spyflying
@RicardLee Since the download of renewed annotation from google drive always fails, could you kindly provide the link of Baidu Cloud? Thanks a lot for the nice work.
We have fixed the checkpoint of ReferIt and updated Baidu Drive link.
It seems to be a cudnn version error. Try cuda 9.0 & cudnn 7.0.3.
Here is the setting of our conda environment. Hope it would be helpful to you. ``` channels: - conda-forge - defaults dependencies: - _libgcc_mutex=0.1=main - attrs=19.3.0=py_0 - backports=1.0=py_2 - backports.functools_lru_cache=1.5=py_2...
It can be found at this link: https://download.openmmlab.com/pretrain/third_party/resnet50_v1c-2cccc1ad.pth
应该是因为有一条数据的标签超范围了，上面修改标签的那行代码是在合并左右curb，并没有对超出范围的数据做判断。可以把数据过一遍，超出范围的数据直接删掉。
It takes 4092MiB GPU resources when testing with only 1 image and 4546MiB GPU resources when testing with a batch of 16 images.
Training with anchor3dlane.py occupies about 11.26G GPU resources for batch_size=16. The total training time for 60k iterations on 4 A100 GPUs is about 12~13 hours.
你好，因为在验证集上bench时并不会用到这里的pickle文件，而是使用了原始标注，所以这里验证集上车道线的数目是多少其实没有关系。不过为了避免运行时报错，我更新了一下tools/convert_datasets/once.py，当超过8条车道线时，只保留前8条。
It seems that you can train normally for several iterations, but there was an error when saving the checkpoint. I feel that this error message is incomplete. Maybe you can...