slug resolver is registered to always return None, which will change interpolations like ${now:%y%m%d-%H%M%S}_${slug:2} and can alter ${paths.run_dir} / trainer.default_root_dir when loading a saved training config. This can lead to predictions writing/logging under a different run directory than the checkpoint’s parent. Align this resolver with train.py (use coolname.impl.generate_slug) so the composed config resolves identically to training runs.
Originally posted by @Copilot in #43 (comment)
slugresolver is registered to always returnNone, which will change interpolations like${now:%y%m%d-%H%M%S}_${slug:2}and can alter${paths.run_dir}/trainer.default_root_dirwhen loading a saved training config. This can lead to predictions writing/logging under a different run directory than the checkpoint’s parent. Align this resolver withtrain.py(usecoolname.impl.generate_slug) so the composed config resolves identically to training runs.Originally posted by @Copilot in #43 (comment)