search

dongbo811 / AFFormer

119 stars 12 forks source link

代码 #2

Closed zuo-ux closed 1 year ago

zuo-ux commented 1 year ago

这个与AFFormer有关的代码不完整吧

dongbo811 commented 1 year ago

Would you like to describe in detail?

zuo-ux commented 1 year ago

对的，想仔细看看里面的细节！

dongbo811 commented 1 year ago

This repository is the completed. Can you describe your problem in detail?

Haoru commented 1 year ago

Parallel Heterogeneous Architecture的代码能否公开一下？

dongbo811 commented 1 year ago

You can refer to ./tools/afformer.py

If you have more questions, please open this issue.

Haoru commented 1 year ago

冒昧问一下，论文中Clustering是哪一部分代码？

dongbo811 commented 1 year ago

Considering that traditional clustering methods, such as Kmeans, requires multiple iterations, which greatly increases computational complexity, we adopt a simple local clustering strategy. Each local cluster center is initialized with its corresponding region by Equation 1, and the initialization weights can be further optimized during training. Considering the computational complexity and actual running speed, we currently rely on convolution to simplify the implementation of this strategy, which can also be implemented in other similar ways. The code can refer to afformer.py#L193.