codes.train package

codes.train package#

Submodules#

codes.train.progress module#

Progress messages from training workers and their parent-owned display.

class codes.train.progress.EventProgress(iterable, desc='')#

Bases: object

Small tqdm-compatible iterable that reports completed epochs to the parent.

close()#
set_postfix(ordered_dict=None, refresh=True, **kwargs)#
update(n=1)#
class codes.train.progress.ProgressDisplay(tasks, slots, mode='auto', stream=None, log_dir=None)#

Bases: object

Render worker events centrally, with stable bars or plain log lines.

close()#
property elapsed#
event(slot, payload)#
finish(slot, success, detail='')#
log(slot, text)#
start(slot, task, device)#
codes.train.progress.current_reporter()#

Return the current worker’s message callback, if reporting is enabled.

codes.train.progress.worker_reporting(send)#

Route a worker’s progress and Python output to its parent process.

codes.train.tasks module#

Run individual training tasks independently of their scheduler.

codes.train.tasks.execute_training_task(task, *, device='cpu', position=0, cpu_threads=1)#

Configure a worker process and train and save exactly one model.

task contains (surrogate, mode, metric, training_id, seed, epochs), as produced by create_task_list_for_surrogate. The saved configuration is read from trained/<training_id>/config.yaml. device is a logical device string within the worker’s allocation; cpu_threads sets PyTorch’s intra-operation thread count.

Local spawned processes and future external workers can both call this function. It propagates errors and does not update pending-task files or completion markers; those belong to the scheduler invoking it.

codes.train.tasks.train_and_save_model(surr_name, mode, metric, training_id, seed=None, epochs=None, device='cpu', position=1, threadlock=None)#

Train and save a model for a specific benchmark mode.

Parameters:

  • surr_name (str) – The name of the surrogate model.

  • mode (str) – The benchmark mode.

  • metric (int) – The metric for the benchmark mode.

  • training_id (str) – The training ID for the current training session.

  • seed (int | None) – Random seed for training.

  • epochs (int | None) – Number of training epochs.

  • device (str) – Device to run training on.

  • position (int) – Model position in the task list.

  • threadlock (optional) – Legacy lock used around seeded initialisation.

codes.train.tasks.validate_execution_settings(device, cpu_threads)#

Validate worker settings without selecting a CUDA context.

CUDA indices are logical indices within the process’s visible devices. CPU threads must fit within both its affinity mask and any Slurm allocation.

codes.train.train_fcts module#

exception codes.train.train_fcts.TrainingRunError#

Bases: RuntimeError

One or more local training tasks did not complete successfully.

codes.train.train_fcts.create_task_list_for_surrogate(config, surr_name)#

Creates a list of training tasks for a specific surrogate model based on the configuration file.

Parameters:

  • config (dict) – The configuration dictionary taken from the config file.

  • surr_name (str) – The name of the surrogate model.

Returns:

A list of training tasks for the surrogate model.

Return type:

list

codes.train.train_fcts.parallel_training(tasks, device_list, task_list_filepath, *, cpu_threads=1, progress='auto')#

Run each task in a fresh process, with one process per device slot.

codes.train.train_fcts.sequential_training(tasks, device_list, task_list_filepath, *, cpu_threads=1, progress='auto')#

Run each task in a fresh process using the first device slot.

Module contents#

exception codes.train.TrainingRunError#

Bases: RuntimeError

One or more local training tasks did not complete successfully.

codes.train.create_task_list_for_surrogate(config, surr_name)#

Creates a list of training tasks for a specific surrogate model based on the configuration file.

Parameters:

  • config (dict) – The configuration dictionary taken from the config file.

  • surr_name (str) – The name of the surrogate model.

Returns:

A list of training tasks for the surrogate model.

Return type:

list

codes.train.execute_training_task(task, *, device='cpu', position=0, cpu_threads=1)#

Configure a worker process and train and save exactly one model.

task contains (surrogate, mode, metric, training_id, seed, epochs), as produced by create_task_list_for_surrogate. The saved configuration is read from trained/<training_id>/config.yaml. device is a logical device string within the worker’s allocation; cpu_threads sets PyTorch’s intra-operation thread count.

Local spawned processes and future external workers can both call this function. It propagates errors and does not update pending-task files or completion markers; those belong to the scheduler invoking it.

codes.train.parallel_training(tasks, device_list, task_list_filepath, *, cpu_threads=1, progress='auto')#

Run each task in a fresh process, with one process per device slot.

codes.train.sequential_training(tasks, device_list, task_list_filepath, *, cpu_threads=1, progress='auto')#

Run each task in a fresh process using the first device slot.

codes.train.train_and_save_model(surr_name, mode, metric, training_id, seed=None, epochs=None, device='cpu', position=1, threadlock=None)#

Train and save a model for a specific benchmark mode.

Parameters:

  • surr_name (str) – The name of the surrogate model.

  • mode (str) – The benchmark mode.

  • metric (int) – The metric for the benchmark mode.

  • training_id (str) – The training ID for the current training session.

  • seed (int | None) – Random seed for training.

  • epochs (int | None) – Number of training epochs.

  • device (str) – Device to run training on.

  • position (int) – Model position in the task list.

  • threadlock (optional) – Legacy lock used around seeded initialisation.