extract_representation_cpu was broken after latest changes, I fixed it, probably not all the latest functionality will work, but I successfully tested it on feature extraction from imagenet/CLS_net_20140621074703.
second commit: I think it'll be useful to have cpu implementation, as you did in deepnet.
In this case we don't need to have separate code for feature_extractor (just switch between matrix class implementations), can run on cpu, use it for tests, etc.
So I ported eigenmat from deepnet (with minimal changes) and slightly unified CPUMatrix and matrix interfaces.
Hello, @nitishsrivastava
extract_representation_cpu was broken after latest changes, I fixed it, probably not all the latest functionality will work, but I successfully tested it on feature extraction from imagenet/CLS_net_20140621074703.
second commit: I think it'll be useful to have cpu implementation, as you did in deepnet. In this case we don't need to have separate code for feature_extractor (just switch between matrix class implementations), can run on cpu, use it for tests, etc. So I ported eigenmat from deepnet (with minimal changes) and slightly unified CPUMatrix and matrix interfaces.