Enum: ObservationReductionMode
How an observation is evaluated over the trajectory (see Observation.reduce). Absent (the default) keeps the post-scan pipeline: the observation is computed from a materialised trajectory. streaming opts the observation into an incremental reducer that is folded into the integrator carry via prepare(reduce=…), so the trajectory is never held — byte-identical to the post-scan value (to f64 rounding), but with O(block) instead of O(n_time) peak memory. Required for whole-brain fits whose per-stage simulation is long enough that the full trajectory would not fit in memory (e.g. the Schirner 2023 BOLD/FC group fit).
URI: tvbo:enum/ObservationReductionMode
Permissible Values
| Value | Meaning | Description |
|---|---|---|
| streaming | None | Fold this observation into the integrator carry as an (init, update, finalize… |
| trials | None | Evaluate the pipeline across the trial ensemble rather than per solve: the (s… |
Slots
| Name | Description |
|---|---|
| reduce | Opt into streaming evaluation of this observation: fold it into the integrato… |
Identifier and Mapping Information
Schema Source
- from schema: https://w3id.org/tvbo
LinkML Source
name: ObservationReductionMode
description: 'How an observation is evaluated over the trajectory (see Observation.reduce).
Absent (the default) keeps the post-scan pipeline: the observation is computed from
a materialised trajectory. `streaming` opts the observation into an incremental
reducer that is folded into the integrator carry via prepare(reduce=...), so the
trajectory is never held — byte-identical to the post-scan value (to f64 rounding),
but with O(block) instead of O(n_time) peak memory. Required for whole-brain fits
whose per-stage simulation is long enough that the full trajectory would not fit
in memory (e.g. the Schirner 2023 BOLD/FC group fit).'
from_schema: https://w3id.org/tvbo
rank: 1000
permissible_values:
streaming:
text: streaming
description: Fold this observation into the integrator carry as an (init, update,
finalize) reducer; never materialise the trajectory. Supported for the HRF-Volterra
BOLD pipeline (hemodynamic convolution, whose kernel / downsample stride / TR
stride / Volterra scaling the resolver lifts from the declared pipeline), cumulative
mean/std/variance aggregations, and a matrix co-moment FC (compute_fc as a running
covariance).
trials:
text: trials
description: 'Evaluate the pipeline across the trial ensemble rather than per
solve: the (single) source observation is computed per trial as usual, and after
the trial map this observation''s (single) pipeline stage runs host-side on
the trial-stacked source (n_trials, ...) — e.g. a pattern-entropy over an ensemble
of per-trial evoked patterns. Requires a trial-only exploration (n_trials without
parameter axes).'