Source code for tbp.monty.frameworks.models.evidence_matching.hypotheses_updater

# Copyright 2025-2026 Thousand Brains Project
#
# Copyright may exist in Contributors' modifications
# and/or contributions to the work.
#
# Use of this source code is governed by the MIT
# license that can be found in the LICENSE file or at
# https://opensource.org/licenses/MIT.

from __future__ import annotations

import logging
from typing import Any, ContextManager, Dict, Literal, Optional, Protocol

import numpy as np
import numpy.typing as npt
from typing_extensions import Self

from tbp.monty.frameworks.models.evidence_matching.channels import (
    all_usable_input_channels,
)
from tbp.monty.frameworks.models.evidence_matching.feature_evidence.calculator import (
    DefaultFeatureEvidenceCalculator,
    FeatureEvidenceCalculator,
)
from tbp.monty.frameworks.models.evidence_matching.feature_evidence.scorer import (
    DefaultFeatureEvidenceScorer,
)
from tbp.monty.frameworks.models.evidence_matching.features_for_matching.selector import (  # noqa: E501
    DefaultFeaturesForMatchingSelector,
    FeaturesForMatchingSelector,
)
from tbp.monty.frameworks.models.evidence_matching.graph_memory import (
    EvidenceGraphMemory,
)
from tbp.monty.frameworks.models.evidence_matching.hypotheses import Hypotheses
from tbp.monty.frameworks.models.evidence_matching.hypotheses_displacer import (
    DefaultHypothesesDisplacer,
)
from tbp.monty.frameworks.utils.graph_matching_utils import (
    get_initial_possible_poses,
    possible_sensed_directions,
)
from tbp.monty.frameworks.utils.spatial_arithmetics import (
    align_multiple_orthonormal_vectors,
)
from tbp.monty.geometry import Rotation

logger = logging.getLogger(__name__)

HypothesesUpdateTelemetry = Optional[Dict[str, Any]]
HypothesesUpdaterTelemetry = Dict[str, Any]


