train resnet imagenet with caffe

All models are trained on 4 GPUs with a minibatch size of 32.

resnet-18

coming soon

resnet-50

coming soon

resnet-32

This is a bottleneck architecture,
Since there's no strong data augmentation and 10-crop test in caffe, the results maybe a bit low.
test accuracy: accuracy@1 = 0.67892, accuracy@5 = 0.88164
training loss for resnet-32 is shown below:
the trained model is provided in release