Closed sdaulton closed 1 year ago
Hi, thanks for the suggested edit, but I do not think this is a useful addition to the repo. The decision to avoid calling mkdir
to create parent directories when downloading the files was made consciously and intentionally in order to avoid some practical issues that can crop up when working on compute clusters with various restrictions on file system usage. Essentially, we explicitly count on users of the repository to ensure that, before they make a call to JAHS-Bench-201 functions, all top-level directory structures needed (namely, save_dir
in this case) before doing so are in place. In this case, extracting the data should automatically generate the required directories as long as save_dir
already exists. This is not so significant for the surrogate model but extremely important for the tabular data or the live training APIs. Essentially, given any top-level directory that JAHS-Bench-201 has sufficient rights for, it will only ever create sub-directories as and when needed but never go higher up the directory tree.
I'm closing this pull request for the moment, but will leave the comments open.
Actually, this prompted me to check and figure out that the directory structures for assembled_models.tar was incorrect in the latest release. I have now fixed it.
see title.
To test, I successfully ran
python -m jahs_bench_examples.minimal