Spaces:
Sleeping
Sleeping
File size: 20,224 Bytes
82614b0 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 | """BrainRL region selection environment implementation."""
from __future__ import annotations
import json
import math
import os
import random
from typing import Any
from uuid import uuid4
try:
from openenv.core.env_server.mcp_environment import MCPEnvironment
from openenv.core.env_server.types import Action, Observation, State
except ImportError: # pragma: no cover - keeps local smoke tests lightweight
class MCPEnvironment: # type: ignore[no-redef]
def __init__(self, mcp: Any | None = None):
self.mcp = mcp
def step(self, action: Any, timeout_s: float | None = None, **kwargs: Any) -> Any:
return self._step_impl(action, timeout_s=timeout_s, **kwargs)
async def step_async(self, action: Any, timeout_s: float | None = None, **kwargs: Any) -> Any:
return self._step_impl(action, timeout_s=timeout_s, **kwargs)
class Action: # type: ignore[no-redef]
pass
class Observation: # type: ignore[no-redef]
def __init__(self, done: bool = False, reward: float = 0.0, metadata: dict | None = None):
self.done = done
self.reward = reward
self.metadata = metadata or {}
class State: # type: ignore[no-redef]
def __init__(self, episode_id: str, step_count: int = 0):
self.episode_id = episode_id
self.step_count = step_count
try:
from fastmcp import FastMCP
except ImportError: # pragma: no cover
class FastMCP: # type: ignore[no-redef]
def __init__(self, name: str):
self.name = name
def tool(self, func):
return func
from data_loader import (
BrainSubset,
DEFAULT_CONFIG_PATH,
RegionCandidate,
_parse_simple_yaml,
candidate_table,
load_brain_subset,
)
from rewards import format_reward_breakdown, reward_columns, score_region_action
from stimulus_loader import (
compact_for_prompt as compact_stimulus_for_prompt,
n_windows as stimulus_n_windows,
stimulus_bias_for_group,
summarize_window as summarize_stimulus_window,
)
def _env_int(name: str, default: int) -> int:
raw = os.environ.get(name)
if not raw:
return default
try:
return int(raw)
except (TypeError, ValueError):
return default
def _yaml_stimulus_defaults() -> tuple[int, int]:
"""Read stimulus_window_size / stimulus_top_words from subset_config.yaml."""
try:
config = _parse_simple_yaml(DEFAULT_CONFIG_PATH)
except Exception: # pragma: no cover - defensive: never fail env construction
config = {}
window = int(config.get("stimulus_window_size", 30) or 30)
top = int(config.get("stimulus_top_words", 10) or 10)
return max(1, window), max(1, top)
_yaml_window, _yaml_top = _yaml_stimulus_defaults()
_DEFAULT_STIMULUS_WINDOW_SIZE = _env_int("BRAINRL_STIMULUS_WINDOW_SIZE", _yaml_window)
_DEFAULT_STIMULUS_TOP_WORDS = _env_int("BRAINRL_STIMULUS_TOP_WORDS", _yaml_top)
TASKS = {
"roi_selection": {
"description": (
"Sequentially select brain regions that improve prediction of auditory "
"stimulus responses from a compact Le Petit Prince fMRI summary."
),
"difficulty": "medium",
"reward": "independent_verifier_components",
"valid_actions": "Any unselected candidate region_id",
}
}
class BrainRegionSelectionEnvironment(MCPEnvironment):
"""OpenEnv-style environment for active brain region acquisition."""
def __init__(self):
mcp = FastMCP("brain_rl_region_selection")
@mcp.tool
def get_selection_state() -> dict:
"""Return selected regions, current score, budget, and candidates."""
return self._build_selection_state()
@mcp.tool
def get_task_info() -> dict:
"""Return task metadata and reward definition."""
return self._build_task_info()
@mcp.tool
def take_action(region_id: str) -> dict:
"""Select the next region and advance one acquisition step."""
