Loss can't be below 1.

[BlueAnthony]

The loss can not decrease under 1. It will stop and jitter around some number, like 5 to 6 or 12 to 13, when the iteration is around 50.
I already try different base learning rate, like 0.001, 0.0001, 0.00001, 0.000001.
And the loss start from about 1000.
I have 2 classes(Car and pedestrian), 3712 images for training and 3769 images for validation.
I use yolov3.weight as pretrained.

Thank you!!

---

I use the code from pjreddie/darknet and try to fine-tune with yolov3.weight.
The command I use is below.
"./darknet detector train cfg/kitti.data cfg/yolov3-kitti.cfg model/yolov3.weights"
Yes, I use random=1. My cfg is modified from yolov3.cfg of pjreddie/darknet.

And why I use these learning rate and steps?
It's because the yolov3.weights seems to remember the max iteration number, the max_batches for fine-tuning must be larger than 500200 and the fine-tuning just can be start.
The loss start about 1000 and stop decreasing about "500200+50" iterations.
Do I misunderstand something?

@AlexeyAB Really thank you for your patience.

[AlexeyAB]

Use these params:

darknet/cfg/yolov3.cfg

Lines 18 to 23 in e29fcb7

	learning_rate=0.001
	burn_in=1000
	max_batches = 500200
	policy=steps
	steps=400000,450000
	scales=.1,.1

And train about 2000 iterations.
If it doesn't help, then something wrong with classes number or with your dataset. Check it using this software: https://github.com/AlexeyAB/Yolo_mark

• Do you use random=1 in cfg-file?
• Do you use the latest version of this GitHub repository? https://github.com/AlexeyAB/darknet

[BlueAnthony]

@AlexeyAB When I change to use "AlexeyAB/darknet", "./darknet detector train cfg/kitti.data cfg/yolov3-kitti.cfg model/yolov3.weights" will directly save original model without training

[ghost]

Your learning rate is about 1e-8 which is too small. Try using the option of -clear, then the iteration will restart from 0, if you'd like to use yolov3.weights as pre-trained weights.

[BlueAnthony]

@panda9095 Thank you for your response.
Could you tell me more detail about "-clear"? When will "-clear" restart?

[ghost]

@BlueAnthony ./darknet detector train cfg/kitti.data cfg/yolov3-kitti.cfg model/yolov3.weights -clear

By doing so, the step number will start from 0 instead of 500200. Then you can use @AlexeyAB 's parameters for training.

[BlueAnthony]

@panda9095 Really thank you for your helping! I will try.

[AlexeyAB]

@BlueAnthony
Properly commands for training:

1. ./darknet detector train cfg/kitti.data cfg/yolov3-kitti.cfg model/yolov3.weights -clear

2. ./darknet detector train cfg/kitti.data cfg/yolov3-kitti.cfg darknet53.conv.74

3. ./darknet detector train cfg/kitti.data cfg/yolov3-kitti.cfg yolov3.conv.105
   Pre-trained file yolov3.conv.105 you can get by using this command:
   ./darknet partial cfg/yolov3.cfg yolov3.weights yolov3.conv.105 105

darknet/build/darknet/x64/partial.cmd

Line 21 in fb9fcfb

darknet.exe partial cfg/yolov3.cfg yolov3.weights yolov3.conv.105 105

[AbhimanyuAryan]

@AlexeyAB does this mean that I can further train my last trained model(on my dataset)....with new data?