These are chat archives for beniz/deepdetect

7th
Nov 2016
ProFire
@ProFire
Nov 07 2016 06:02

Hey Beniz. I'm here to ask questions again. I'm still on xgboost trying to do regression based on xgb-tutorial. I'm also following the API ref to modify the curl calls to do regression instead of classification. Now I'm encountering an error I'm quite clueless.

value 0for Parameter num_class should be greater equal to 1

I can't find in the API ref on num_class. What is num_class and what should be the value?

Emmanuel Benazera
@beniz
Nov 07 2016 06:25
hi @ProFire you'd need to put your API calls into a gist so that I can review them
Emmanuel Benazera
@beniz
Nov 07 2016 06:58
your goal (regression) is not aligned with your model and loss (multi:softprob). See https://github.com/dmlc/xgboost/tree/master/demo/regression.
ProFire
@ProFire
Nov 07 2016 07:06
Ah, so that's what objective is. I shall give that a try and see how it goes.
ProFire
@ProFire
Nov 07 2016 07:20

I am now having a different error which I can't find information on the API ref.

negative supervised discrete target (e.g. wrong use of label_offset ?

Is this due to too little training data or I used label_offset wrongly?

Emmanuel Benazera
@beniz
Nov 07 2016 07:28
there's no discrete target in regression mode, and no need for an offset
bbl
ProFire
@ProFire
Nov 07 2016 07:50
Oh wow! It works! No errors from creating service to prediction. Except that the prediction result isn't what I'm expecting. I gave val1 12.8 to predict, it returned 12.798982620239 instead. When I gave val1 0.1, it returned 0.40083363652229.
Could I have misconfigured the training paramaters somewhere, or do I need to have a better understanding of xgboost before I optimise it to predict better?
I shall wait for your return.
Emmanuel Benazera
@beniz
Nov 07 2016 07:51
follow your current path, loom at the loss, etc
getting a model right takes time and trials. Enjoy ^^
ProFire
@ProFire
Nov 07 2016 07:56
Ahaha... I know what you mean... I'm guessing I should take things from here. At least I have made a functional prediction. Perhaps a guide before I spread my wings and fly on my own?
Or any parting words of wisdoms?
ProFire
@ProFire
Nov 07 2016 07:59
Ahaha... Alright then. I shall try to live on... =P
Thanks for all the assist so far!
Emmanuel Benazera
@beniz
Nov 07 2016 08:00
ProFire
@ProFire
Nov 07 2016 08:01
Alright!!!!!! Now that's something more exciting to carry forward!
=D
Emmanuel Benazera
@beniz
Nov 07 2016 08:01
that's OK, some days you could trade Blade Runner for Andrew Ng class, just not always :)