Yolov3: training without pre-trained weight option? (train from scratch)

[Sibozhu]

Hi @glenn-jocher thank you so much for this amazing repo on YoloV3. Right now I'm trying to train with my own custom cfg file, therefore, I'm wondering if there's a way to train from scratch instead of using any other pre-trained weights?

[glenn-jocher]

@Sibozhu yes of course. Training from scratch is the simplest option. The weights and biases are initialized at random, though constrained to certain statistical distributions by PyTorch based on the layer type. To skip loading of a pretrained backbone simply comment out these lines:

https://github.com/ultralytics/yolov3/blob/c1c09eb3ccf18c74ddf0c4217ab306e0a60c0775/train.py#L70-L76

[hxy1051653358]

0/99 156/156 0.253 0.723 19.8 2.07 22.9 6 2.84 0 0 0
1/99 156/156 0.248 0.492 2.02 1.81 4.57 10 0.284 0 0 0
2/99 156/156 0.244 0.444 1.96 1.72 4.37 13 0.267 0.000337 0.000472 0.000337
Training from the beginning, is this result normal?

[glenn-jocher]

@hxy1051653358 looks fine. Keep in mind that YOLOv3 needs about 250-300 epochs for full training, so 2 epochs won't get you any usable results.

[riyaj8888]

0/99 156/156 0.253 0.723 19.8 2.07 22.9 6 2.84 0 0 0
1/99 156/156 0.248 0.492 2.02 1.81 4.57 10 0.284 0 0 0
2/99 156/156 0.244 0.444 1.96 1.72 4.37 13 0.267 0.000337 0.000472 0.000337
Training from the beginning, is this result normal?
where do i find that train.py file ..