Posting here (per Florian's suggestion, with Ashesh's agreement) so the clarifications from a recent email thread are publicly available to other users reproducing the Table 1 tasks. Answers below are paraphrased from the authors' replies.
Q1 — Which repo should be used for configs vs. data splits?
Use juglab/MicroSplit only to obtain the train/val/test splits. Use CAREamics/microSplit-reproducibility for training and configurations. (The juglab/MicroSplit repo is a trimmed clone of ashesh-0/Disentangle and still contains a lot of task-irrelevant material, so it is not the right source for configs.)
Q2 — How should a config be adapted for a given task?
Always start from a provided microSplit-reproducibility config that has the same 2D/3D setting as the target task, then change only what the manuscript/supplement specifies. The changes are generally:
(a) whether Lateral Context (LC) is used,
(b) whether a noise model is used,
(c) on the data side, whether the input is real or synthetic.
Q3 — What exactly is the Chicago-Sch23 task in Table 1?
It is a 4-channel splitting task: one model outputs all four structures as targets. The supplement states the task "is generated from the Chicago-Sch23 dataset and uses all four structures." Since the dataset provides four channels and the model predicts four, there is no dedicated input channel — the input is synthetic:
channel_idx_list=[0,1,2,3], input_idx=None, target_idx_list=[0,1,2,3]. No noise model is used for this task.
Happy to move this into the docs / reproducibility notebooks instead of an issue if the maintainers prefer.
Thanks to the authors for the detailed answers.
Posting here (per Florian's suggestion, with Ashesh's agreement) so the clarifications from a recent email thread are publicly available to other users reproducing the Table 1 tasks. Answers below are paraphrased from the authors' replies.
Q1 — Which repo should be used for configs vs. data splits?
Use
juglab/MicroSplitonly to obtain the train/val/test splits. UseCAREamics/microSplit-reproducibilityfor training and configurations. (Thejuglab/MicroSplitrepo is a trimmed clone ofashesh-0/Disentangleand still contains a lot of task-irrelevant material, so it is not the right source for configs.)Q2 — How should a config be adapted for a given task?
Always start from a provided
microSplit-reproducibilityconfig that has the same 2D/3D setting as the target task, then change only what the manuscript/supplement specifies. The changes are generally:(a) whether Lateral Context (LC) is used,
(b) whether a noise model is used,
(c) on the data side, whether the input is real or synthetic.
Q3 — What exactly is the Chicago-Sch23 task in Table 1?
It is a 4-channel splitting task: one model outputs all four structures as targets. The supplement states the task "is generated from the Chicago-Sch23 dataset and uses all four structures." Since the dataset provides four channels and the model predicts four, there is no dedicated input channel — the input is synthetic:
channel_idx_list=[0,1,2,3],input_idx=None,target_idx_list=[0,1,2,3]. No noise model is used for this task.Happy to move this into the docs / reproducibility notebooks instead of an issue if the maintainers prefer.
Thanks to the authors for the detailed answers.