss2/top-1"
type: "Accuracy"
bottom: "loss2/classifier"
bottom: "label"
top: "loss2/top-1"
include {
phase: TEST
}
}
layer {
name: "loss2/top-5"
type: "Accuracy"
bottom: "loss2/classifier"
bottom: "label"
top: "loss2/top-5"
include {
phase: TEST
}
accuracy_param {
top_k: 5
}
}
layer {
name: "inception_4e/1x1"
type: "Convolution"
bottom: "inception_4d/output"
top: "inception_4e/1x1"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
convolution_param {
num_output: 256
kernel_size: 1
weight_filler {
type: "xavier"
}
bias_filler {
type: "constant"
value: 0.2
}
}
}
layer {
name: "inception_4e/relu_1x1"
type: "ReLU"
bottom: "inception_4e/1x1"
top: "inception_4e/1x1"
}
layer {
name: "inception_4e/3x3_reduce"
type: "Convolution"
bottom: "inception_4d/output"
top: "inception_4e/3x3_reduce"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
convolution_param {
num_output: 160
kernel_size: 1
weight_filler {
type: "xavier"
}
bias_filler {
type: "constant"
value: 0.2
}
}
}
layer {
name: "inception_4e/relu_3x3_reduce"
type: "ReLU"
bottom: "inception_4e/3x3_reduce"
top: "inception_4e/3x3_reduce"
}
layer {
name: "inception_4e/3x3"
type: "Convolution"
bottom: "inception_4e/3x3_reduce"
top: "inception_4e/3x3"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
convolution_param {
num_output: 320
pad: 1
kernel_size: 3
weight_filler {
type: "xavier"
}
bias_filler {
type: "constant"
value: 0.2
}
}
}
layer {
name: "inception_4e/relu_3x3"
type: "ReLU"
bottom: "inception_4e/3x3"
top: "inception_4e/3x3"
}
layer {
name: "inception_4e/5x5_reduce"
type: "Convolution"
bottom: "inception_4d/output"
top: "inception_4e/5x5_reduce"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
convolution_param {
num_output: 32
kernel_size: 1
weight_filler {
type: "xavier"
}
bias_filler {
type: "constant"
value: 0.2
}
}
}
layer {
name: "inception_4e/relu_5x5_reduce"
type: "ReLU"
bottom: "inception_4e/5x5_reduce"
top: "inception_4e/5x5_reduce"
}
layer {
name: "inception_4e/5x5"
type: "Convolution"
bottom: "inception_4e/5x5_reduce"
top: "inception_4e/5x5"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
convolution_param {
num_output: 128
pad: 2
kernel_size: 5
weight_filler {
type: "xavier"
}
bias_filler {
type: "constant"
value: 0.2
}
}
}
layer {
name: "inception_4e/relu_5x5"
type: "ReLU"
bottom: "inception_4e/5x5"
top: "inception_4e/5x5"
}
layer {
name: "inception_4e/pool"
type: "Pooling"
bottom: "inception_4d/output"
top: "inception_4e/pool"
pooling_param {
pool: MAX
kernel_size: 3
stride: 1
pad: 1
}
}
layer {
name: "inception_4e/pool_proj"
type: "Convolution"
bottom: "inception_4e/pool"
top: "inception_4e/pool_proj"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
convolution_param {
num_output: 128
kernel_size: 1
weight_filler {
type: "xavier"
}
bias_filler {
type: "constant"
value: 0.2
}
}
}
layer {
name: "inception_4e/relu_pool_proj"
type: "ReLU"
bottom: "inception_4e/pool_proj"
top: "inception_4e/pool_proj"
}
layer {
name: "inception_4e/output"
type: "Concat"
bottom: "inception_4e/1x1"
bottom: "inception_4e/3x3"
bottom: "inception_4e/5x5"
bottom: "inception_4e/pool_proj"
top: "inception_4e/output"
}
layer {
name: "pool4/3x3_s2"
type: "Pooling"
bottom: "inception_4e/output"
top: "pool4/3x3_s2"
pooling_param {
pool: MAX
kernel_size: 3
stride: 2
}
}
layer {
name: "inception_5a/1x1"
type: "Convolution"
bottom: "pool4/3x3_s2"
top: "inception_5a/1x1"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
convolution_param {
num_output: 256
kernel_size: 1
weight_filler {
type: "xavier"
}
bias_filler {
type: "constant"
value: 0.2
}
}
}
layer {
name: "inception_5a/relu_1x1"
type: "ReLU"
bottom: "inception_5a/1x1"
