Summary
Follow-up to review feedback on src/sampleworks/data/cc89_epoch415.yaml, introduced in PR #300 to provide a packaged default Protpardelle training config.
The file currently bundles developer-local absolute paths (e.g. /scratch/users/...) for keys such as pdb_paths, full_mpnn_model_path, ckpt_path, and shapes_path. Since this config ships as part of the installed package and is used as a fallback config path in get_model_and_device (src/sampleworks/utils/guidance_script_utils.py), these values are not portable and will not resolve on other machines.
Rationale
- Local absolute paths baked into a packaged default config break portability for any user/environment other than the original author's machine.
- Sections such as
data (dataset/training-specific settings: sigma calculation, pdb_paths, mixing ratios, etc.) appear to only be relevant during model training and may not be needed for inference/sampling usage of the package.
Proposed changes
- Replace
pdb_paths, full_mpnn_model_path, ckpt_path, and shapes_path with safe empty defaults or package-relative resource references wherever possible.
- Audit the config for sections/keys that are only used for training (starting with the
data section) and remove them if not required for the packaged inference/sampling use case.
- Confirm which keys are actually read by the Protpardelle wrapper/sampling path (src/sampleworks/models/protpardelle/wrapper.py, src/sampleworks/utils/guidance_script_utils.py) before removing anything, to avoid breaking sampling.
Affected areas
src/sampleworks/data/cc89_epoch415.yamlsrc/sampleworks/utils/guidance_script_utils.py (consumer of the default config path)src/sampleworks/models/protpardelle/wrapper.py (consumer of config values)
Acceptance criteria
- No developer-local absolute paths remain in the packaged default config.
- Unused/training-only sections are removed or clearly documented as unused, without breaking existing sampling functionality/tests.
- Existing Protpardelle wrapper tests continue to pass.
References
Summary
Follow-up to review feedback on
src/sampleworks/data/cc89_epoch415.yaml, introduced in PR #300 to provide a packaged default Protpardelle training config.The file currently bundles developer-local absolute paths (e.g.
/scratch/users/...) for keys such aspdb_paths,full_mpnn_model_path,ckpt_path, andshapes_path. Since this config ships as part of the installed package and is used as a fallback config path inget_model_and_device(src/sampleworks/utils/guidance_script_utils.py), these values are not portable and will not resolve on other machines.Rationale
data(dataset/training-specific settings: sigma calculation, pdb_paths, mixing ratios, etc.) appear to only be relevant during model training and may not be needed for inference/sampling usage of the package.Proposed changes
pdb_paths,full_mpnn_model_path,ckpt_path, andshapes_pathwith safe empty defaults or package-relative resource references wherever possible.datasection) and remove them if not required for the packaged inference/sampling use case.Affected areas
src/sampleworks/data/cc89_epoch415.yamlsrc/sampleworks/utils/guidance_script_utils.py(consumer of the default config path)src/sampleworks/models/protpardelle/wrapper.py(consumer of config values)Acceptance criteria
References