return self._process_action(region_id)
super().__init__(mcp)
self._state = State(episode_id=str(uuid4()), step_count=0)
self._rng = random.Random(42)
self._subset: BrainSubset = load_brain_subset()
self._task = "roi_selection"
self._episode_id = str(uuid4())
self._timestep = 0
self._selected_region_ids: list[str] = []
self._current_r2 = 0.0
self._last_feedback = "Environment initialized"
self._subject_id: str | None = None
self._run_id: str | None = None
self._condition: str | None = None
self._stimulus_window_size: int = _DEFAULT_STIMULUS_WINDOW_SIZE
self._stimulus_top_words: int = _DEFAULT_STIMULUS_TOP_WORDS
self._stimulus_window_index: int | None = None
self._stimulus_features: dict | None = None
self._effective_base_r2: dict[str, float] = self._compute_effective_base_r2()
def reset(
self,
seed: int | None = None,
episode_id: str | None = None,
task: str = "roi_selection",
subject_id: str | None = None,
run_id: str | None = None,
condition: str | None = None,
stimulus_window: int | None = None,
stimulus_window_size: int | None = None,
stimulus_top_words: int | None = None,
**_: Any,
) -> Observation:
if task not in TASKS:
raise ValueError(f"Unknown task={task}. Valid tasks: {sorted(TASKS)}")
if seed is not None:
self._rng = random.Random(seed)
self._subset = load_brain_subset()
self._task = task
self._episode_id = episode_id or str(uuid4())
self._state = State(episode_id=self._episode_id, step_count=0)
self._timestep = 0
self._selected_region_ids = []
self._current_r2 = 0.0
self._subject_id = subject_id
self._run_id = run_id
self._condition = condition
if stimulus_window_size is not None:
self._stimulus_window_size = max(1, int(stimulus_window_size))
if stimulus_top_words is not None:
self._stimulus_top_words = max(1, int(stimulus_top_words))
self._stimulus_window_index = (
int(stimulus_window) if stimulus_window is not None else None
)
self._stimulus_features = self._build_stimulus_features(seed=seed)
# ``_effective_base_r2`` depends on the active stimulus window so it
# has to be recomputed every reset(), not just on subject/condition
# changes.
self._effective_base_r2 = self._compute_effective_base_r2()
ctx = self._context_label()
self._last_feedback = f"Select the first brain region. ({ctx})" if ctx else "Select the first brain region."
return Observation(done=False, reward=0.0, metadata=self._build_observation())
def _build_stimulus_features(self, *, seed: int | None) -> dict | None:
"""Resolve the stimulus window for this episode (if data is available)."""
if not self._condition:
return None
# Deterministic key: episode varies stimulus across resets even when
# subject/run repeat, while staying reproducible for a given seed.
deterministic_key = (
self._subject_id or "_",
self._run_id or "_",
int(seed) if seed is not None else 0,
self._episode_id,
)
return summarize_stimulus_window(
self._condition,
window_index=self._stimulus_window_index,
window_size=self._stimulus_window_size,
top_words=self._stimulus_top_words,
deterministic_key=deterministic_key,
)
def _step_impl(
self,
action: Action,
timeout_s: float | None = None,
**kwargs: Any,
) -> Observation:
region_id = getattr(action, "region_id", None)
if region_id is None and isinstance(action, dict):
region_id = action.get("region_id")
result = self._process_action(str(region_id))
return Observation(
done=bool(result["done"]),
reward=float(result["reward"]),
metadata=self._build_observation(extra=result),
)
def step(self, action: Action, timeout_s: float | None = None, **kwargs: Any) -> Observation:
self._state.step_count += 1
return super().step(action, timeout_s=timeout_s, **kwargs)
async def step_async(
self,
action: Action,
timeout_s: float | None = None,
**kwargs: Any,
) -> Observation:
self._state.step_count += 1
return await super().step_async(action, timeout_s=timeout_s, **kwargs)
@property
def state(self) -> State:
return self._state
def _candidate_by_id(self) -> dict[str, RegionCandidate]:
return {candidate.region_id: candidate for candidate in self._subset.candidates}
def _selected_candidates(self) -> list[RegionCandidate]:
by_id = self._candidate_by_id()
return [by_id[region_id] for region_id in self._selected_region_ids if region_id in by_id]
def _context_label(self) -> str:
parts = [
f"subject={self._subject_id}" if self._subject_id else "",
f"run={self._run_id}" if self._run_id else "",
f"condition={self._condition}" if self._condition else "",
]
return ", ".join(p for p in parts if p)
def _condition_boost(self, candidate: RegionCandidate) -> float:
"""Per-condition multiplier so train/test conditions reward differently.
