search

pluskid / Mocha.jl

Deep Learning framework for Julia
Other
1.29k stars 254 forks source link

add and accuracy layer for regression, add an example of regression #184

Closed mcreel closed 5 years ago

mcreel commented 8 years ago

Two files are added: a regression accuracy layer, and an example of regression. The regression accuracy layer reports RMSE for all outputs, as well as RMSE averaged over outputs. One file src/layers.jl is modified, to add support for the new regression accuracy layer.

pluskid commented 8 years ago

@mcreel Thanks for the PR! What is the difference between this new layer and the existing SquareLossLayer?

mcreel commented 8 years ago

Hi, This layer reports RMSE for each of the outputs, as well as the average over all outputs. In some cases, certain outputs may be more important than others, so a user might like to know RMSE for each output. So, it gives more output. Also, as far as I understand it, the proposed RegressionAccuracyLayer is analogous to AcurracyLayer, that is, it is a statistics layer, not a loss layer. Can the SquareLossLayer be used for computing accuracy of a test sample, as well as for loss in the training step? Best, Michael

On Fri, Feb 26, 2016 at 3:56 PM, Chiyuan Zhang notifications@github.com wrote:

@mcreel https://github.com/mcreel Thanks for the PR! What is the difference between this new layer and the existing SquareLossLayer?

— Reply to this email directly or view it on GitHub https://github.com/pluskid/Mocha.jl/pull/184#issuecomment-189313034.

pluskid commented 8 years ago

@mcreel Yes, it actually could act both as a loss layer and as a statistics layer. It report the "overall" mean squared error. So in the case of multi-output regression, the statistics here will be the same as the objective function value (unlike in the case of classification).

mcreel commented 8 years ago

I have tried out using SquareLossLayer as a statistics layer, and yes, it does work, which is good to know. However, when doing validation for a regression net with multiple outputs, the information given by SquareLossLayer is quite sparse. In my work, I would like to the the RMSEs for each output, which is what the proposed layer adds.