AVX and SSE4 instructions. Pretty much any Intel or AMD cpu made after 2011 should be fine.
Alternately, you can re-complile dlib yourself with these features disabled: https://github.com/ageitgey/face_recognition/issues/11#issuecomment-287398611
That error means the program tried to call a CPU instruction that wasn’t supported by your CPU. Unfortunately that most likely means your computer’s CPU is too old to support the vector math instructions being used. Sorry about that.
That was a typo! I meant to say “training iterations”. Thanks for catching it.
I wrote this so long ago that I don’t remember if I meant ‘epochs’ or ‘batches’, but I’m assuming from context that I meant “several thousand more training batches” and not several thousand full epochs.
You could use a simple math function like mean error instead, but it would weight error for all 128 values the same.
An SVM could figure out if some of the numbers were more important than others and weight them more to maybe get a little higher accuracy.
You could just try it both ways and see what worked better for your data.
You almost always scale the image to some known size instead of only looking at the original size straight out the camera.
One common solution is to is scale the image into several different pixel sizes. This is called an ‘image pyramid’. Then you run this algorithm on each different-sized copy of the image. That eliminates…
For face recognition of a few people from 20 images each, CNNs as described here are probably overkill.
Given than you only have 20 pictures of each person, you’d have to split those groups up into sub-groups of say 15 training pictures and 5 evaluation pictures (to test your model) of each…
full_dataset.pkl just needs to be in the same folder where you run the script. That file contains the combined data sets.
I didn’t include the file originally because it was 5.5gb. But I managed to zip it down to about 450mb and upload it here:
Thanks, I fixed the /gizmos text :)
Yep, you need a pathPrefix. Like you said, you can use “/” if you want it for your whole site. But in my experience, doing that is risky because you might accidentally trigger the app for pages that your app doesn’t know how to handle.