search

dicarlolab / archconvnets

Architecturally optimized neural networks trained with regularized backpropagation
5 stars 5 forks source link

Codifying logs, and code to recreate them #2

Open ardila opened 11 years ago

ardila commented 11 years ago

I'm having some trouble getting the logs into the repo right now because of some git issues. I'll commit them as soon as possible.

mnist logs are in here (inside figure 2 or 3)

http://cs.nyu.edu/~wanli/dropc/dropnn-exp.tar.gz

cifar logs are in here. (something something run##_log.txt)

http://cs.nyu.edu/~wanli/dropc/cifar10-9_32.tar.gz

ardila commented 11 years ago

@dicarlolab/daseibert it looks like the mnist training procedure was done only on a few fully connected layers _on features extracted by a model trained on cifar!! Besides this being very interesting, I'm pretty sure that means that a) we should not test the mnist dataset, or its procedures b) this resulting extra computation time freed up by not doing this should be used by testing different parameters/procedures which have been used for cifar. We should codify alex's procedure: https://code.google.com/p/cuda-convnet/wiki/Methodology

And any other ones that we can find. If we could figure out any approximation of the methodology used for imagenet, then we should use that.

ardila commented 10 years ago

@yamins81 Do you have any code for automatically adjusting the learning rates?