如何反向传播训练Learnable Texture Extractor？

researchmm / TTSR

[CVPR'20] TTSR: Learning Texture Transformer Network for Image Super-Resolution

MIT License

765 stars 115 forks source link

Closed Achhhe closed 2 years ago

Achhhe commented 2 years ago

你好，计算attention时需要argmax，而argmax是不可导的，请问反向传播时，如果将梯度回传给LTE啊

FuzhiYang commented 2 years ago

argmax indeed cannot backpropagate gradients. The patches chosen by the argmax can backpropagate gradients