HTC customize data training bugs

I am using HTC model with customize dataset and get this error (mycode running on Gooble Colab and install mmdet with openmim). Please tell me how to fix this. Thanks RuntimeError: one of the variables needed for gradient computation has been modified by an inplace operation: [torch.cuda.FloatTensor [2, 256, 100, 136]], which is output 0 of ReluBackward0, is at version 4; expected version 0 instead. Hint: enable anomaly detection to find the operation that failed to compute its gradient, with torch.autograd.set_detect_anomaly(True).

[jaideep11061982]

@AndreaPi num_classes=183 FusedSemanticHead num_classes=80- Shared2FCBBoxHead/HTCMaskHead

i cant relate these in standard coco.

[AndreaPi]

@AndreaPi each image have one kind of object and background, so num_classes in FusedSemanticHead is 3+1=4. Is this correct?

I don’t understand: “each image has one kind of object and background”, so how do you get to 3? Wouldn’t that be 1+1?Anyway, I don’t use COCOstuff data, so I can’t help you out here. If I were in you, I’d have a look at these two config files:

https://github.com/open-mmlab/mmdetection/blob/master/configs/htc/htc_without_semantic_r50_fpn_1x_coco.py https://github.com/open-mmlab/mmdetection/blob/master/configs/htc/htc_r50_fpn_1x_coco.py

note down the number of classes resp. for the Shared2FCBBoxHead/HTCMaskHead and for the FusedSemanticHead, and compare these numbers with the number of classes in the COCO & COCOstuff datasets, in order to understand how to set these numbers in your config file.