| Detailed description |
|
|
Hello,
I am a graduate student researcher working on long-timescale MD simulation, using MSMBuilder to reduce the dimensionality of my data and construct an MSM. I have a couple questions about issues I've been running into when applying MSMBuilder to large datasets:
If the directory in which to write transformed .npy files already exists, a ValueError is raised. However, this has caused a lot of troubles when trying to featurize my data, even if transforming one trajectory at a time or queueing individual trajectory transformations through a job manager. When a job has been evicted from a node and returns to execution, the new directory has already been created, the error is raised, and the job is terminated. Is there any way to circumvent this using method parameters or apply a conditional statement which would append a suffix to the directory name? Does this happen to simply be unintended behavior or is this something I would need to pull from github and edit for my own use?
Additionally, is there a version of the methods which parallelizes calculations for running on multiple cores? Or perhaps one that outputs each transformed trajectory in the dataset as it is transformed? In the MSMBuilder docs, a method called 'LandmarkHierarchical' is mentioned in the FAQ as a heuristic to determine an optimal tradeoff between the volume of data/memory and accuracy, but it is mentioned nowhere else in MSMBuilder documentation. Where can I find information on this module? Since it may crash out due to Memory Error before writing any transformed trajectories, this method would be extremely helpful for me to avoid "guess-and-check-ing" for how large the initial data can be at each step, then having to fully restart or manually check each file to maintain progress. My apologies for asking these last two questions in the bugs email, but all of the MSMBuilder forums seem to be archived.
Thank you for your time and the use of this excellent tool!
Mikaela Farrugia
mfarrugi@nd.edu
(Sorry, resubmitted due to the lack of an email field here) |
|
