Zilong Huang
Zilong Huang
Yes, you are right. Have you run this repo with the Cityscapes IMG_MEAN. You are welcomed to report this result.
The performance is not stable. Maybe we can run 5 times for each setting and then compare the mean of them.
We haven't made such a comparison. For ImageNet classification 224x224, the computation costs of Swin and ours are not in the same order of magnitude. To some extent, it is...
Besides reducing the dim of Q and K, we use the multi-head self-attention rather than the 1-head self-attention used in the Non-Local block.
The inserted activation function is intended initially to increase non-linearity. However, we found that removing the activation function could achieve slightly better performance. I suggest removing it when using Topformer.
Your question is heavily related to the MMSegmentation framework and ONNX. It could be better to raise an issue under the corresponding repositories.
@vozhuo 我们有对一些TNN不支持的操作做了些处理，简单替换成相同功能和相同计算量的操作，比如改变softmax中dim等。 [tnn_models.zip](https://github.com/hustvl/TopFormer/files/9661282/tnn_models.zip) 压缩包里是SegFormer和Semantic FPN ConvMLP-S的tnnproto文件，希望对你有帮助。
Hi @inderpreetsingh01, thanks for your interest in our work. I saw you use the simple transform as the image preprocess, it is mismatched with the preprocess used in the Seaformer...
Hi @lindadamama, We hadn't tested our code on the Window system. We recommend using Unix-like/Linux system to run this code. For the mmcv-full installation on the Window system, you could...
Please refer to https://github.com/fudan-zvg/SeaFormer for the ImageNet training script.