cnn
===
This is a matlab-code implementation of convolutional neural network.
Functionality
---
* supported layertypes : 'conv', 'sigmoid', 'maxpool', 'meanpool', 'relu', 'tanh', 'softmax', 'stack2line', 'softsign'
* supported loss function : 'crossEntropy'
* supported training method : 'SGD'
* supported computing device : 'GPU', 'CPUonly'
* debug tools : deconvnet, display\_training, gradent\_check
* supported demo dataset : 'MNIST', 'GENKI-R2009a'
Usage
---
The structure of convolutional neural network is
conv pool [conv pool] stack2line ['nonlinear']
[] means optional, and can be replicated for many times.
Layer
---
### conv ###
implement convolution computing. To make codes flexible, I do not implemente non-linear functions after convlution. You can add a layer to complete the non-linear instead.
To use 'conv' layer, you should specify the following parameters:
**filterDim**
**numFilters**
**nonlineartype**
If the inputs has multimaps, then you may specify the connection table between the input maps and the output maps:
**conn\_matrix**
If you don't specify the connection table, then each output map is connected to all input maps.
### pool/pool ###
'maxpool' and 'meanpool' are both pooling layer. To use pooling layer, the following parameters should be specified:
**poolDim**
**pooltypes**
### relu/tanh/sigmoid/softmax/softsign ###
These four types of layers mainly do the non-linear function to the input.
y = max(0,x)
y = tanh(x)
y = 1/exp(-x)
y = softmax(x)
y = x/(1+abs(x))
To use them, the following parameters should be specified:
**size**
Besides, the softmax layer is usually used as output layer.
### stack2line ###
After convlution and pooling, the multi-dimention "outputs" usually are converted to a vector to be used as the inputs of the densely connected non-linear layers. And stack2line layer is to indicate this converting.
Training Method
---
### SGD ###
Computing Device
---
### GPU ###
### CPUonly ###
Debug Tools
---
### deconvnet ###
### display\_training ###
### gradient\_check ###
Dataset
---
### MNIST ###
### GENKI-R2009a ###
没有合适的资源?快使用搜索试试~ 我知道了~
用CNN用来对自己的数据集分类
共29个文件
m:23个
train-images-idx3-ubyte:1个
t10k-images-idx3-ubyte:1个
4星 · 超过85%的资源 需积分: 50 578 下载量 87 浏览量
2016-06-02
09:26:38
上传
评论 29
收藏 11.08MB ZIP 举报
温馨提示
这个CNN工具箱只用改一两个地方就可以对自己的数据集分类了 比github上的deeplearning的工具箱里的CNN改动要简单
资源推荐
资源详情
资源评论
收起资源包目录
newcnn-master.zip (29个子文件)
cnn-master
DebugTools
display_network.m 3KB
computeNumericalGradient.m 1KB
samplePatches.m 733B
grad_check.m 679B
thetaChange.m 3KB
README.md 2KB
Dataset
MNIST
loadMNISTLabels.m 516B
train-labels-idx1-ubyte 59KB
t10k-labels-idx1-ubyte 10KB
loadMNISTImages.m 826B
train-images-idx3-ubyte 44.86MB
t10k-images-idx3-ubyte 7.48MB
layer
Test.m 4KB
cnnPool.m 2KB
cnnParamsToStack.m 1KB
TestPool.m 231B
nonlinear.m 988B
cnnConvolve.m 3KB
Demo
cnnTrain.m 3KB
configTestGradient.m 1KB
config.m 794B
LICENSE 18KB
cnnCost.m 8KB
TrainingMethod
minFuncSGD.m 2KB
cnnInitParams.m 3KB
Testing
test.m 102B
testThetaChange.m 651B
testGradCom.m 2KB
testInit.m 93B
共 29 条
- 1
元气少女缘结神
- 粉丝: 6068
- 资源: 65
上传资源 快速赚钱
- 我的内容管理 展开
- 我的资源 快来上传第一个资源
- 我的收益 登录查看自己的收益
- 我的积分 登录查看自己的积分
- 我的C币 登录后查看C币余额
- 我的收藏
- 我的下载
- 下载帮助
安全验证
文档复制为VIP权益,开通VIP直接复制
信息提交成功
- 1
- 2
前往页