Fine-tune in R: end-to-end example

[statist-bhfz]

This is tutorial mentioned in #5149

[matt32106]

@statist-bhfz out of sheer curiosity, how much accuracy do you get on your validation set?

[statist-bhfz]

@matt32106 It was around 86% on validation set and 83% on test set (public leaderboard) after 3 or 5 epochs. I'm still confused with R session crashes, so I can't perform more experiments.

[matt32106]

I had some trouble (R crashes) with the mean image because it needs to be the same size as data.shape, i don't know if this apply to your crashes!? (check issue #5214) I did not try public LB yet but i get about 95% accuracy with bn-28, error rate: http://pasteboard.co/FJjZNMJ8E.png

…

________________________________ From: Andrey Ogurtsov - [email protected] <git.m7.213feed42b.notifications#[email protected]> Sent: Sunday, March 5, 2017 10:45 AM To: [email protected] Cc: matt; Mention Subject: Re: [dmlc/mxnet] Fine-tune in R: end-to-end example (#5253) [email protected] exclusive) @matt32106<https://github.com/matt32106> It was around 86% on validation set and 83% on test set (public leaderboard) after 3 or 5 epochs. I'm still confused with R session crashes, so I can't perform more experiments. — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub<#5253 (comment)>, or mute the thread<https://github.com/notifications/unsubscribe-auth/AHNTYsSBrQ-MfbJ2JmgfwhTfijeuC7yiks5ripIugaJpZM4MTUmx>.

[statist-bhfz]

No, I had another issue because I didn't subtract mean image.