Spaces:
Running
Running
File size: 8,633 Bytes
31226fd | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 | """Observation-only memory/prediction integration with independent ablations."""
from __future__ import annotations
import copy
import hashlib
import json
import os
import tempfile
from dataclasses import dataclass, field
from pathlib import Path
from typing import Any
from src.embodiment.models import ActionCommand, EnvironmentObservation, SensorFrame
from .store import MemoryStore, PredictionRecord
from .world_model import TransitionWorldModel, WorldPrediction
class MemoryWorldModelError(ValueError):
"""Raised when the memory/world-model runtime contract is misused."""
def _boolean(value: Any, name: str) -> bool:
if type(value) is not bool:
raise MemoryWorldModelError(f"{name} must be a boolean")
return value
@dataclass(slots=True)
class MemoryWorldModel:
"""Observer only: no prediction/error is silently injected into SNN learning."""
store: MemoryStore
world_model: TransitionWorldModel
run_id: str
enabled: bool = True
persistence_path: Path | None = None
_episode_id: str = "episode-0"
prediction_enabled: bool = True
learning_enabled: bool = True
_previous: dict[str, Any] | None = field(default=None, repr=False)
last_error_components: dict[str, Any] | None = field(default=None, repr=False)
def __post_init__(self) -> None:
_boolean(self.enabled, "enabled")
_boolean(self.prediction_enabled, "prediction_enabled")
_boolean(self.learning_enabled, "learning_enabled")
if self.run_id != self.store.run_id:
raise MemoryWorldModelError("store and predictor run identities differ")
def reset_episode(self, episode_id: str) -> None:
self._episode_id = episode_id
self._previous = None
self.last_error_components = None
def predict(
self, frame: SensorFrame, action: ActionCommand | None, tick: int
) -> WorldPrediction | None:
if not self.enabled or not self.prediction_enabled:
return None
previous_state = None
if (
self._previous is not None
and self._previous["sensor_id"] == frame.sensor_id
and self._previous["modality"] == frame.modality
):
previous_state = self._previous["state"]
return self.world_model.predict(
frame, action, target_tick=tick + 1, persistence_state=previous_state
)
def complete(
self,
frame: SensorFrame,
action: ActionCommand | None,
observation: EnvironmentObservation | None,
tick: int,
prediction: WorldPrediction | None,
) -> None:
if not self.enabled:
return
self.store.record(frame, action, observation, episode_id=self._episode_id)
self.last_error_components = None
if observation is None:
return
# Score the pre-action prediction before any model update (prequential).
if prediction is not None:
self.last_error_components = self.world_model.error_components(
prediction.predicted_state, observation.state
)
if self.store.write_enabled:
self.store.record_prediction(
PredictionRecord(
self.run_id,
self._episode_id,
tick,
prediction.target_tick,
prediction.source,
prediction.predicted_state,
observation.state,
self.world_model.error(
prediction.predicted_state, observation.state
),
prediction.uncertainty,
)
)
if self.learning_enabled:
self.world_model.update(frame, action, observation)
# A single prior observation is independent of episodic read/write flags.
self._previous = {
"sensor_id": frame.sensor_id,
"modality": frame.modality,
"state": copy.deepcopy(observation.state),
}
if self.persistence_path is not None:
self.save(self.persistence_path)
def state_dict(self) -> dict[str, Any]:
state = {
"schema_version": 2,
"owner": "memory.world_model.integration",
"run_id": self.run_id,
"enabled": self.enabled,
"episode_id": self._episode_id,
"prediction_enabled": self.prediction_enabled,
"learning_enabled": self.learning_enabled,
"previous_observation": copy.deepcopy(self._previous),
"last_error_components": copy.deepcopy(self.last_error_components),
"world_model": self.world_model.state_dict(),
"memory": self.store.state_dict(),
}
unsigned = json.dumps(
state,
sort_keys=True,
separators=(",", ":"),
ensure_ascii=True,
allow_nan=False,
)
state["integrity_digest"] = hashlib.sha256(unsigned.encode("utf-8")).hexdigest()
return state
def save(self, path: Path | None = None) -> Path:
destination = path or self.persistence_path
if destination is None:
raise MemoryWorldModelError("coupled persistence path is not configured")
payload = json.dumps(
self.state_dict(),
sort_keys=True,
separators=(",", ":"),
ensure_ascii=True,
allow_nan=False,
).encode("utf-8")
destination.parent.mkdir(parents=True, exist_ok=True)
fd, temporary = tempfile.mkstemp(
prefix=f".{destination.name}.", dir=str(destination.parent)
)
try:
with os.fdopen(fd, "wb") as stream:
stream.write(payload)
stream.flush()
os.fsync(stream.fileno())
os.replace(temporary, destination)
finally:
if os.path.exists(temporary):
os.unlink(temporary)
return destination
@classmethod
def load(cls, path: Path) -> "MemoryWorldModel":
try:
state = json.loads(path.read_text(encoding="utf-8"))
if not isinstance(state, dict) or state.get("schema_version") != 2:
raise MemoryWorldModelError(
"unsupported coupled state schema; retain legacy file and rebuild from a provenance-bound replay"
)
unsigned = dict(state)
digest = unsigned.pop("integrity_digest", None)
expected = hashlib.sha256(
json.dumps(
unsigned,
sort_keys=True,
separators=(",", ":"),
ensure_ascii=True,
allow_nan=False,
).encode("utf-8")
).hexdigest()
if (
state.get("owner") != "memory.world_model.integration"
or digest != expected
):
raise MemoryWorldModelError("coupled state integrity check failed")
previous = state.get("previous_observation")
if previous is not None and (
not isinstance(previous, dict)
or not isinstance(previous.get("sensor_id"), str)
or not isinstance(previous.get("modality"), str)
or not isinstance(previous.get("state"), dict)
):
raise MemoryWorldModelError("invalid previous observation")
return cls(
store=MemoryStore.from_state_dict(state["memory"]),
world_model=TransitionWorldModel.from_state_dict(state["world_model"]),
run_id=str(state["run_id"]),
enabled=_boolean(state["enabled"], "enabled"),
persistence_path=path,
_episode_id=str(state["episode_id"]),
prediction_enabled=_boolean(
state["prediction_enabled"], "prediction_enabled"
),
learning_enabled=_boolean(
state["learning_enabled"], "learning_enabled"
),
_previous=copy.deepcopy(previous),
last_error_components=copy.deepcopy(state.get("last_error_components")),
)
except MemoryWorldModelError:
raise
except (OSError, KeyError, TypeError, ValueError) as error:
raise MemoryWorldModelError(
"coupled state could not be restored"
) from error
|