mAP is 0% and the chart goes up and down...

[hechoel]

#1 - Installation information
CUDA-version: 11010 (12000), cuDNN: 8.9.7, CUDNN_HALF=1, GPU count: 1
CUDNN_HALF=1
OpenCV version: 4.6.0
0 : compute_capability = 890, cudnn_half = 1, GPU: NVIDIA GeForce RTX 4060 Laptop GPU
net.optimized_memory = 0
mini_batch = 1, batch = 64, time_steps = 1, train = 0

#2 - Command
darknet detector train data/projectname4/obj.data data/projectname4/yolov4-custom.cfg yolov4.conv.137 -map

#3 - cfg file
yolov4-custom.txt

#4 - data file
projectname4.zip

#5 - problem

A few things I tried to solve

1. Change from v3 to v4
2. Read and change anchors values
3. Recheck labeling image
4. Edit batch, subdivision size (64 -> 32/There is an issue where it turns off due to an unknown error)

Is the image the problem?..
What is the problem?
I need help..

[stephanecharette]

Please try the new Darknet/YOLO repo: https://github.com/hank-ai/darknet?tab=readme-ov-file#table-of-contents

[hechoel]

새로운 다크넷/욜로(Darknet/YOLO) 저장소를 사용해 보세요 https://github.com/hank-ai/darknet?tab=readme-ov-file#table-of-contents.

I training the tiny model with the new version, but the symptoms are the same.
Is it because there are not enough sample images?
[yolov4-tiny-custo
projectname6.zip
m.txt](https://github.com/AlexeyAB/darknet/files/15328169/yolov4-tiny-custom.txt)

[hechoel]

Please try the new Darknet/YOLO repo: https://github.com/hank-ai/darknet?tab=readme-ov-file#table-of-contents

"darknet detector train data/projectname6/obj.data data/projectname6/yolov4-tiny-custom.cfg -map"
The order is as follows:
Older versions include pre-training files. Is it that important?

[stephanecharette]

Join us on the discord server if you'd like more assistance, as these issue tickets aren't great for conversations. https://discord.gg/zSq8rtW

What would be best is if you can share your dataset so we can replicate your results.