cellmap_flow.finetune.job_manager

Finetune jobs, from the dashboard’s side: submitting them, following them until they end, restarting them, and finding them again after the dashboard itself restarts.

  • manager: FinetuneJobManager, which holds the jobs and is what the finetune routes call.

  • state: a job’s record (FinetuneJob), its statuses, and what moves a job from one to another.

  • submit: what a submission runs and where: the model’s type and settings, the trainer’s command line, and launching it on LSF or here.

  • persistence: each run’s metadata.json and export, and finding a session’s jobs again from them (rehydration).

  • monitor: following a job until it ends, from its scheduler’s answers and its log, and telling the listeners.

  • tailer: the trainer’s log as the manager reads it.

  • listener: what the manager tells its listeners (the dashboard’s viewer) of each job, and how.

  • restart: asking a job that waits for a restart to train again.

The trainer’s side of the same protocol (its markers, its restart signal, the layout of its run directory) is finetune.markers, finetune.session_loop and finetune.run_outputs.

Nothing here imports torch, Flask, neuroglancer or cellmap_flow.globals at module level. Import the submodule you need; this package imports none of them itself.

Submodules