dongbo811 / afformer Goto Github PK

你好，CLS的decoder返回结果，还需要经过什么操作怎么才能转成语义分割中每一类的像素概率呢，我单独测试的模型返回结果有负数值，你能解答我的疑惑吗？谢谢

Hello, thank you for the excellent paper, I also try to reproduce the model in the paper, but when I load the model it will appear KeyError: "EncoderDecoder: 'afformer_base is not in the models registry'". I hope you can answer my questions in your busy schedule

Thanks for your great work. Are there possible configs of large models that still have competitive performance?

In the manuscript, you used the key k and value v to calculate the self-attention weight matrix. But most transformers used the query and key to calculate the weight matrix. There is no difference in codes, but the expression ( k and v for computing self-attention weight) is confusing. Could you give me an explanation in-depth?

excuse me, could you share the pretrained model in aliyunpan ?

@dongbo811 Thank you for your work! I trained 150000epoch on the model based on cityscape data as required, but the final miou was only about 74.23, which is still a little short of the 78.7 mentioned in your paper. Here is my training log file, could you please help me check the problem? Thank you very much!
20230802_150709.log

can't open file demo\image_demo.py': No such file or directory

Thanks for your fantastic work!
I am curious that how many GPUs (maybe V100?) were used in the training (because I am trying to explore some downstream tasks based on your model.)
I have checked out the configuration folder and found that samples_per_gpu is set to 2 for Cityscapes and 4 for ADE20K. You said that the batch size is set to 16 and 8 for Cityscapes and ADE20K respectively, so you probably used 4 V100?

mmseg - WARNING - The model and loaded state dict do not match exactly
size mismatch for stem.0.conv.weight: copying a param with shape torch.Size([32, 3, 3, 3]) from checkpoint, the shape in current model is torch.Size([16, 3, 3, 3]).
size mismatch for stem.0.bn.weight: copying a param with shape torch.Size([32]) from checkpoint, the shape in current model is torch.Size([16]).
size mismatch for stem.0.bn.bias: copying a param with shape torch.Size([32]) from checkpoint, the shape in current model is torch.Size([16]).
size mismatch for stem.0.bn.running_mean: copying a param with shape torch.Size([32]) from checkpoint, the shape in current model is torch.Size([16]).
好像是预训练权重文件 和模型不匹配，请问这个怎么修改呢？

您好，我使用了配置文件中的AFFormer_small_cityscapes.py并从下载了您提供的AFFormer_small_cityscapes.pth并将其加载到了配置文件中，为什么还是会有这个警告，并抛出StopIteration异常？




非常渴望能获得您的帮助

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

Dear author, thank you for your contribution. I have a question: the size of the four stages of the deformer has been reduced by half in sequence. In contrast, the stages 2-4 of this model are all H/8 * W/8 in size. What are the considerations for this design？

后续还会继续更新、优化吗

这个与AFFormer有关的代码不完整吧

Hi@dongbo811 ，
You mentioned parallel heterogeneous architecture in your article. In which aspect or module is the concept of heterogeneity reflected? Does it refer to the cluster center of each grid on prototype G? Please point out the error in my understanding. Thank you

Hi, there. Thanks for your wonderful work.
I have a question about the mmsegmenation version when cuda11.3 pytorch 1.10.2 and mmcv==1.5.0
Looking forward to your reply as soon as possible.
Thank you in advance.