Doesn't seem like it. The GPU module required for training CNNs is sold separately[1]:
> A GPU capable special SOD release which is not available to the public is required in order to train your own CNN model.
There are some bits included for training "RealNet" models[2] but I'm not sure what those are or how they work. The documentation suggests that RealNet models can be trained in reasonable timeframes with a CPU.
"Q16.16" seems to refer to fixed floating point precision (https://en.wikipedia.org/wiki/Q_(number_format)). The parent comment is unhappy that SOD doesn't support floats with 16 integer bits and 16 fractional bits. I'm unsure why lack of support for 32 bit floats is such a problem considering that SOD is suppose to run on such resource constrained devices. Is there a particular domain where this is necessary?
> "Q16.16" seems to refer to fixed floating point precision
Floating is the opposite of fixed in this context. The GP is unhappy that it doesn't support fixed point numbers of a common fixed point number format.
This is not just nitpicking, because the difference in time and implementation cost between floating and fixed point can be huge on systems where floating point operations aren't offloaded to a pipelined FPU. Fixed point arithmetic on the other hand, even if your hardware doesn't support it directly, can easily be implemented in terms of integers.
I was interested until I saw you have to purchase something to train your own models. I am wondering how this compares to Darknet by Joseph Redmon [1]. You can do most of what I see on the Sod site, and train your own models. It's in C too.
Just a quick heads-up that Joseph Redmon isn't working on the development on Darknet any longer, and the YOLO fork by AlexeyAB [0] (now on YOLOv4) is one people should also consider using moving forward[1].
Honestly, I'm not sure. It seems like they cover slightly different fields, despite having some overlap. AFAIK, Darknet doesn't natively do a bunch of the image processing work that Sod can do, like edge detection with Canny, because if you wanted that, you'd just use OpenCV.
I'd imagine that based on the licensing, this is probably closer to the open-sourcing (well, sort of, since the commercial applications and GPU training requires a paid license) of an internal tool that the company is using on the backend of their own API-driven machine learning business. So, it covers some things that Darknet isn't really designed to do, since real-time object detection isn't the main area they're really interested in.
Real-time object detection is what YOLO does and darknet implements it, but I see where you are going with the edge detection in SOD vs. darknet. You add in OpenCV if you want to use it.
There is nothing to purchase at all if you want to train your own realnet models on the CPU such as Pedestrian, Car, etc. Just follow the documentation at https://sod.pixlab.io/api.html#sod_realnet_trainer on how to do so.
The site says "The syntax is based on the darknet framework...".
What are the differences in capabilities? Not sure yet. I'll have to look more into it, but I was hoping somebody here could tell me if they have experience with both.
Why would I use this over darknet and the variations of Yolo - YOLOv3 or YOLOv4?
Knowing that some schools and universities are testy to use GIMP because it’s a derogatory term, I’m wondering how “SOD” will fare. “Sod off” and “you’re a real sod” are not exactly terms of endearment and are very much still in current usage (particularly by the kind of people who might be offended by using something that refers to them).
Except it's much milder, there's no problem with saying "sod off" on tv during the day, for example. Apparently the British Communications Regulator, Ofcom[0], sees it as "Mild language, generally of little concern", whereas fuck is "Strongest language, unacceptable pre-watershed. Seen as strong, aggressive and vulgar. Older participants more likely to consider the word unacceptable."
It is indeed interesting! I had some use case in mind so I tried to find more about company. sysmic is owned by Pixlab, but both have zero info about people involved. Interestingly, even their linkedin page (../company/pixlabio) doesn't lists any individual and given the work they are doing, I doubt it's single person effort. Well, someone mentioned darknet project, looking into that instead.
I don't have an answer, but it's not always a bad sign. This guy was a powerhouse in the Ruby community, for instance, and was anonymous for as long as he was active there: https://en.wikipedia.org/wiki/Why_the_lucky_stiff
They are a commercial operation. If this is indeed a single guy, by obfuscating that, he can fool some people into thinking that they're dealing with a larger organization (and hence be more willing to hand over money).
30 comments
[ 4.3 ms ] story [ 73.7 ms ] thread> A GPU capable special SOD release which is not available to the public is required in order to train your own CNN model.
There are some bits included for training "RealNet" models[2] but I'm not sure what those are or how they work. The documentation suggests that RealNet models can be trained in reasonable timeframes with a CPU.
[1] https://sod.pixlab.io/cnn_train.html
[2] https://sod.pixlab.io/c_api/sod_realnet_train_start.html
Floating is the opposite of fixed in this context. The GP is unhappy that it doesn't support fixed point numbers of a common fixed point number format.
This is not just nitpicking, because the difference in time and implementation cost between floating and fixed point can be huge on systems where floating point operations aren't offloaded to a pipelined FPU. Fixed point arithmetic on the other hand, even if your hardware doesn't support it directly, can easily be implemented in terms of integers.
[1] https://pjreddie.com/darknet/
[0] https://github.com/AlexeyAB/darknet
[1] https://mobile.twitter.com/pjreddie/status/12538910781821992...
I'd imagine that based on the licensing, this is probably closer to the open-sourcing (well, sort of, since the commercial applications and GPU training requires a paid license) of an internal tool that the company is using on the backend of their own API-driven machine learning business. So, it covers some things that Darknet isn't really designed to do, since real-time object detection isn't the main area they're really interested in.
What are the differences in capabilities? Not sure yet. I'll have to look more into it, but I was hoping somebody here could tell me if they have experience with both.
Why would I use this over darknet and the variations of Yolo - YOLOv3 or YOLOv4?
1: https://www.urbandictionary.com/define.php?term=SO%20D
It comes from sodomy. It's interchangeable with fuck in many contexts.
[0] https://www.mirror.co.uk/news/uk-news/read-what-ofcom-thinks...
http://symisc.net/contact.html
Maybe they regard it as a hobby side project. Or maybe not everybody wants to attract public attention to all of their activities.