[docs]class HypothesesUpdater(ContextManager[Self], Protocol):
[docs] def reset(self) -> None: """Resets updater at the beginning of an episode."""
[docs] def update_hypotheses( self, hypotheses: Hypotheses, features: dict, displacement: npt.NDArray[np.float64] | None, graph_id: str, evidence_update_threshold: float, ) -> tuple[Hypotheses | None, HypothesesUpdateTelemetry]: """Update hypotheses based on sensor displacement and sensed features. Args: hypotheses: Hypothesis space for the graph. features: Input features keyed by channel name. displacement: LM displacement between the current and previous input. graph_id: ID of the graph being updated. evidence_update_threshold: Evidence update threshold. Returns: Updated graph hypothesis space (or None if no channels available) and telemetry. """ ...
[docs]class DefaultHypothesesUpdater(HypothesesUpdater):
[docs] def __init__( self, feature_weights: dict, graph_memory: EvidenceGraphMemory, max_match_distance: float, tolerances: dict, evidence_threshold_config: float | str = "all", # noqa: ARG002 feature_evidence_calculator: type[FeatureEvidenceCalculator] = ( DefaultFeatureEvidenceCalculator ), feature_evidence_increment: int = 1, features_for_matching_selector: type[FeaturesForMatchingSelector] = ( DefaultFeaturesForMatchingSelector ), initial_possible_poses: Literal["uniform", "informed"] | list[Rotation] = "informed", max_nneighbors: int = 3, past_weight: float = 1, present_weight: float = 1, umbilical_num_poses: int = 8, ): """Initializes the DefaultHypothesesUpdater. Args: feature_weights: How much each feature should be weighted when calculating the evidence update for a hypothesis. Weights are stored in a dictionary with keys corresponding to features (same as keys in tolerances). graph_memory: The graph memory to read graphs from. max_match_distance: Maximum distance of a tested and stored location for them to be matched. tolerances: How much can each observed feature deviate from the stored features while still being considered a match. evidence_threshold_config: How to decide which hypotheses should be updated. When this parameter is either '[int]%' or 'x_percent_threshold', then this parameter is applied to the evidence for the Most Likely Hypothesis (MLH) to determine a minimum evidence threshold in order for other hypotheses to be updated. Any hypotheses falling below the resulting evidence threshold do not get updated. The other options set a fixed threshold that does not take MLH evidence into account. In [int, float, '[int]%', 'mean', 'median', 'all', 'x_percent_threshold']. Defaults to 'all'. Ignored by the DefaultHypothesesUpdater. Required for compatibility with the HypothesesUpdater protocol. feature_evidence_calculator: Class to calculate feature evidence for all nodes. Defaults to the default calculator. feature_evidence_increment: Feature evidence (between 0 and 1) is multiplied by this value before being added to the overall evidence of a hypothesis. This factor is only multiplied with the feature evidence (not the pose evidence, unlike the present_weight). Defaults to 1. features_for_matching_selector: Class to select if features should be used for matching. Defaults to the default selector. initial_possible_poses: Initial possible poses to test. Defaults to "informed". max_nneighbors: Maximum number of nearest neighbors to consider in the radius of a hypothesis for calculating the evidence. Defaults to 3. past_weight: How much should the evidence accumulated so far be weighted when combined with the evidence from the most recent observation. Defaults to 1. present_weight: How much should the current evidence be weighted when added to the previous evidence. If past_weight and present_weight add up to 1, the evidence is bounded and can't grow infinitely. Defaults to 1. NOTE: Right now this doesn't give as good performance as with unbounded evidence since we don't keep a full history of what we saw. With a more efficient policy and better parameters, that may be possible to use and could help when moving from one object to another and generally make setting thresholds more intuitive. umbilical_num_poses: Number of sampled rotations in the direction of the plane perpendicular to the surface normal. These are sampled at umbilical points (i.e., points where PC directions are undefined). """ self.graph_memory = graph_memory self.initial_possible_poses = get_initial_possible_poses(initial_possible_poses) self.umbilical_num_poses = umbilical_num_poses # TODO: Dependency inject a constructed FeatureEvidenceScorer instead self._feature_evidence_scorer = DefaultFeatureEvidenceScorer( graph_memory=self.graph_memory, feature_weights=feature_weights, tolerances=tolerances, feature_evidence_calculator=feature_evidence_calculator, feature_evidence_increment=feature_evidence_increment, features_for_matching_selector=features_for_matching_selector, ) # TODO: Dependency inject a constructed HypothesesDisplacer instead self._hypotheses_displacer = DefaultHypothesesDisplacer( feature_weights=feature_weights, graph_memory=self.graph_memory, max_match_distance=max_match_distance, max_nneighbors=max_nneighbors, past_weight=past_weight, present_weight=present_weight, feature_evidence_scorer=self._feature_evidence_scorer, ) self._initialized_channels: dict[str, set[str]] = {}
def __enter__(self) -> Self: """Enter context manager, runs before updating the hypotheses. Returns: Self: The context manager instance. """ return self def __exit__(self, exc_type, exc_val, exc_tb) -> None: """Exit context manager, runs after updating the hypotheses."""
[docs] def reset(self) -> None: """Resets updater at the beginning of an episode.""" self._initialized_channels = {}
[docs] def update_hypotheses( self, hypotheses: Hypotheses, features: dict, displacement: npt.NDArray[np.float64] | None, graph_id: str, evidence_update_threshold: float, ) -> tuple[Hypotheses | None, HypothesesUpdateTelemetry]: """Update hypotheses based on sensor displacement and sensed features. Updates the existing hypothesis space or initializes a new hypothesis space if one does not exist (i.e., at the beginning of the episode). Updating the hypothesis space includes displacing the hypotheses' possible locations, as well as updating their evidence scores. Evidence from all available input channels is computed and summed by the HypothesesDisplacer. Args: hypotheses: Hypothesis space for the graph. features: Input features keyed by channel name. displacement: LM displacement between the current and previous input. graph_id: Identifier of the graph being updated. evidence_update_threshold: Evidence update threshold. Returns: Updated hypothesis space (or None if no channels available) and telemetry. """ # Get all usable input channels # NOTE: We might also want to check the confidence in the input channel # features, but this information is currently not available here. # TODO S: Once we pull the observation class into the LM we could add