faq.md

Problem you may face and how to solve

mask2coco.py

• RecursionError: maximum recursion depth exceeded in comparison
  • Go to your conda environment imantics package annotation.py, and change the code:

  def index(self, image):
  
      annotation_index = image.annotations
      category_index = image.categories
  
      if self.id < 1:
          self.id = len(annotation_index) + 1
  
      # Increment index until not found
      if annotation_index.get(self.id):
          self.id = max(annotation_index.keys()) + 1
  
      annotation_index[self.id] = self
  
      # Category indexing should be case insenstive
      category_name = self.category.name.lower()
  
      # Check if category exists
      category_found = category_index.get(category_name)
      if category_found:
          # Update category
          self.category = category_found
      else:
          # Index category
          category_index[category_name] = self.category

mmdetction

• All the things work previously, but suddenly get ModuleNotFoundError: No module named 'mmcv'
  • sometimes this happens when your computer didn't get the package from the right path. You can uninstall mmcv and mmdet and reinstall it again by following the step below:

  pip uninstall mmdet
  pip uninstall mmcv
  rm -rf ~/.local/lib/python3.7/site-packages
  pip install openmim
  mim install mmdet
  cd mmdetection
  python setup.py install
  

• ImportError: libGL.so.1: cannot open shared object file: No such file or directory
  • sudo apt-get update && sudo apt-get install libgl1
• RuntimeError: CUDA error: out of memory
  • There are some scenarios when there are large amount of ground truth boxes, which may cause OOM during target assignment. You can set gpu_assign_thr=N in the config of assigner thus the assigner will calculate box overlaps through CPU when there are more than N GT boxes.
  • Set with_cp=True in the backbone. This uses the sublinear strategy in PyTorch to reduce GPU memory cost in the backbone.
  • Try mixed precision training using following the examples in config/fp16. The loss_scale might need further tuning for different models.