你好，我在使用您的模型在ADE20K上进行训练时，发现达不到您在论文中的mIoU指标，发现有10个点的下降

[InMyDreammer]

这是我的日志，mIoU只有31.31，但在您的论文的Table1中mIoU却是41.8 20231028_132107.log

[dongbo811]

Thank you very much for trying AFFormer. From your log, I find that it is mainly caused by the mismatch between the model and the pre-trained weights. Please check the version network and pre-trained weights. If you have any questions, please feel free to contact~

在 2023-10-28 17:28:05，"InMyDreammer" @.***> 写道：

这是我的日志，mIoU只有31.31，但在您的论文的Table1中mIoU却是41.8 20231028_132107.log

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

[InMyDreammer]

我使用您的Imagenet1k作为预训练权重，是否应该使用在ade20k上的预训练权重呢？发自我的 iPhone在 2023年10月28日，17:35，TimZ @.***> 写道： Thank you very much for trying AFFormer. From your log, I find that it is mainly caused by the mismatch between the model and the pre-trained weights. Please check the version network and pre-trained weights. If you have any questions, please feel free to contact~

在 2023-10-28 17:28:05，"InMyDreammer" @.***> 写道：

这是我的日志，mIoU只有31.31，但在您的论文的Table1中mIoU却是41.8 20231028_132107.log

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

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

[InMyDreammer]

Thank you very much for trying AFFormer. From your log, I find that it is mainly caused by the mismatch between the model and the pre-trained weights. Please check the version network and pre-trained weights. If you have any questions, please feel free to contact~ 在 2023-10-28 17:28:05，"InMyDreammer" @.> 写道： 这是我的日志，mIoU只有31.31，但在您的论文的Table1中mIoU却是41.8 20231028_132107.log — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you are subscribed to this thread.Message ID: @.>

我仔细核对了一下，我发现我使用的是base版本的AFFormer，加载的权重也是base版本的AFFormer

[InMyDreammer]

您可以再看一下吗？因为我核对了很多遍版本是没有问题的，使用的是base版本的AFFormer，预训练文件也是base版本的Imagenet1K

[dongbo1998]

Sorry, I have re-updated the pre-training weights. Please download and try again. If you have more questions, please contact me.

[InMyDreammer]

Sorry, I have re-updated the pre-training weights. Please download and try again. If you have more questions, please contact me.

我尝试了您新的预训练权重，但是mIoU结果只有39.12，距离您41.8还有差距，我查看了日志发现仍然有一部分参数不匹配：unexpected key in source state_dict: cls_head.cls.weight, cls_head.cls.bias，您确定您上传的预训练权重没有问题吗？ 这是我的训练日志 20231029_115047.log

[dongbo1998]

First of all, it is normal for cls-related parameters not to match. The pre-training weights are obtained by training on imagenet-1k. There will be a classification layer and there is no need to load it when doing segmentation.

Secondly, there are two reasons why the training did not achieve the results of the paper:

1. Environmental issues. According to previous experience, segmentation tasks are sensitive to different environment versions, GPU and CUDA versions, etc., which will introduce certain uncertainties;
2. Parameter randomization problem. The initial parameters of each training are different, and the training effect will be different;

Since the experiment has been going on for a long time, there are currently not enough resources to reproduce it again, but I have saved the experiment log at that time. If you are interested, please add my WeChat (15316357479) and I will send it to you privately.