MIoU is less

[chris2000520]

Why is the miou of each of my models 10% less than the table? I ran 200 rounds of training.

[zh320]

Hi, Chris. Please share the configs for your experiments.

[chris2000520]

Hi, Chris. Please share the configs for your experiments.

I used the same configs as you. I ran 200 rounds of epochs on BiseNetv2, but the best miou is 68.65%. The GPU is RTX 2080 Ti. There are some pictures about my result.

[zh320]

BiseNetv2 is not pretrained on imagenet dataset, so you need to train it more epochs (e.g. 800 epochs) to achieve better results. Actually you may switch this repo from main branch to archive branch, you will see the mIoU for BiseNetv2 is 69.36 (200 epochs) and 72.15 (800 epochs). Other tricks may increase the performances are: use larger crop-size (e.g. 1024) and use larger batch size.

Best

[chris2000520]

BiseNetv2 is not pretrained on imagenet dataset, so you need to train it more epochs (e.g. 800 epochs) to achieve better results. Actually you may switch this repo from main branch to archive branch, you will see the mIoU for BiseNetv2 is 69.36 (200 epochs) and 72.15 (800 epochs). Other tricks may increase the performances are: use larger crop-size (e.g. 1024) and use larger batch size.

Best

Thank you so much！It helps me a lot. I'll try it again.
Best wishes.

[zhy168]

Hi, Chris. Please share the configs for your experiments.

I used the same configs as you. I ran 200 rounds of epochs on BiseNetv2, but the best miou is 68.65%. The GPU is RTX 2080 Ti. There are some pictures about my result.

兄弟你用过这里面STDC的Detail head吗，训练无法收敛

[chris2000520]

没有-------- 原始邮件 --------发件人： zhy61718 ***@***.***>日期： 2024年6月27日周四 09:32收件人： zh320/realtime-semantic-segmentation-pytorch ***@***.***>抄送： 书灏 ***@***.***>, Author ***@***.***>主 题： Re: [zh320/realtime-semantic-segmentation-pytorch] MIoU is less (Issue #9) Hi, Chris. Please share the configs for your experiments. I used the same configs as you. I ran 200 rounds of epochs on BiseNetv2, but the best miou is 68.65%. The GPU is RTX 2080 Ti. There are some pictures about my result. 兄弟你用过这里面STDC的Detail head吗，训练无法收敛 —Reply to this email directly, view it on GitHub, or unsubscribe.You are receiving this because you authored the thread.Message ID: ***@***.***>