``single_m`` (single male narrator) emphasizes auditory/language ROIs;
``single_f`` does the same with a slight twist; ``mixed_*`` rewards a
broader set of association regions.
"""
condition = (self._condition or "").lower()
group = candidate.redundancy_group
if condition == "single_m":
return 1.20 if group == "auditory_temporal" else (1.05 if group == "inferior_frontal" else 0.95)
if condition == "single_f":
return 1.18 if group == "auditory_temporal" else (1.04 if group == "inferior_frontal" else 0.96)
if condition == "mixed_m":
return 1.10 if group in {"auditory_temporal", "inferior_frontal"} else 1.02
if condition == "mixed_f":
return 1.08 if group in {"auditory_temporal", "association"} else 1.0
return 1.0
def _subject_perturbation(self, candidate: RegionCandidate) -> float:
"""Deterministic per-subject jitter in [0.7, 1.3].
Hash-based so the same (subject, parcel) always gives the same value
but different subjects experience different reward landscapes - which
is what makes train/test generalization meaningful.
"""
if not self._subject_id:
return 1.0
h = abs(hash((self._subject_id, candidate.region_id))) % 10_000
return 0.7 + (h / 10_000.0) * 0.6
def _stimulus_bias(self, candidate: RegionCandidate) -> float:
"""Per-window stimulus multiplier for this candidate's group.
Bounded to a small range so it nudges the policy toward parcels
that match the current stimulus content (e.g. nouns/density →
auditory_temporal, function/syntactic words → inferior_frontal)
without overwhelming the underlying base_r2 ranking.
"""
return stimulus_bias_for_group(candidate.redundancy_group, self._stimulus_features)
def _compute_effective_base_r2(self) -> dict[str, float]:
effective: dict[str, float] = {}
for candidate in self._subset.candidates:
value = (
candidate.base_r2
* self._condition_boost(candidate)
* self._subject_perturbation(candidate)
* self._stimulus_bias(candidate)
)
effective[candidate.region_id] = float(max(0.001, min(0.30, value)))
return effective
def _score_regions(self, region_ids: list[str]) -> float:
by_id = self._candidate_by_id()
selected = [by_id[region_id] for region_id in region_ids if region_id in by_id]
if not selected:
return 0.0
total = 0.0
group_counts: dict[str, int] = {}
for candidate in selected:
group_count = group_counts.get(candidate.redundancy_group, 0)
diminishing_return = 0.72 ** group_count
base = self._effective_base_r2.get(candidate.region_id, candidate.base_r2)
total += base * diminishing_return
group_counts[candidate.redundancy_group] = group_count + 1
# Bound cumulative explained variance to keep rewards stable.
return float(1.0 - math.exp(-total))
def _process_action(self, region_id: str) -> dict:
by_id = self._candidate_by_id()
previous_r2 = self._current_r2
if self._timestep >= self._subset.selection_budget:
self._last_feedback = "Budget exhausted; episode already complete."
return self._result(
done=True,
error="budget_exhausted",
previous_r2=previous_r2,
cost_penalty=0.0,
)
if region_id not in by_id:
self._last_feedback = f"Invalid region_id={region_id}."
return self._result(
done=False,
error="invalid_region",
previous_r2=previous_r2,
cost_penalty=0.0,
)
if region_id in self._selected_region_ids:
self._last_feedback = f"Region {region_id} was already selected."
self._timestep += 1
return self._result(
done=self._is_done(),
error="duplicate_region",
previous_r2=previous_r2,
cost_penalty=0.0,
)
candidate = by_id[region_id]
self._selected_region_ids.append(region_id)
self._timestep += 1
self._current_r2 = self._score_regions(self._selected_region_ids)
delta_r2 = self._current_r2 - previous_r2
cost_penalty = self._subset.cost_penalty * candidate.cost
done = self._is_done()
reward_breakdown = score_region_action(
previous_r2=previous_r2,
current_r2=self._current_r2,
cost_penalty=cost_penalty,
error=None,
done=done,
selected_count=len(self._selected_region_ids),
)
self._last_feedback = (
f"Selected {region_id}: delta_r2={delta_r2:.4f}, "
f"reward_components=[{format_reward_breakdown(reward_breakdown.as_dict())}]."
