ResNet on CIFAR-10 with Caffe

GitHub - yihui-he/resnet-cifar10-caffe: ResNet-20/32/44/56/110 on CIFAR-10 with Caffe

Testing

~/caffe/build/tools/caffe test -gpu 0 -iterations 100 -model resnet-20/trainval.prototxt -weights resnet-20/snapshot/solver_iter_64000.caffemodel

Model	Acc	Claimed Acc
https://github.com/yihui-he/resnet-cifar10-caffe/releases/download/1.0/resnet20_iter_64000.caffemodel	91.4%	0.9125
https://github.com/yihui-he/resnet-cifar10-caffe/releases/download/1.0/resnet32_iter_64000.caffemodel	92.48%	0.9248999999999999
ResNet-44	%	0.9283
https://github.com/yihui-he/resnet-cifar10-caffe/releases/download/1.0/resnet56_iter_64000.caffemodel	92.9%	0.9303
ResNet-110	%	0.9339

Citation

If you find the code useful in your research, please consider citing:

@InProceedings{He_2017_ICCV,
author = {He, Yihui and Zhang, Xiangyu and Sun, Jian},
title = {Channel Pruning for Accelerating Very Deep Neural Networks},
booktitle = {The IEEE International Conference on Computer Vision (ICCV)},
month = {Oct},
year = {2017}
}

Training

#build caffe
git clone https://github.com/yihui-he/resnet-cifar10-caffe
./download_cifar.sh
./train.sh [GPUs] [NET]
#eg., ./train.sh 0 resnet-20
#find logs at resnet-20/logs

Visualization

specify caffe path in cfgs.py and use plot.py to generate beautful loss plots.

python plot.py PATH/TO/LOGS

Results are consistent with original paper. seems there’s no much difference between resnet-20 and plain-20. However, from the second plot, you can see that plain-110 have difficulty to converge.

How I generate prototxts:

use net_generator.py to generate solver.prototxt and trainval.prototxt, you can generate resnet or plain net of depth 20/32/44/56/110, or even deeper if you want. you just need to change n according to depth=6n+2

How I generate lmdb data:

./create_cifar.sh

create 4 pixel padded training LMDB and testing LMDB, then create a soft link ln -s cifar-10-batches-py in this folder. - get cifar10 python version - use data_utils.py to generate 4 pixel padded training data and testing data. Horizontal flip and random crop are performed on the fly while training.

Other models in Caffe