Generalization problem

[SchweitzerGAO]

Description of trained model: 10 epoch on tusimple dataset, accuracy 95.5% I did some experiments on this trained model with my own images and images of CULane, it performed badly: my own image:

image in CULane I am wondering why it's performing like this. Is increasing the number of epoch useful to solve this?

[SchweitzerGAO]

Is this something like overfitting?

[Turoad]

It is underfitting. Tusimple only has near 3000 training images and it only has highway scene. Training using Tusimple cannot get good generalization for other complex datasets.

[SchweitzerGAO]

It is underfitting. Tusimple only has near 3000 training images and it only has highway scene. Training using Tusimple cannot get good generalization for other complex datasets.

Thanks but still I am wondering how its performance can be improved. Is increasing the number of epochs or finetune the model with other datasets useful for this? By the way, are there any accessible pretrained models trained by CULane dataset?

[MySuperSoul]

You can find all pretrained models in here: https://github.com/Turoad/CLRNet/releases/tag/models

[SchweitzerGAO]

You can find all pretrained models in here: https://github.com/Turoad/CLRNet/releases/tag/models

Thanks, I will then test with the pretrained models to check if it's better

[SchweitzerGAO]

I tested with the pretrained model(DLA34 CULane), it still performed badly...... on Tusimple There's even no output on my own image: I am wondering why it is performing like this. Is there any advice to improve the generalization performance?

[MySuperSoul]

I test tusimple with culane-pretrained model with command as this:

python main.py configs/clrnet/clr_resnet18_tusimple.py --load_from culane_r18.pth --validate --view --gpus 0

It seems that it performs quite good, you can have a try like this.

And also you may to decrease the conf threshold in config, since you test one new dataset with culane-pretrained model, the conf score usually will lower than normal.

[SchweitzerGAO]

I will try this and thanks! By the way, how did you test the image outside the dataset with main.py? I think it didn't provide this function. I just test this with detect.py in the lanedet repo.

[MySuperSoul]

I will try this and thanks! By the way, how did you test the image outside the dataset with main.py? I think it didn't provide this function. I just test this with detect.py in the lanedet repo.

I haven't test image outside the dataset yet, but I think the detect.py in lanedet is also useful~

[SchweitzerGAO]

Oh, I think I tested images from the Internet, but it performed badly. I am wondering how the performance on images outside the dataset can be improved

[SchweitzerGAO]

I tried this command: python main.py configs/clrnet/clr_resnet18_tusimple.py --load_from culane_r18.pth --validate --view --gpus 0 but it failed with the error I wonder why it happened

[MySuperSoul]

I tried this command: python main.py configs/clrnet/clr_resnet18_tusimple.py --load_from culane_r18.pth --validate --view --gpus 0 but it failed with the error I wonder why it happened

Just modify to num_classes = 4 + 1 in config clr_resnet18_tusimple.py

[SchweitzerGAO]

OK, I will do this and test 

------------------ 原始邮件 ------------------ 发件人: "Yifei @.>; 发送时间: 2022年5月15日(星期天) 下午2:25 收件人: @.>; 抄送: @.>; @.>; 主题: Re: [Turoad/CLRNet] Generalization problem (Issue #17)

I tried this command: python main.py configs/clrnet/clr_resnet18_tusimple.py --load_from culane_r18.pth --validate --view --gpus 0 but it failed with the error I wonder why it happened

Just modify to num_classes = 4 + 1 in config clr_resnet18_tusimple.py

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: @.***>

[Turoad]

@SchweitzerGAO You can refer this issue https://github.com/Turoad/CLRNet/issues/19.

[SchweitzerGAO]

Thank you

------------------ 原始邮件 ------------------ 发件人: @.>; 发送时间: 2022年5月15日(星期天) 晚上8:26 收件人: @.>; 抄送: @.>; @.>; 主题: Re: [Turoad/CLRNet] Generalization problem (Issue #17)

@SchweitzerGAO You can refer this issue #19.

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you were mentioned.Message ID: @.***>