this. input_channels_to_use = all_usable_input_channels( features, self.graph_memory.get_input_channels_in_graph(graph_id) ) if graph_id not in self._initialized_channels: self._initialized_channels[graph_id] = set() if len(input_channels_to_use) == 0: logger.info( f"No input channels observed for {graph_id} that are stored in the " "model. Not updating evidence." ) return None, {} if hypotheses.count == 0: all_hyps = [] for channel in input_channels_to_use: channel_hyps = self._get_initial_hypothesis_space( channel_features=features[channel], graph_id=graph_id, input_channel=channel, ) all_hyps.append(channel_hyps) self._initialized_channels[graph_id].add(channel) return Hypotheses.concatenate(all_hyps), {} # We only displace existing hypotheses since the newly sampled # hypotheses should not be affected by the displacement from the last # sensory input. displaced_hypotheses, displacer_telemetry = ( self._hypotheses_displacer.displace_hypotheses_and_compute_evidence( displacement=displacement, features=features, evidence_update_threshold=evidence_update_threshold, graph_id=graph_id, possible_hypotheses=hypotheses, ) ) # Initialize hypotheses for newly available channels. new_channels = [ ch for ch in input_channels_to_use if ch not in self._initialized_channels[graph_id] ] if new_channels: displaced_hypotheses = self._initialize_new_channels( displaced_hypotheses, features, new_channels, graph_id ) telemetry = {"mlh_prediction_error": displacer_telemetry.mlh_prediction_error} return displaced_hypotheses, telemetry
def _initialize_new_channels( self, hypotheses: Hypotheses, features: dict, new_channels: list[str], graph_id: str, ) -> Hypotheses: """Initialize hypotheses for channels that haven't been seen before. When a new channel starts reporting after other channels have already accumulated evidence, the new channel's hypotheses are initialized with the current mean evidence added to give them a fighting chance. Args: hypotheses: Current unified hypotheses. features: Input features keyed by channel name. new_channels: Channels to initialize. graph_id: Identifier of the graph being updated. Returns: Hypotheses with new channel hypotheses appended. """ current_mean_evidence = np.mean(hypotheses.evidence) new_hyps = [] for channel in new_channels: channel_hyps = self._get_initial_hypothesis_space( channel_features=features[channel], graph_id=graph_id, input_channel=channel, ) # Add current mean evidence to give the new hypotheses a fighting # chance against existing hypotheses that have been accumulating # evidence from other channels. new_hyps.append( Hypotheses( evidence=channel_hyps.evidence + current_mean_evidence, locations=channel_hyps.locations, poses=channel_hyps.poses, possible=channel_hyps.possible, ) ) self._initialized_channels[graph_id].add(channel) return Hypotheses.concatenate([hypotheses, *new_hyps]) def _get_all_informed_possible_poses( self, graph_id: str, sensed_channel_features: dict, input_channel: str ): """Initialize hypotheses on possible rotations for each location. This is similar to _get_informed_possible_poses but doesn't require looping over nodes. For this, we use the surface normal and curvature directions and check how they would have to be rotated to match between sensed and stored vectors at each node. If principal curvature is similar in both directions, the direction vectors cannot inform this and we have to uniformly sample multiple possible rotations along this plane. Note: In general this initialization of hypotheses determines how well the matching later on does and if an object and pose can be recognized. We should think about whether this is the most robust way to initialize hypotheses. Returns: The possible locations and rotations. """ all_possible_locations = np.zeros((1, 3)) all_possible_rotations = np.zeros((1, 3, 3)) logger.debug(f"Determining possible poses using input from {input_channel}") node_directions = self.graph_memory.get_rotation_features_at_all_nodes( graph_id, input_channel ) sensed_directions = sensed_channel_features["pose_vectors"] # Check if PCs in patch are similar -> need to sample more directions if ( "pose_fully_defined" in sensed_channel_features and not sensed_channel_features["pose_fully_defined"] ): possible_s_d = possible_sensed_directions( sensed_directions, self.umbilical_num_poses ) else: possible_s_d = possible_sensed_directions(sensed_directions, 2) for s_d in possible_s_d: # Since we have orthonormal vectors and know their correspondence we can # directly calculate the rotation instead of using the Kabsch estimate # used in Rotation.align_vectors r = align_multiple_orthonormal_vectors(node_directions, s_d, as_scipy=False) all_possible_locations = np.vstack( [ all_possible_locations, np.array( self.graph_memory.get_locations_in_graph( graph_id, input_channel ) ), ] ) all_possible_rotations = np.vstack([all_possible_rotations, r]) return all_possible_locations[1:], all_possible_rotations[1:] def _get_initial_hypothesis_space( self, channel_features: dict, graph_id: str, input_channel: str ) -> Hypotheses: if self.initial_possible_poses is None: # Get initial poses for all locations informed by pose features ( initial_possible_channel_locations, initial_possible_channel_rotations, ) = self._get_all_informed_possible_poses( graph_id, channel_features, input_channel ) else: initial_possible_channel_locations = [] initial_possible_channel_rotations = [] all_channel_locations = self.graph_memory.get_locations_in_graph( graph_id, input_channel ) # Initialize fixed possible poses (without using pose features) for rotation in self.initial_possible_poses: for node_id in range(len(all_channel_locations)): initial_possible_channel_locations.append( all_channel_locations[node_id] ) initial_possible_channel_rotations.append(rotation.as_matrix()) initial_possible_channel_locations = np.array( initial_possible_channel_locations ) initial_possible_channel_rotations = np.array( initial_possible_channel_rotations ) # There will always be two feature weights (surface normal and curvature # direction). If there are no more weights we are not using features for # matching and skip this step. Doing matching with only morphology can # currently be achieved in two ways. Either we don't specify tolerances # and feature_weights or we set the global feature_evidence_increment to 0. node_feature_evidence = self._feature_evidence_scorer( graph_id=graph_id, input_channel=input_channel, query_features=channel_features, ) # Stack node_feature_evidence to match possible poses nwmf_stacked = [] for _ in range( len(initial_possible_channel_rotations) // len(node_feature_evidence) ): nwmf_stacked.extend(node_feature_evidence) evidence = np.array(nwmf_stacked) # New hypotheses cannot be possible initial_possible_hyps = np.zeros_like(evidence, dtype=np.bool_) return Hypotheses( evidence=evidence, locations=initial_possible_channel_locations, poses=initial_possible_channel_rotations, possible=initial_possible_hyps, )