1.数据集准备
ImageNet: http://www.image-net.org/
CIFAR10/CIFAR100: http://www.cs.utoronto.ca/~kriz/cifar.html
MNIST: http://yann.lecun.com/exdb/mnist/
Labeled Faces in the Wild:http://vis-www.cs.umass.edu/lfw/
more dataset: http://deeplearning.net/datasets/
==================================================================================================
2.深度学习竞赛及典型网络模型
ILSVRC(ImageNet Large Scale Visual Recognition Challenge)比赛;
网络:
(1)Alexnet(Alex Krizzhevsky)
(2)LeNet-5(Yan LeCun) http://yann.lecun.com/exdb/lenet/
(3)CaffeNet,GoogLeNet,VGG
Alexnet,LeNet-5,CaffeNet,GoogLeNet 在Caffe有;
VGG:http://www.robots.ox.ac.uk/~vgg/
==================================================================================================
3.深度学习框架
Big 3:Caffe,Torch,Theano;
TensorFlow(Google),CNTK(MS),SysytemML(IBM),Wills(Samsung),FAIR modukle on Torch(FB);
Caffe(http://caffe.berkeleyvision.org/):
是一个不需要编程的框架,只需要使用.prototxt 文件去描述网络;
train models:
./build/tools/caffe train -- solver=models/bvlc_reference_caffenet/solver.prototxt
use GPU:
./build/tools/caffe train -- solver=models/bvlc_reference_caffenet/solver.prototxt -gpu 0
Fine-Tune in Caffe:
对已经训练好的模型加上已经得到的权值参数等再进行重新训练;
Caffe Model Zoo:http://github.com/BVLC/caffe/wiki/Model-Zoo
Places-205:http://places.csail.mit.edu/demo.html
Caffe on SuperVesse 用户指导l:http://services.ptopenlab.com/mediawiki/index.php/
Torch:http://torch.ch/ 科学计算框架
Theano(http://deeplearning.net/software/theano/)python科学计算库
TensorFlow(http://www.tensorflow.org/)
谷歌开源计算框架;数据流图,节点计算单元,边是输入输出的关系;C/C++/Python API;ML && DL;
MNIST数据集的训练例子:http://www.tensorflow.org/versions/ro.7/tutorials/mnist/beginners/index.html#minst-for-ml-beginners
对比:
灵活性:Theano?TensorFlow>Torch>Caffe
初学者:Caffe>Theano>Torch(可以分布式运算)>TensorFlow
深度学习可视化工具DIGITS:
==================================================================================================
4.深度学习实验与实践