sequince.checkpoint

sequince.checkpoint#

Attributes#

CHECKPOINT_VERSION

Classes#

`CheckpointWriter`	Base class for persisting committed SMC checkpoints.
`StepState`	Checkpoint state associated with one configured SMC step.
`SMCCheckpoint`	Detached end-of-iteration state for resuming an SMC run.
`HDF5CheckpointWriter`	Persist committed SMC checkpoints to HDF5 at a fixed cadence.
`PickleCheckpointWriter`	Atomically persist committed SMC checkpoints as pickle files.

Functions#

`capture_step_states`(steps)	Capture mutable state from a configured step pipeline.
`restore_steps`(steps, states)	Restore step state after validating pipeline identity and order.

Module Contents#

sequince.checkpoint.CHECKPOINT_VERSION = 1#

class sequince.checkpoint.CheckpointWriter(*, every=1, save_rng=True)[source]#

Bases: abc.ABC

Base class for persisting committed SMC checkpoints.

Parameters:

every (int)
save_rng (bool)

requires_detached_checkpoint = True#: Whether the writer may retain the checkpoint after write returns.

every = 1#

save_rng = True#

is_due(next_iteration)[source]#

Return whether a checkpoint is due after this iteration.

Parameters:: next_iteration (int)
Return type:: bool

write(checkpoint)[source]#

Apply this writer’s options and persist the checkpoint.

Parameters:: checkpoint (SMCCheckpoint)
Return type:: None

class sequince.checkpoint.StepState[source]#

Checkpoint state associated with one configured SMC step.

index: int#

module: str#

qualname: str#

state: dict[str, Any]#

classmethod from_step(index, step)[source]#

Parameters:

index (int)
step (Any)

Return type:

StepState

to_dict()[source]#

Return type:: dict[str, Any]

classmethod from_dict(payload)[source]#

Parameters:: payload (collections.abc.Mapping[str, Any])
Return type:: StepState

sequince.checkpoint.capture_step_states(steps)[source]#

Capture mutable state from a configured step pipeline.

Parameters:: steps (collections.abc.Sequence[Any])
Return type:: tuple[StepState, …]

sequince.checkpoint.restore_steps(steps, states)[source]#

Restore step state after validating pipeline identity and order.

Parameters:

steps (collections.abc.Sequence[Any])
states (collections.abc.Sequence[StepState])

Return type:

None

class sequince.checkpoint.SMCCheckpoint[source]#

Detached end-of-iteration state for resuming an SMC run.

states: tuple[sequince.state.ParticleState, ...]#

next_iteration: int#

next_observation_index: int#

metadata: dict[str, Any]#

outputs: tuple[tuple[Any, ...], ...]#

step_states: tuple[StepState, ...]#

evaluation_counts: sequince.target.EvaluationCounts#

rng_state: dict[str, Any] | None = None#

observation: Any = None#

initial_state: sequince.state.ParticleState | None = None#

version: int = 1#

property state: sequince.state.ParticleState#

Return the current state represented by this checkpoint.

Return type:: sequince.state.ParticleState

classmethod from_runtime(state, context, result, steps, rng=None, *, detach=True)[source]#

Capture a checkpoint after a completed SMC iteration.

Parameters:

state (sequince.state.ParticleState)
context (sequince.steps.context.StepContext)
result (sequince.result.SMCResult)
steps (collections.abc.Sequence[Any])
rng (Any | None)
detach (bool)

Return type:

SMCCheckpoint

to_dict()[source]#

Return a detached plain-Python checkpoint representation.

Return type:: dict[str, Any]

classmethod from_dict(payload, *, copy_values=True)[source]#

Reconstruct a checkpoint from to_dict() output.

Parameters:

payload (collections.abc.Mapping[str, Any])
copy_values (bool)

Return type:

SMCCheckpoint

restore_rng()[source]#

Return a live ORNG generator restored from this checkpoint.

Return type:: Any | None

result(shared_context=None)[source]#

Return the result states represented by this checkpoint.

Parameters:: shared_context (collections.abc.Mapping[str, Any] | None)
Return type:: sequince.result.SMCResult

save_pickle(file_path, protocol=None)[source]#

Save this checkpoint as a versioned pickle payload.

Parameters:

file_path (str | pathlib.Path)
protocol (int | None)

Return type:

None

classmethod load_pickle(file_path)[source]#

Load a trusted checkpoint from a pickle file.

Parameters:: file_path (str | pathlib.Path)
Return type:: SMCCheckpoint

save_hdf5(h5_file, path='checkpoint', dsetname='runtime', protocol=None, *, replace_states=False)[source]#

Append states and commit their count with the runtime payload.

Interrupted writes leave the previous runtime readable. This protects against incomplete writes in a readable HDF5 file, not file corruption.

Parameters:

path (str)
dsetname (str)
protocol (int | None)
replace_states (bool)

Return type:

None

classmethod load_hdf5(h5_file, path='checkpoint', dsetname='runtime')[source]#

Load a checkpoint from its recursive HDF5 representation.

Parameters:

path (str)
dsetname (str)

Return type:

SMCCheckpoint

class sequince.checkpoint.HDF5CheckpointWriter(file_path, *, every=1, path='checkpoint', dsetname='runtime', save_rng=True)[source]#

Bases: CheckpointWriter

Persist committed SMC checkpoints to HDF5 at a fixed cadence.

Parameters:

file_path (str | pathlib.Path)
every (int)
path (str)
dsetname (str)
save_rng (bool)

requires_detached_checkpoint = False#: Whether the writer may retain the checkpoint after write returns.

file_path#

path = 'checkpoint'#

dsetname = 'runtime'#

class sequince.checkpoint.PickleCheckpointWriter(file_path, *, every=1, save_rng=True, protocol=None)[source]#

Bases: CheckpointWriter

Atomically persist committed SMC checkpoints as pickle files.

Parameters:

file_path (str | pathlib.Path)
every (int)
save_rng (bool)
protocol (int | None)

requires_detached_checkpoint = False#: Whether the writer may retain the checkpoint after write returns.

file_path#

protocol = None#