Discussion about training models

[stavros-k]

Hey, first of all thanks for bringing this app! I'm hopping this question isn't answered anywhere in the wiki (i didn't find anything atleast).

While i want to contribute, my coding skill level is far far behind. I'm fairly ok doing simple (very simple) apps but not at this level. But i'd like to try to contribute on the model. Is there anyway we could train the model and contribute it here? In a way to expand the model and make it even more accurate (or maybe faster?). In case there IS a way to do that, if i use my photos, is there a way to "reverse engineer" the model and recreate those photos? If not, i have a ton of photos to let the model train from!

[matiasdelellis]

Hey, first of all thanks for bringing this app!

Hey @stavros-k Thanks for the words of support.. :grimacing:

I'm hopping this question isn't answered anywhere in the wiki (i didn't find anything atleast).

No .. It's not anywhere, that's why I'm going to leave it open until I document it. :sweat_smile:

While i want to contribute, my coding skill level is far far behind. I'm fairly ok doing simple (very simple) apps but not at this level.

Don't worry, sometimes the reports are so much more important than the development.. :wink:

But i'd like to try to contribute on the model. Is there anyway we could train the model and contribute it here? In a way to expand the model and make it even more accurate (or maybe faster?).

Information about how to train the model is widely available. Here you can read the official announcement.

• http://blog.dlib.net/2017/02/high-quality-face-recognition-with-deep.html

You can read that the current model was trained with 3 million images. Well, the current model has its problems, the author says that it is biased in terms of race, or age, but to improve it, we have to obtain at least the same amount of images, with a better distribution of people and it is very unlikely.. :disappointed:

On the other hand, you need hardware according to the task, which I personally cannot afford to buy. :disappointed: :cry:

In case there IS a way to do that, if i use my photos, is there a way to "reverse engineer" the model and recreate those photos?

The code is here:

• https://github.com/davisking/dlib/blob/master/examples/dnn_face_recognition_ex.cpp

If not, i have a ton of photos to let the model train from!

You can have a lot of photos, but you can only share YOUR photos. Not those of your family or friends without their consent.. Training images are a very sensitive subject. :thinking:

[stavros-k]

@matiasdelellis Thanks for taking the time to reply! I do have an nvidia 1060. but now i'm thinking it.. this not gonna do much. Is one thing to use the model and another to train it :P

Don't worry, sometimes the reports are so much more important than the development.. 😉

I'll throw in any ideas that might come to my mind that will improve the UX/UI :) !!