Description
Context:
i am trying to run ECLIPSE on cityscapes dataset on 15(base classes)-2 (incremental classes) configuration.
For step0 training, I used BASE_CLS = 15 and INC_CLS = 2.
Training is completed successfully and PQ is around 51.
But when I run the eval_only code, I have to select BASE_CLS = 14 because IncrementalClassifier's layer 1 MLP's out_feature = 15 and it consider 1 bkg and 14 Base classes.
If i select BASE_CLS = 15 during inference, the model doesn't load properly and give following warning:
WARNING [07/29 15:58:12 fvcore.common.checkpoint]: Skip loading parameter 'sem_seg_head.predictor.class_embed.cls.1.layers.2.bias' to the model due to incompatible shapes: (15,) in the checkpoint but (16,) in the model! You might want to double check if this is expected.
Selecting BASE_CLS = 14, the accuracy of each class is okay but the PQ for "road" (trainId=0) is 0.0.
Should I modify the cityscpaes dataset and start the trainId from 1. Or there any easy way to solve this?