Lv Yang

Results 10 issues of Lv Yang

My train loss has always decrease, but the val_mIoU decreased at the first time, then it began increase. ![val_mIoU](https://user-images.githubusercontent.com/91008145/194453445-1962823d-609b-456f-befd-5827f8aad37f.svg)

I found it diffcult to understand this two file. #1 local_configs/_base_/models/segformer.py #2 local_configs/_base_/models/segformer.py I knew that #1 means the base of #2, but I find the code in file 2:...

Geforce RTX 3060 must use torch11.0+, and this means I must use cuda11.3, and the minimum version of mmcv is 1.4.7 So I got this error: "AssertionError: MMCV==1.4.7 is used...

![image](https://user-images.githubusercontent.com/91008145/220935010-0a7f349d-6390-4cc5-90e3-5dd5239c9ed3.png) ![image](https://user-images.githubusercontent.com/91008145/220935131-2253b6a2-886d-48c4-bee9-f8fd3ff06d92.png) 参数量只有5.85M,这个不太正常太小了呀,请问一下我该从那个方向去排查一下?

**Describe the current behavior** 我在使用网络做分割任务时,CEloss的结果比BCEloss结果更好,请问这个是有什么讲究吗?因为我映像中二分类任务一般使用BCEloss,那么我用CEloss处理二分类任务是不是很奇怪?

我输入了自己的模板和测试图像,一直无法匹配。请问大佬模板图像有什么特殊的要求吗?注意到scale_test,noise_test,我将图1作为模板训练测试noise_test是正常的,但是图1放在scale_test就报错?请问这个可能是什么原因呢? ![train](https://user-images.githubusercontent.com/91008145/219575594-9727767b-7302-4f5c-8d56-469145613e45.png) ![test](https://user-images.githubusercontent.com/91008145/219575699-c18fa5f5-edb0-46b2-8300-6d3367a28f1f.png) @meiqua

I find these links are not valid('404' or 'Well... This page cannot be accessed') 1) www.dropbox.com/s/gpzm15ipyt01mis/DA_Seg_models.zip?dl=0 2) http://vllab.ucmerced.edu/ytsai/CVPR18/DeepLab_resnet_pretrained_init-f81d91e8.pth Could you please give me some suggestions? I'll appreciate it very mush!

Have you ever test the params of the net(dp3+,mobilenetv2)?I got 7.5M. But when I use the code in mmsegmentaion, I got 15M? I think the params is only related with...

非常感谢up主的分享,请问B0-B5权值文件能否直接使用官方提供的文件。 我注意到up主提供了B0-B3的预训练文件,剩余的B4,B5,我能否直接使用官方提供的文件? 期待回复,一直在线哈哈。 Thank you very much for the code. Can the B0-B5 weight file directly use the official file? I noticed that you provides the pre-training files of B0-B3....

如题,我训练官方voc数据集是正常的,但是使用自己的数据集训练,loss会减小至负值。 交叉熵损失我的理解是不可能有负值,请问Up主能不能提供一些解决思路?非常感谢。