After single-class training, nothing can be detected.

[MarkDana]

I've used this model to train for a single-class dataset, which contains 300 images. After 500 epochs of training, total loss has been reduced to a fairly low level (min=0.011), with precision up to 0.887.

However when I use the generated weights to detect images in the train-val set, nothing can be detected for each image. When running detect.py,

detections = model(input_imgs)
detections = non_max_suppression(detections, 1, opt.conf_thres, opt.nms_thres)
print(detections)

All images show [None].

Below is some changes I made to the original project:

• 1.in yolov3.cfg,

batch=16
subdivisions=4
filters=18(you know where..)
classes=1

• 2.in detect.py, to test the probability, I edited non_max_suppression(detections, num_classes=1, opt.conf_thres=0, opt.nms_thres=0), but nothing detected, still.

One of the reason may be that, the weights were initialized from yolov9.weights, and that may be different in network structure I edited in yolov3.cfg.

What's more, exploding gradient problem happened 3 times in the 500 epochs.

Anyone in same situation or can help me? thks.

[fourth-archive]

@MarkDana try this tutorial.
https://github.com/ultralytics/yolov3/wiki/Example:-Train-Single-Class

The https://github.com/ultralytics/yolov3 repository also includes a YOLOv3 realtime iPhone app:

[MarkDana]

@MarkDana try this tutorial.
https://github.com/ultralytics/yolov3/wiki/Example:-Train-Single-Class

The https://github.com/ultralytics/yolov3 repository also includes a YOLOv3 realtime iPhone app:

Thks and I've tried this tutorial. Training now, waiting to see the result.
What's more, the two bugs mentioned in issue127 also work for me.

[gfjiangly]

I had same questiones with me before.
I think the reason is that "save_weights" function in trian.py doesn't match "load_weights" function in detect.py. You can replace
model.apply(weights_init_normal) with
model.apply(weights_init_normal)
model.load_state_dict(torch.load(opt.weights_path))
and replace
model.save_weights("%s/%d.weights" % (opt.checkpoint_dir, epoch)) with
torch.save(model.state_dict(), "%s/%d.weights" % (opt.checkpoint_dir, epoch)) in train.py.

in detect.py, replace
model.load_weights(opt.weights_path) with
model.load_state_dict(torch.load(opt.weights_path))

[0merjavaid]

Use transfer learning as mentioned by @gfjiangly it wont show results otherwise

[MarkDana]

Both @gfjiangly and @fourth-archive 's method work well. I'll close this issue, thx.