[15:13:33] <leila>	 miriam_: o/
[15:14:02] <leila>	 miriam_: do you know if we keep the deleted images in Commons forever or for a short time?
[15:31:44] <leila>	 isaacj: o/
[15:32:09] <leila>	 isaacj: do we have the finalized list of questions for wtwrw on meta? I'm being asked about the specifics. 
[15:38:07] <isaacj>	 leila: not the wording but the demographics yes. at this point, I see no reason to add/remove any questions, so I'd feel comfortable adding the English text for each question if you think that'd be useful
[15:43:18] <leila>	 isaacj: if you can do that, that'd be great. I have a question from two chapters, and it would be best if I can link them to the page on meta.
[15:44:08] <isaacj>	 leila: great -- current descriptions are here ( https://meta.wikimedia.org/wiki/Research:Characterizing_Wikipedia_Reader_Behaviour/Demographics_and_Wikipedia_use_cases#Reader_Surveys ) and i'll add the question text to that section today
[15:48:46] <apergos>	 if you mean deleted as in a sysadmin on the project deletes them, the image is kept forever in case it's decided to undelete, except for legal issues (I'm sure you can think of some) and those get blasted with fire never to return
[16:00:23] <miriam_>	 ciao leila, sorry I was in meeting :) thanks apergos for your detailed reply! :)
[16:00:41] <apergos>	 sure thing
[16:54:03] <isaacj>	 fyi leila: question text/answers up: https://meta.wikimedia.org/wiki/Research:Characterizing_Wikipedia_Reader_Behaviour/Demographics_and_Wikipedia_use_cases#English-language_demographics_survey_questions
[18:31:03] <leila>	 isaacj: thanks for the questions on wiki. 
[19:13:17] <ebernhardson>	 anyone have a tensorflow model of at least medium complexity put together with a performant data input pipeline? Looking to test the GPU with something that wont be bottlenecked on preparing input data
[23:14:41] <miriam>	 ebernhardson I might be able to give you a model to finetune with few thousands images?
[23:21:30] <ebernhardson>	 miriam: sure, i suppose it doesn't matter that much what it is, i just wanted something to run on cpu from stat1007 and gpu from stat1005 to compare both runtime performance, and loss function performance
[23:29:34] <ebernhardson>	 one thing i was hoping for was to use something with tfrecords though. In tests on an nvidia card i have at home i've seen ~5x runtime perf improvement using tfrecord + dataset, rather than feed_dict + placeholder.