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 |
91.4% | 0.9125 | |
92.48% | 0.9248999999999999 | |
ResNet-44 | % | 0.9283 |
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+2How 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.