These are chat archives for beniz/deepdetect

12th
Jul 2016
Matt Zeiger
@mattzeiger
Jul 12 2016 06:01
hi @beniz is there an existing or known issue with dede not compiling with the newest version of protobuf ?
Emmanuel Benazera
@beniz
Jul 12 2016 06:16
not that I know of but you may want to post the out put here or in an issue
Emmanuel Benazera
@beniz
Jul 12 2016 19:46
@Isaacpm see #158 and FYI I've been able to reproduce the near inf row in the confusion matrix. This happens when a class was not predicted at all across all testing examples. I will check more deeply and see whether there's a way to avoid this (it's basically a division by near 0). In all cases, it is a bad sign for the classifier, meaning the class is never predicted.
danielgollas
@danielgollas
Jul 12 2016 21:53
@beniz I added a comment on #157, I don't think the proposed output format is valid JSON.
Emmanuel Benazera
@beniz
Jul 12 2016 21:54
yeahhh.... I should sleep more...
danielgollas
@danielgollas
Jul 12 2016 21:54
hahaha, sorry
Emmanuel Benazera
@beniz
Jul 12 2016 21:54
so I actually had an alternative
my problem is that I can't have an array of array of double easily
danielgollas
@danielgollas
Jul 12 2016 21:55
aah
Emmanuel Benazera
@beniz
Jul 12 2016 21:55
I need to have a key for each line
danielgollas
@danielgollas
Jul 12 2016 21:56
It's kind of hacky but how about this:
"cmfull": [
{"label1":[1,0,0,0]},
{"label2":[0,1,0,0]},
...
]
Order is enforced by the top level array, and you keep the label next to it's row
Emmanuel Benazera
@beniz
Jul 12 2016 21:57
thats what I had yes
danielgollas
@danielgollas
Jul 12 2016 21:57
You had an array of key value pairs
I just added the { } to make objects around the pairs
Or was that always in the code and the error was just in the github comment?
Emmanuel Benazera
@beniz
Jul 12 2016 21:59
ah ok yes, mine was a bit different. can you add it to the ticket ?
there s an internal data structure that is converted to Json and maybe the converter is limited, I need to look at it
wait I believe the code is correct and my comment not
I'm on my phone but if you look at the unit test the output looks good
sorry about that
danielgollas
@danielgollas
Jul 12 2016 22:06
Aaah, correct. If the tests are passing against "\"cmfull\":[{\"zero\":[0.5,0.5,0.0,0.0]},{\"one\":[0.0,1.0,0.0,0.0]},{\"two\":[0.0,1.0,0.0,0.0]},{\"three\":[2.696539702293474e308,2.696539702293474e308,2.696539702293474e308,2.696539702293474e308]}]" then all seems good.