3
$\begingroup$

pjreddie's official darknet version (link from official website here) has been forked several times. In particular I've come accross AlexeyAB's fork through this tutorial. I assume the tutorial's author used AlexeyAB's fork because he wanted to use it on a Windows machine, which pjreddie's darknet cannot do AFAIK.

I am not really concerned about that (I am a linux user), but I am very interested about the half precision option (CUDNN_HALF) that AlexeyAB's darknet has, whereas pjreddie's darknet does not. Of course I've checked that this option was handled by the graphic card (RTX2080) we use at my office.

Nevertheless, I wonder: how stable/robust is that fork? Of course I want high-performing software, but I also want a certain level of stability! On the other hand, the latest commit on pjreddie's darknet is back from September 2018 (ie 1 year old), whereas AxeleyAB's darknet is active…

More broadly, there seems to be a lot of darknet forks: which ones to prefer?

What does the neural network community think?

Improve this question
$\endgroup$

1 Answer 1

Reset to default
3
$\begingroup$

I used Trieu's Darkflow and it was trained well up to around 30 to 50 epoch with maximum 30 GB dataset. At the same time, it was quite often crashed after these epochs.

Half precision...I don't remember Darkflow have that.

One way to estimate each fork reputation is by number of Star and Issues, for example Alexey's Darknet 5,000 and 2,500 and Darkflow 5,000, 539 respectively.

Reference: OS Ubuntu 16.04

Improve this answer
$\endgroup$
1
  • 1
    $\begingroup$ Thanks for sharing your user experience with darkflow! :) $\endgroup$
    – Gael Lorieul
    Sep 27, 2019 at 14:17

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .