Zilong Huang
Zilong Huang
@swjtulinxi 1.你的意思是训练过程中每个epoch都测试吧。这种操作不是必需的，速度会很慢。 2.在shell脚本里没法实现这个功能，需要在train.py里添加eval的逻辑。
没有做模型筛选，直接取网络最后的保存的pth。 NoneType一般是由于提供给opencv的路径不对导致的。 请确保dataloader是可以访问cityscapes的图像，你可以打印一下dataloader读取的路径是否正确，并敲代码试一下是否可以读取图像数据。
超参都是校验过的，没必要在代码里体现。若你有需要，可以自行添加校验逻辑。
It may be feasible. It's difficult to predict the result of the local and sparser attention. This is [ICCV17 paper](http://www0.cs.ucl.ac.uk/staff/I.Kokkinos/pubs/Adam_segmentation_embeddings_ICCV2017.pdf) about local attention, hope it helps.
The issue may occur when the machine is out of GPU memory. You can reduce the input size and try again.
Actually, It's a bug of package Inplace-abn. It needs to run with the context of torch.distributed.
@DotWang We do not have a plan to make it support mixed-precision training. It will be great if you can achieve it. Maybe there is an alternative way to achieve...
@mingminzhen Thanks for asking. They could achieve the same performance when both use OHEM. Without OHEM, the new version could achieve 78.5+ mIOU, which is lower than the previous one....
@yong-qiang 见[Issue 3](https://github.com/speedinghzl/CCNet/issues/3)
Hi @hustcc19860606 Maybe the [pure-python](https://github.com/speedinghzl/CCNet/tree/pure-python) or [>Pytorch 1.5](https://github.com/speedinghzl/CCNet/issues/94) could solve your problem.