Thanks for releasing the SegAnyPET code.
I found that compute_epistemic_uncertainty(stacked_preds) is called in train_cpcl.py, but this function seems to be missing from the repository.
According to the paper, the CPCL strategy estimates uncertainty from multiple prompt-based predictions using predictive entropy, and then uses it for label rectification. However, the current code only contains compute_uncertainty(predictions), which appears to compute variance instead.
Could you please provide the missing implementation of compute_epistemic_uncertainty, or release the complete CPCL training code used in the paper?
Thanks.
Thanks for releasing the SegAnyPET code.
I found that
compute_epistemic_uncertainty(stacked_preds)is called intrain_cpcl.py, but this function seems to be missing from the repository.According to the paper, the CPCL strategy estimates uncertainty from multiple prompt-based predictions using predictive entropy, and then uses it for label rectification. However, the current code only contains
compute_uncertainty(predictions), which appears to compute variance instead.Could you please provide the missing implementation of
compute_epistemic_uncertainty, or release the complete CPCL training code used in the paper?Thanks.