GitHub - dailiang/cifar-10: Using caffe, the best accuracy on test set is 80%.

说明

CIFAR-10 分类模型(基于 caffe ), 在测试集上准确率可达 80%

###框架
3 层卷积层 + 2 层全连接层 + softmax

###新技术

所有非线性变换均采用 ReLU, 前两层卷积层后面加入 pool 和 local response nomalization, 两个全连接层均采用了 dropout 技术.

###模型细节

pad1-conv1-relu1-pool1-lrn1-pad2-conv2-relu2-pool2-lrn2-conv3-relu3-fc4-relu4-dropout4-fc5-relu5-dropout5-softmax

###曲线

训练过程中的 error 和 acccuracy 曲线如下:

Name		Name	Last commit message	Last commit date
Latest commit History 10 Commits
img		img
.gitignore		.gitignore
README.md		README.md
README.md.html		README.md.html
TODO.md		TODO.md
accuracy.png		accuracy.png
cifar_solver.prototxt		cifar_solver.prototxt
cifar_solver_1.prototxt		cifar_solver_1.prototxt
cifar_test.prototxt		cifar_test.prototxt
cifar_test_1.prototxt		cifar_test_1.prototxt
cifar_train.prototxt		cifar_train.prototxt
cifar_train_1.prototxt		cifar_train_1.prototxt
convert_cifar_data.cpp		convert_cifar_data.cpp
create_cifar.sh		create_cifar.sh
error.png		error.png
train_cifar.sh		train_cifar.sh
train_cifar_1.sh		train_cifar_1.sh