top: "inception_5a/1x1"
}
layer {
name: "inception_5a/3x3_reduce"
type: "Convolution"
bottom: "pool4/3x3_s2"
top: "inception_5a/3x3_reduce"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
convolution_param {
num_output: 160
kernel_size: 1
weight_filler {
type: "xavier"
}
bias_filler {
type: "constant"
value: 0.2
}
}
}
layer {
name: "inception_5a/relu_3x3_reduce"
type: "ReLU"
bottom: "inception_5a/3x3_reduce"
top: "inception_5a/3x3_reduce"
}
layer {
name: "inception_5a/3x3"
type: "Convolution"
bottom: "inception_5a/3x3_reduce"
top: "inception_5a/3x3"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
convolution_param {
num_output: 320
pad: 1
kernel_size: 3
weight_filler {
type: "xavier"
}
bias_filler {
type: "constant"
value: 0.2
}
}
}
layer {
name: "inception_5a/relu_3x3"
type: "ReLU"
bottom: "inception_5a/3x3"
top: "inception_5a/3x3"
}
layer {
name: "inception_5a/5x5_reduce"
type: "Convolution"
bottom: "pool4/3x3_s2"
top: "inception_5a/5x5_reduce"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
convolution_param {
num_output: 32
kernel_size: 1
weight_filler {
type: "xavier"
}
bias_filler {
type: "constant"
value: 0.2
}
}
}
layer {
name: "inception_5a/relu_5x5_reduce"
type: "ReLU"
bottom: "inception_5a/5x5_reduce"
top: "inception_5a/5x5_reduce"
}
layer {
name: "inception_5a/5x5"
type: "Convolution"
bottom: "inception_5a/5x5_reduce"
top: "inception_5a/5x5"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
convolution_param {
num_output: 128
pad: 2
kernel_size: 5
weight_filler {
type: "xavier"
}
bias_filler {
type: "constant"
value: 0.2
}
}
}
layer {
name: "inception_5a/relu_5x5"
type: "ReLU"
bottom: "inception_5a/5x5"
top: "inception_5a/5x5"
}
layer {
name: "inception_5a/pool"
type: "Pooling"
bottom: "pool4/3x3_s2"
top: "inception_5a/pool"
pooling_param {
pool: MAX
kernel_size: 3
stride: 1
pad: 1
}
}
layer {
name: "inception_5a/pool_proj"
type: "Convolution"
bottom: "inception_5a/pool"
top: "inception_5a/pool_proj"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
convolution_param {
num_output: 128
kernel_size: 1
weight_filler {
type: "xavier"
}
bias_filler {
type: "constant"
value: 0.2
}
}
}
layer {
name: "inception_5a/relu_pool_proj"
type: "ReLU"
bottom: "inception_5a/pool_proj"
top: "inception_5a/pool_proj"
}
layer {
name: "inception_5a/output"
type: "Concat"
bottom: "inception_5a/1x1"
bottom: "inception_5a/3x3"
bottom: "inception_5a/5x5"
bottom: "inception_5a/pool_proj"
top: "inception_5a/output"
}
layer {
name: "inception_5b/1x1"
type: "Convolution"
bottom: "inception_5a/output"
top: "inception_5b/1x1"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
convolution_param {
num_output: 384
kernel_size: 1
weight_filler {
type: "xavier"
}
bias_filler {
type: "constant"
value: 0.2
}
}
}
layer {
name: "inception_5b/relu_1x1"
type: "ReLU"
bottom: "inception_5b/1x1"
top: "inception_5b/1x1"
}
layer {
name: "inception_5b/3x3_reduce"
type: "Convolution"
bottom: "inception_5a/output"
top: "inception_5b/3x3_reduce"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
convolution_param {
num_output: 192
kernel_size: 1
weight_filler {
type: "xavier"
}
bias_filler {
type: "constant"
value: 0.2
}
}
}
layer {
name: "inception_5b/relu_3x3_reduce"
type: "ReLU"
bottom: "inception_5b/3x3_reduce"
top: "inception_5b/3x3_reduce"
}
layer {
name: "inception_5b/3x3"
type: "Convolution"
bottom: "inception_5b/3x3_reduce"
top: "inception_5b/3x3"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
convolution_param {
num_output: 384
pad: 1
kernel_size: 3
weight_filler {
type: "xavier"
}
bias_filler {
type: "constant"
value: 0.2
}
}
}
layer {
name: "inception_5b/relu_3x3"
type: "ReLU"
