在birdview_vihicles和logo数据集达不到预期

[nikei946]

与notebook完全一样的参数 训练中val的loss较大，达不到notebook中给的效果 python train.py -c 0 -p vehicles --head_only True --lr 5e-3 --batch_size 32 --load_weights weights/efficientdet-d0.pth --num_epochs 10 --save_interval 100

python train.py -c 0 -p vehicles --head_only False --lr 1e-3 --batch_size 16 --load_weights last --num_epochs 16 --save_interval 100

[zylo117]

1，不排除运气问题。 2，看看你的各种库的版本，我试过无数次因为这版本不同导致结果有大差异 3，我用colab跑的，你试试

[landskris]

I cannot reproduce the logo.ipynb tutorial results in colab either @zylo117 Which versions are you referring to?

! python train.py -c 0 -p logo --head_only True --lr 5e-3 --batch_size 32 --load_weights weights/efficientdet-d0.pth --num_epochs 10 --save_interval 100

! python train.py -c 0 -p logo --head_only False --lr 1e-3 --batch_size 16 --load_weights last --num_epochs 30 --save_interval 100

[zylo117]

What? Did evil company do something secretly? I'll check on it.

[zylo117]

I tested it. It is luck or something, lower than I did a few month before by a little bit but still way higher than yours.

notebook

I think it's unstable with these few epochs. It's only tutorials. By training more epochs, you can expect higher mAP.