Darknet: how to use yolov4-tiny_contrastive.cfg?

Created on 30 Oct 2020 · 2Comments · Source: AlexeyAB/darknet

What is the new config file yolov4-tiny_contrastive.cfg?
When do we use it?
How do we use it?

I see the filters=... on line 413 is: (number_of_classes + 5) * 9. (The mask on line 419 contains 9 mask indexes, 0-8.)

So with the network I'm training with 1 class, I'm using filters=54.

But when I try to train with it, I get a segfault after 1017 iterations:

v3 (iou loss, Normalizer: (iou: 0.07, cls: 1.00) Region 52 Avg (IOU: 0.810298, GIOU: 0.805515), Class: 0.990291, Obj: 0.720371, No Obj: 0.000382, .5R: 0.987805, .75R: 0.804878, count: 82, class_loss = 2.362441, iou_loss = 170.592484, total_loss = 172.954926 
Command terminated by signal 11

> dmesg | grep darknet
[2482552.554782] traps: darknet[9947] general protection ip:5624934fa568 sp:7f35b68fa278 error:0 in darknet[562493468000+3aa000]

Source

stephanecharette

Most helpful comment

@stephanecharette

In general Contrastive learning (Contrastive-loss is an extended Triple-loss) is suitable for un-supervised learning and for multi-camera object tracking.
I just currently don't have enough time to implement it well to get top results.

This post is about this: https://github.com/AlexeyAB/darknet/issues/6004

There is experimental YOLOv4-tiny-contrastive model that is trained on MSCOCO:

./darknet detector demo cfg/coco.data cfg/yolov4-tiny_contrastive.cfg yolov4-tiny_contrastive_last.weights test.avi -out_filename out_test.avi -ext_output

AlexeyAB on 4 Nov 2020

🚀2 👍2

All 2 comments

@AlexeyAB can you comment on this one? Thanks.

stephanecharette on 2 Nov 2020