sjiang95
commented
1 year ago Reason found.
In PR #9078, a new entry seg_suffix is added for class CustomDataset. mmdetection/mmdet/datasets/custom.py at e71b499608e9c3ccd4211e7c815fa20eeedf18a2 · open-mmlab/mmdetection
def __init__(self,
ann_file,
pipeline,
classes=None,
data_root=None,
img_prefix='',
seg_prefix=None,
seg_suffix='.png', # new entry
proposal_file=None,
test_mode=False,
filter_empty_gt=True,
file_client_args=dict(backend='disk'))But in mmdet/datasets/cityscapes.py#L264, the initialization of class CocoDataset remains.
self_coco = CocoDataset(self.ann_file, self.pipeline.transforms,
None, self.data_root, self.img_prefix,
self.seg_prefix, self.proposal_file,
self.test_mode, self.filter_empty_gt)There is an arguments mismatch between the arguments of self_coco = CocoDataset in mmdet/datasets/cityscapes.py with the new __init__ in mmdet/datasets/custom.py. self.proposal_file in CocoDataset class in mmdet/datasets/cityscapes.py is assigned to seg_suffix in CustomDataset.__init__() in mmdet/datasets/custom.py.
I'll create a PR to fix this.
Prerequisite
Task
I'm using the official example scripts/configs for the officially supported tasks/models/datasets.
Branch
master branch https://github.com/open-mmlab/mmdetection
Environment
env0
limited by insufficient GPU memory, img resolution is tuned to
env1
On both env0&1, pytorch is installed by
Environment variables on env1:
$PATH
$LD_LIBRARY_PATH
$PYTHONPATH
Reproduces the problem - code sample
The cityscapes dataset is converted by 1: Inference and train with existing models and standard datasets — MMDetection 2.25.1 documentation
Single gpu (on env0):
Distributed (on env1):
Reproduces the problem - command or script
same as above
Reproduces the problem - error message
The error messages are same on env0&1.
Additional information
Training and evaluation on voc and coco are totally fine. For cityscapes, everything works like a charm in training stage, so I believe the dataset conversion has nothing to do with the error.