These are chat archives for beniz/deepdetect

4th
Nov 2017
Emmanuel Benazera
@beniz
Nov 04 2017 07:43
the new docker gpu should be available
Gavin Gilmour
@gaving
Nov 04 2017 11:36
hey, bit a dumb question but does anybody know how I'd go about getting a corresp.txt file for the vgg_face model? vgg_face comes bundled with a 'names.txt' here for LFW but as I understand I need to convert that to a corresp.txt file for DD to then map my numeric categories i.e. "1013" to say, textual labels...
Emmanuel Benazera
@beniz
Nov 04 2017 11:39
Gavin Gilmour
@gaving
Nov 04 2017 11:39
Ah, so it's as simple as just numerically ordering them. Doh! Thanks.
Emmanuel Benazera
@beniz
Nov 04 2017 11:40
Hopefully...
Gavin Gilmour
@gaving
Nov 04 2017 11:40
I had a look at corresp_ilsvrc12.txt and worried about all the 'n01440764' stuff etc and thought it was more involved :)
Gavin Gilmour
@gaving
Nov 04 2017 11:50
works great, thanks again
Gavin Gilmour
@gaving
Nov 04 2017 12:52
FWIW wrote [and plagiarised :smile:] some words getting DD playing with VGG_face: http://gavin.coffee/2017/11/04/face-recognition/
Emmanuel Benazera
@beniz
Nov 04 2017 12:59
Great thanks, I could link to it from somewhere as external resource. Let us know how the fine-tuning goes, its easy to do with DD if this is the path you intend to follow.
Gavin Gilmour
@gaving
Nov 04 2017 13:10
To be honest, I'm a front-end guy and knew absolutely nothing about ML up until a few days ago. Just trying to get my head around stuff and I'll write some stuff up if I think it's of any help
The training/fine-tuning stuff looks quite complicated, just end up googling around for blogs or github issues for any help I can get :)
I've basically got a ton of images in work and wanted ultimately to build a nice front-end around uploading an image and showing matches/similarities... looks like DD fits my use case for that nicely
Facial images, sorry
(the elastic search blog post is great btw)
I'm thinking some sort of searchkit interface -> elasticsearch populated via DD for this, but I'll see how the fine-tuning goes