csjliang / LPTN

Official implementation of the paper 'High-Resolution Photorealistic Image Translation in Real-Time: A Laplacian Pyramid Translation Network' in CVPR 2021
Apache License 2.0
413 stars 47 forks source link

关于论文中金字塔层数与性能的问题请教! #2

Closed 24werewolf closed 3 years ago

24werewolf commented 3 years ago

在论文中有这样的描述: Take the task on 1080p images an an example, the PSNR of the LPTN is just reduced from 22.09 to 21.95 when the L is increased from 3 to 5, yet the model achieves a speed-up of more than x2 and takes about 1/16 of memory usage. 以及对应的table 里面也可以看出L = 5 时比L =3 时要快。这个时间的衡量,指的是整个推理的时间?按照常理理解不应该是L越大,层数越多,推理时间越慢!

csjliang commented 3 years ago

你好!金字塔层数越多,低频分量的分辨率越低,低频分量过网络所需要的处理时间越短(每多一层,低频分量的长宽分别减少一半)。因为本文关注低频(亮度、颜色等)相关任务,模型计算量集中在低频处理网络,因此层数越多,对推理时间增益越大。

24werewolf commented 3 years ago

你好!金字塔层数越多,低频分量的分辨率越低,低频分量过网络所需要的处理时间越短(每多一层,低频分量的长宽分别减少一半)。因为本文关注低频(亮度、颜色等)相关任务,模型计算量集中在低频处理网络,因此层数越多,对推理时间增益越大。

感谢回复,已经理解