Closed VikasRajashekar closed 1 year ago
Hello @VikasRajashekar
The num_classes
in the config file needs three places need to be changed, such as the line 23, line 94 and line 103 in this confg.
Besides the num_classes, you also need to change the training step
and max_iters
, according to the number of your training images to satisfy the 50 epoch.
Our model is based on MaskFormer framework, which can achieve the panoptic segmentation. In the box2mask, we only do the instance segmentation, so num_stuff_classes=0. We also explore the box2mask to achieve the panoptic segmentation recently.
Yes, the commands are same as the mmdetection.
@LiWentomng Thanks for the reply. I almost got it to run but facing the following issue.
"File "/opt/conda/lib/python3.8/site-packages/mmcv/ops/focal_loss.py", line 56, in forward
ext_module.sigmoid_focal_loss_forward(
RuntimeError: sigmoid_focal_loss_forward_impl: implementation for device cuda:7 not found."
what is mmcv version you used? I am using 1.5.0.
Any help to resolve this issue would be of great help.
@VikasRajashekar
It seems your installed mmcv-full is not compatible with your envs.
You can try to re-install mmcv-full by the following,
pip install -U openmim
mim install mmcv-full==1.5.0
@LiWentomng I am using the right version of mmcv I believe.
TorchVision: 0.10.0a0
OpenCV: 3.4.11
MMCV: 1.5.0
MMCV Compiler: GCC 9.3
MMCV CUDA Compiler: not available
MMDetection: 2.25.0+
Do I need to compile it?
Closing the issue as I am able to run after building it from the source.
First of all very cool work.
I would like to know for my dataset in coco format with just one class, I would like to try your model.
Do you recommend any changes other than num_classes?
Also what is num_things_classes=80,num_stuff_classes=0, in box2Mask?
Also the commands to run are the same as mmdetection right?