Performance issue of images in Val set of Cityscapes

[scyyx1]

Hi,

I have used your provided pretrained model of DDRNet23 slim on Cityscapes dataset. However, for the validation set, I have only reached 41.49% mIoU instead of 77.8%. Did I forget to tune any parameters?

[sithu31296]

The mIoU results are from official papers and pretrained weights from official repo. I haven't tested each of them with my codebase. Though evaluation image size should be 1024x2048 for DDRNet. And I don't remember are there any more parameters you need to change. The results are only included for comparability of different models. Still 41.49 is too low.

[scyyx1]

I have tried both your provided one "ddrnet_23slim_city.pth" and the one in the official paper. They produced the same results, which are 41.49%. I've also tried to set the image to different size. For image size of [2048, 1024], I get lower accuracy, while if I change it to [1024, 2048], I received the same result.

[sithu31296]

Hello, sorry for late. I will evaluate the model and see if there is an issue with my implementation.

[scl666]

我已经尝试了您提供的“ddrnet_23slim_city.pth”和官方文件中的那个。他们产生了相同的结果，为41.49%。我还尝试将图像设置为不同的大小。对于[2048，1024]的图像大小，我得到的精度较低，而如果我将其更改为[1024，2048]，我会收到相同的结果。

你好，方便提供cityscapes数据集的DDR23slim的预训练权重吗？邮箱2358919283@qq.com

[scl666]

您好，很抱歉迟到了。我将评估模型，看看我的实现是否存在问题。

你好，请问我在使用DDR模型遇到这个错误‘RuntimeError: transform: failed to synchronize: cudaErrorIllegalAddress: an illegal memory access’，请教一下遇到这样子错误吗？网上搜有说pytorch版本，还有标注文件 / 标签 等没有设置正确 以及网络模型、训练数据（图片和标签）没有放置到GPU上，我感觉这些都没有问题啊。期待您的回复