Hello,
I’d like to ask about the minimum segment length. Based on the manuscript this is 3. The default label_filter value for min_seg is 3, but when I generate trajectories using datasets_challenge I will often receive label changes 1 or 2 frames apart. Is it safe to remove these trajectories from our training sets? Even in the provided starting kit data there are a few trajectories (26 of them) containing these short segments.
Also note that in the challenge_two_submission.ipynb and in the documentation for the datasets_challenge it says the segments will be made of at least 5 frames (maybe this was confused with the filtering window size).
Thank you,
Alvaro
Dear Alvaro,
thanks for the feedback. We have updated the notebook to match the paper's information. We have identified the bug producing these shorter segments. We will let you and the rest of the participants know when the issue is resolved.
Gorka
Posted by: gorka.munoz @ May 31, 2024, 5:20 p.m.Hey, wanted to follow up on this. I am using the github package (specifically challenge_phenom_dataset) and obtained tracks with 2 segments of a specific state? I installed a new conda environment recently (today) and pip install andi-datasets. As I understand the min segment length should be 3 as per manuscript? As an example,
360 15 360.0 88.239008 60.829480 0.939862 0.041505 1.0
361 15 361.0 87.937517 61.368659 1.162765 0.941943 2.0 <-------------
362 15 362.0 87.680051 61.829142 1.162765 0.941943 2.0 <------------
363 15 363.0 87.227396 59.605966 0.939862 0.041505 1.0
Yes I've also found some in the updated starting kit set
Posted by: alanza @ June 11, 2024, 10:31 p.m.Hi both! Thanks for the feedback.
1) The solution to the bug of the labels is still not in pip, only on github, sorry that we did not specify that. We plan to update pip this week together with other upgrades of the library. The last version of andi_datasets can always be found in the master branch of our github. For the challenge, we recommend that you clone the library from there and the run "pip install -e . " to locally install the library. The bug is solved there, let me know if you find any problem there.
2) There are still some faulty segments (13) of length 1 in the starting kit that escaped us. I am working on cleaning them, we will keep you posted.
Gorka
Posted by: gorka.munoz @ June 12, 2024, 7:53 a.m.