Dear author, thank you for the excellent work you shared. I have encountered some problems in the reproduction process and would like to ask you for advice. When I tried to reproduce with acdc cardiac dataset, the dice value of each part in the training process began to drop from a very small value (0.01) to 0, only the dice of the background was rising, and the training graph in the log read was basically not a graph, but similar to a noise picture. Is this a normal occurrence of insufficient epoch training or is it my mistake?

Thank you so much for releasing the code. However, I had a doubt with the way the model checkpoints are saved.

ckpt_latest_saver = ModelCheckpoint(save_top_k=1, dirpath=root_dir / LATEST_CHECKPOINT_DIR, monitor="step", mode="max")

Is there a reason not to save checkpoints based on the validation loss and just save the latest checkpoint? As far as I can understand, the monitor=step implies there is no specific criterion for saving the model checkpoints, just its epoch number. Please correct me if I am wrong.

Best,
Chinmay

thanks for sharing.
have you ever tested the segmentation on echcardiacgraphy images ?