search

torch / demos

Demos and tutorials around Torch7.
355 stars 301 forks source link

scale data and externalize potentially useful functions #6

Closed rlowrance closed 8 years ago

rlowrance commented 10 years ago

I notice that the SGD example as written gave terrible prediction accuracy. The cause is that the two input features are on very different scales. The age features averages about 31 and the is-female features averages around .5. Thus SGD is tricky to get to converge on the un-rescaled data, needed a very small step size and many iterations.

I rewrote the example to include scaling, adding the necessary explanations and code.

Also, I decided that some of the code could be reusable if it were in external modules, so I moved the relevant code to separate modules. Now there is a little kit for fitting multinomial logistic regression modules. For pedaggogical reasons, I left all the explanations of how to do logistic regression using the neural network package in the code.

Atcold commented 9 years ago

@soumith, what about this? Why isn't it merged yet? It looks nice...

soumith commented 9 years ago

@atcold because i dont have time to see if these changes still hold, or if they are outdated. can you review and tell me?

Atcold commented 9 years ago

@soumith, shall I clone the original repo and try to merge it locally? Otherwise, what do you mean by "see if these changes still hold"?

soumith commented 9 years ago

Yes thats what I meant. On Mar 21, 2015 19:17, "Alfredo Canziani" notifications@github.com wrote:

@soumith https://github.com/soumith, shall I clone the original repo and try to merge it locally? Otherwise, what do you mean by "see if these changes still hold"?

Reply to this email directly or view it on GitHub https://github.com/torch/demos/pull/6#issuecomment-84470276.