ResNet 20/32/44/56/110 for CIFAR10 with caffe
- get cifar10 python version, then create a soft link
ln -s cifar-10-batches-py
here - 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.
- use net_generator.py to generate
solver.prototxt
andtrainval.prototxt
, you can generate resnet or plain net of depth 20 44 56 110, or even deeper if you want. you just need to changen
according todepth=6n+2
- use train.sh to train it
- specify caffe path in cfgs.py and use plot.py to generate beautful loss plots.
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.