Re:Darknet – A neural network framework written in C and CUDA
by joshvm on Sunday October 28, 2018 @07:55AM
(#18320243)
I've used darknet quite a bit over the last few months. I wouldn't recommend using it for a serious project, unless you happen to need an extremely fast off-the-shelf object detector. The other exception is if you need a fast convnet running on a Pi - Tiny Yolo will do about 1.2FPS with an optimised fork (using nnpack). Other than that you could just buy a Neural Compute Stick. Yolo itself is pretty great - it's very fast and is accurate enough for a lot of things.
The original repo isn't really updated, and while AlexyAB's fork is much improved, it's still a pain to use.
- If you make mistakes, things fail silently. This is by far the biggest problem. Train/test is difficult to get right because it's very difficult to figure out where exactly you've messed up.
- Support for images is arbitrary. Although you can compile with OpenCV, there are internal glob functions which simply ignore certain image types (I had to recompile it with support for TIFF, for example).
- Bounding boxes are stored in an awkward format, which is easy to get wrong. It's referenced to the centre of the box, stored as a fraction of the image width.
- Logging is very basic. Alexey added a loss graph, but that's about it. If you restart training from a checkpoint, you only get a loss curve from where you restarted.
- Retraining on your own data can seem like dark magic. There's a lot of "copy this config file and edit these numbers" and if you get it wrong, you've wasted a day training.
If you need to use Yolo, I'd recommend looking at reimplementations in more mature frameworks like pytorch (e.g. https://eavise.gitlab.io/lightnet/)
Re:Darknet – A neural network framework written in C and CUDA
by petee on Sunday October 28, 2018 @07:37AM
(#18320203)
The License 'stack' made my morning, thank you:
"THIS SOFTWARE LICENSE IS PROVIDED "ALL CAPS" SO THAT YOU KNOW IT IS SUPER
SERIOUS AND YOU DON'T MESS AROUND WITH COPYRIGHT LAW BECAUSE YOU WILL GET IN
TROUBLE"
Re:Darknet – A neural network framework written in C and CUDA
by syntaxing on Saturday October 27, 2018 @10:04PM
(#18318760)
FYI, this is the same person(people?) that came out with the YOLO Objection Detection classfier.I haven't used Darknet before but the Tiny Darknet seems very interesting and I might use that in the future for my small projeccts
Re:Darknet – A neural network framework written in C and CUDA
by wpdev_63 on Sunday October 28, 2018 @01:56PM
(#18321879)
It would be great if these AI/neural frameworks targeted opencl so it would be platform agnostic. AMD/Intel/Arm makes so very cost effective GPUs that run everywhere.
Re:Darknet – A neural network framework written in C and CUDA
by jazoom on Sunday October 28, 2018 @02:50AM
(#18319450)
Not very useful but since it's his own project he can do whatever he likes. It looks like he's having fun, which is great. Except his beehive died. That kinda sucks.
Re:Darknet – A neural network framework written in C and CUDA
by lucb1e on Sunday October 28, 2018 @06:47AM
(#18320066)
That title is very non-descriptive or, at worst, misleading. I thought it would be about an Internet overlay network or something. Repository's description is Convolutional Neural Networks, so perhaps the title could be "Pjreddie/darknet: Convolutional Neural Networks"? (Not sure if pjreddie is supposed to be well-known, there is probably a reason OP added it to the title.)
The original repo isn't really updated, and while AlexyAB's fork is much improved, it's still a pain to use.
- If you make mistakes, things fail silently. This is by far the biggest problem. Train/test is difficult to get right because it's very difficult to figure out where exactly you've messed up.
- Support for images is arbitrary. Although you can compile with OpenCV, there are internal glob functions which simply ignore certain image types (I had to recompile it with support for TIFF, for example).
- Bounding boxes are stored in an awkward format, which is easy to get wrong. It's referenced to the centre of the box, stored as a fraction of the image width.
- Logging is very basic. Alexey added a loss graph, but that's about it. If you restart training from a checkpoint, you only get a loss curve from where you restarted.
- Retraining on your own data can seem like dark magic. There's a lot of "copy this config file and edit these numbers" and if you get it wrong, you've wasted a day training.
If you need to use Yolo, I'd recommend looking at reimplementations in more mature frameworks like pytorch (e.g. https://eavise.gitlab.io/lightnet/)
Anybody remember the resume with the ponies?
https://pjreddie.com/static/Redmon%20Resume.pdf
https://news.ycombinator.com/item?id=17500616
https://news.ycombinator.com/item?id=15060129
"THIS SOFTWARE LICENSE IS PROVIDED "ALL CAPS" SO THAT YOU KNOW IT IS SUPER SERIOUS AND YOU DON'T MESS AROUND WITH COPYRIGHT LAW BECAUSE YOU WILL GET IN TROUBLE"
[0] https://github.com/AlexeyAB/darknet
A particular favourite: https://www.youtube.com/watch?v=VOC3huqHrss
It apparently lacks any ergonomic scripting language bindings, which makes experimentation harder (than with tensorflow).
Or, if it does, it should list them right in the readme.