search

dvlab-research / PanopticFCN

Fully Convolutional Networks for Panoptic Segmentation (CVPR2021 Oral)
Apache License 2.0
391 stars 53 forks source link

About training on Cityscapes #30

Closed wudaxianzi closed 2 years ago

wudaxianzi commented 2 years ago

sorry to bother you , after reading all instructions and documentation about detectronc2 , I still don't know how to modify this project to train on cityscapes

  1. I've done followed dataset/README.md to separated Cityscapes and genarated (cityscapes_panoptic_train.json、cityscapes_panoptic_test.json...)
  2. But in this peoject/configs/...yaml ,there all about coco_train, so whether should I write another config (.yaml)? And I don't know where to modify to locate or link to my cityscapes? 3.About the init WEIGHTS, which should I download to use?(still /data/detectronc2modelzoo/basemodel/R-50-MSRA.pkl) Sorry again if I missed something. Please give me some tips to run this project on training cityscapes. THX a lot ><.
yanwei-li commented 2 years ago

Hi, thanks for your interest. Maybe you can refer to this issue #21.

  1. For the implementation, here is also a fork modified for Cityscapes dataset.
  2. We also provide the R50-based model trained on Cityscapes, you can find the model and metrics.
  3. For the initialized weights, the same Res50 model is used as in /data/detectronc2_modelzoo/basemodel/R-50-MSRA.pkl
wudaxianzi commented 2 years ago

Thank you for your response~ I refered to the issue#21 and tried to run, But it came a new error :
(error in modulated_deformable_im2col_cuda: invalid device function) I wonder whether it's because my CUDA_VERSION(10.1) is not supported your project and I need higer version like 10.2..

yanwei-li commented 2 years ago

Hi, please make sure the Detectron2 has been compiled normally, and the inference process of raw PanopticFCN is conducted normally.

wudaxianzi commented 2 years ago

I can run the demo.py successfully. And so far I tried a few things. I modified yaml DATALOADER_NUM_WORKERS to 0 and set num_gpus =1, then it works . I don't know why I can't use distributed training(my device has 4 1080Ti) ,if I change NUM_WORKERS > 0, it does't work..and come up a error that my process was terminated with signal SIGKILL. when starting training from iteration.

yanwei-li commented 2 years ago

It seems strange and could be an error in the running environment. Maybe you can try some other projects in Detectron2 and find out the reason.