bottom: "inception_5b/3x3"
top: "inception_5b/3x3"
}
layer {
name: "inception_5b/5x5_reduce"
type: "Convolution"
bottom: "inception_5a/output"
top: "inception_5b/5x5_reduce"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
convolution_param {
num_output: 48
kernel_size: 1
weight_filler {
type: "xavier"
}
bias_filler {
type: "constant"
value: 0.2
}
}
}
layer {
name: "inception_5b/relu_5x5_reduce"
type: "ReLU"
bottom: "inception_5b/5x5_reduce"
top: "inception_5b/5x5_reduce"
}
layer {
name: "inception_5b/5x5"
type: "Convolution"
bottom: "inception_5b/5x5_reduce"
top: "inception_5b/5x5"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
convolution_param {
num_output: 128
pad: 2
kernel_size: 5
weight_filler {
type: "xavier"
}
bias_filler {
type: "constant"
value: 0.2
}
}
}
layer {
name: "inception_5b/relu_5x5"
type: "ReLU"
bottom: "inception_5b/5x5"
top: "inception_5b/5x5"
}
layer {
name: "inception_5b/pool"
type: "Pooling"
bottom: "inception_5a/output"
top: "inception_5b/pool"
pooling_param {
pool: MAX
kernel_size: 3
stride: 1
pad: 1
}
}
layer {
name: "inception_5b/pool_proj"
type: "Convolution"
bottom: "inception_5b/pool"
top: "inception_5b/pool_proj"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
convolution_param {
num_output: 128
kernel_size: 1
weight_filler {
type: "xavier"
}
bias_filler {
type: "constant"
value: 0.2
}
}
}
layer {
name: "inception_5b/relu_pool_proj"
type: "ReLU"
bottom: "inception_5b/pool_proj"
top: "inception_5b/pool_proj"
}
layer {
name: "inception_5b/output"
type: "Concat"
bottom: "inception_5b/1x1"
bottom: "inception_5b/3x3"
bottom: "inception_5b/5x5"
bottom: "inception_5b/pool_proj"
top: "inception_5b/output"
}
layer {
name: "pool5/7x7_s1"
type: "Pooling"
bottom: "inception_5b/output"
top: "pool5/7x7_s1"
pooling_param {
pool: AVE
kernel_size: 7
stride: 1
}
}
layer {
name: "pool5/drop_7x7_s1"
type: "Dropout"
bottom: "pool5/7x7_s1"
top: "pool5/7x7_s1"
dropout_param {
dropout_ratio: 0.4
}
}
layer {
name: "loss3/classifier"
type: "InnerProduct"
bottom: "pool5/7x7_s1"
top: "loss3/classifier"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
inner_product_param {
num_output: 1000
weight_filler {
type: "xavier"
}
bias_filler {
type: "constant"
value: 0
}
}
}
layer {
name: "loss3/loss3"
type: "SoftmaxWithLoss"
bottom: "loss3/classifier"
bottom: "label"
top: "loss3/loss3"
loss_weight: 1
}
layer {
name: "loss3/top-1"
type: "Accuracy"
bottom: "loss3/classifier"
bottom: "label"
top: "loss3/top-1"
include {
phase: TEST
}
}
layer {
name: "loss3/top-5"
type: "Accuracy"
bottom: "loss3/classifier"
bottom: "label"
top: "loss3/top-5"
include {
phase: TEST
}
accuracy_param {
top_k: 5
}
}
五、训练测试过程
训练是参照DisBelief【1】,使用异步随机梯度下降法,动量固定为0.9,学习率设置为每迭代8个epoches下降4%。图像采样方法在比赛中一直在变,没有详细的参照方法,但有一个小技巧被证实很有效:采样的图像块大小占原图的8%到100%,长宽比在3/4和4/3中随机选择,光尺度变化(photometric distortions)和若干差值方法也被使用。
在最后的测试过程中,作者用了三个tricks:
(1)融合了7个GoogLeNet版本的模型。这7个模型的权值初始化方式和学习率设置都一样,不同的是采样方法和图片输入顺序;
(2)在测试阶段,将测试图片重采样为256*256, 288*288, 320*320, 352*352 4个尺寸,分别取他们的左、中、右三个正方形部分(肖像图取上中下 三部分)。每一个正方形部分取四个角和中间的224*224部分,并将该正方形图像块重采样为224*224的大小,并取他们的镜像。这样一副测试 图像就可以扩充为4*3*6*2=144张测试图像块;
(3)最后模型融合时采用softmax输出平均的方法
六、实验结果
这里只列出分类的结果,检测的结果可以参考原文。
Table 3可以看出融合的模型越多,测试时裁剪的图像块越多,最后的测试结果越好。
参考论文:
【1】Ng A. Large Scale Distributed Deep Networks.[J]. Advances in Neural Information Processing Systems, 2012:1223-1231.
参考博客:
1、http://blog.csdn.net/shuzfan/article/details/50738394
2、http://blog.csdn.net/lynnandwei/article/details/44458033
3、http://www.cnblogs.com/Allen-rg/p/5833919.html
4、http://blog.csdn.net/whiteinblue/article/details/43635575/
5、http://www.eepw.com.cn/article/201705/359304.htm
博客中哪里写的不对的,欢迎批评指正。