)
return self._result(
done=done,
error=None,
previous_r2=previous_r2,
cost_penalty=cost_penalty,
reward_components=reward_breakdown.as_dict(),
)
def _is_done(self) -> bool:
return self._timestep >= self._subset.selection_budget or (
len(self._selected_region_ids) >= self._subset.n_regions
)
def _result(
self,
done: bool,
error: str | None,
previous_r2: float,
cost_penalty: float,
reward_components: dict[str, float] | None = None,
) -> dict:
if reward_components is None:
reward_components = score_region_action(
previous_r2=previous_r2,
current_r2=self._current_r2,
cost_penalty=cost_penalty,
error=error,
done=done,
selected_count=len(self._selected_region_ids),
).as_dict()
return {
"episode_id": self._episode_id,
"reward": float(reward_components["total_reward"]),
"reward_components": reward_components,
"done": bool(done),
"error": error,
"previous_r2": float(previous_r2),
"current_r2": float(self._current_r2),
"score": float(self._current_r2),
"selection_state": self._build_selection_state(),
"feedback": self._last_feedback,
}
def _build_task_info(self) -> dict:
task = TASKS[self._task]
return {
"task_name": self._task,
"description": task["description"],
"difficulty": task["difficulty"],
"reward": task["reward"],
"reward_components": reward_columns(),
"valid_actions": task["valid_actions"],
"dataset_name": self._subset.dataset_name,
"data_source": self._subset.source,
"candidate_mode": self._subset.candidate_mode,
"atlas": self._subset.atlas,
"selection_budget": int(self._subset.selection_budget),
"prompt_top_k": int(self._subset.prompt_top_k),
"cost_penalty": float(self._subset.cost_penalty),
"n_candidate_regions": int(self._subset.n_regions),
"subject_id": self._subject_id,
"run_id": self._run_id,
"condition": self._condition,
"stimulus": compact_stimulus_for_prompt(self._stimulus_features),
"stimulus_window_size": int(self._stimulus_window_size),
"stimulus_n_windows": int(
stimulus_n_windows(self._condition or "", window_size=self._stimulus_window_size)
if self._condition
else 0
),
}
def _build_selection_state(self) -> dict:
selected_set = set(self._selected_region_ids)
candidates = []
for candidate in self._subset.candidates:
item = candidate.as_dict()
item["selected"] = candidate.region_id in selected_set
candidates.append(item)
payload: dict[str, Any] = {
"episode_id": self._episode_id,
"task_name": self._task,
"timestep": int(self._timestep),
"selection_budget": int(self._subset.selection_budget),
"remaining_budget": int(max(0, self._subset.selection_budget - self._timestep)),
"selected_regions": list(self._selected_region_ids),
"current_r2": float(self._current_r2),
"candidate_regions": candidates,
"candidate_count": int(self._subset.n_regions),
"candidate_mode": self._subset.candidate_mode,
"atlas": self._subset.atlas,
"prompt_top_k": int(self._subset.prompt_top_k),
"dataset_name": self._subset.dataset_name,
"data_source": self._subset.source,
"subject_id": self._subject_id,
"run_id": self._run_id,
"condition": self._condition,
"feedback": self._last_feedback,
}
if self._stimulus_features:
payload["stimulus"] = compact_stimulus_for_prompt(self._stimulus_features)
payload["stimulus_window"] = int(self._stimulus_features.get("window_index", 0))
payload["stimulus_n_windows"] = int(self._stimulus_features.get("n_windows", 1))
else:
payload["stimulus"] = None
return payload
def _build_observation(self, extra: dict | None = None) -> dict:
payload = {
"selection_state": json.dumps(self._build_selection_state()),
"task_name": self._task,
"timestep": int(self._timestep),
"max_timesteps": int(self._subset.selection_budget),
"feedback": self._last_feedback,
"score": float(self._current_r2),
}
if extra:
payload.update(extra)
return payload
def render_text(self) -> str:
selected = ", ".join(self._selected_region_ids) or "none"
return (
f"BrainRL step={self._timestep}/{self._subset.selection_budget} "
f"r2={self._current_r2:.4f} selected=[{selected}]"
)
def load_default_candidates() -> list[dict[str, Any]]:
"""Convenience helper for scripts that only need candidate metadata."""
return candidate_table(load_brain_subset())
|