search

nishnik / ANONS

ANOVA for neural networks
4 stars 11 forks source link

One,two and three-wayANOVA on fashion MNIST #13

Open skyrb opened 5 years ago

skyrb commented 5 years ago

Implemented fashion MNIST and ANOVA on the result w.r.t to learning rate

skyrb commented 5 years ago

please review

nishnik commented 5 years ago

You can use statsmodel for ANOVA. Example: https://plot.ly/python/anova/ As you have already set it up for Fashion MNIST, try doing the ANOVA on it.

12

skyrb commented 5 years ago

I have made changes so please review..

skyrb commented 5 years ago

please review,i think today is last day

nishnik commented 5 years ago

Hi, this looks good to me. We can see that the learning rate has more role to play than the number of layers. There is a minor glitch; you shouldn't run the ANOVA on Mean Accuracy instead on different epochs like either the 5th or the last epoch. Because the model with small learning rate hasn't learned much in the first few epochs and that get counted too. Ideally we should take the epoch with highest val accuracy for each set of experiment.

nishnik commented 5 years ago

And don't worry about KWoC evaluation. They will value your contribution rather than the number of PRs accepted.