src/mcss/mechanism_pilot/rgbd_stream_ttt_context_v6.py
"""Core B V6: does the test-time write need new frames? Context-only TTT next to stream TTT.
PLAN_BEFORE_COREB_V6.md fixed this round. On top of Core B V4/V5's NO_STREAM, OFF and TTT, two
policies supervise the same 10 Adam steps with the 3 warmup context frames instead of the 4 stream
frames: CTX_TTT (no stream cached) and OFF_CTX_TTT (stream cached). TTT is Core B V4/V5's policy
unchanged (same generator); the new policies use generators seeded by (seed, episode, policy).
Everything is sealed before any query (Core B V5's CachedSealBarrier); RGB is clamped to [0, 1].
"""
from __future__ import annotations
import time
from pathlib import Path
import numpy as np
import torch
from mcss.dynamic.types import hash_scene_state, hash_value
from mcss.geometry import transform_cameras
from mcss.measurements import FixedMeasurementRenderer
from mcss.mechanism_pilot.direct_capacity_contracts import CapacityEvaluator
from mcss.mechanism_pilot.direct_capacity_evaluation import _metrics
from mcss.mechanism_pilot.rgbd_stream_evaluation_v3 import preseal_depth_audit
from mcss.mechanism_pilot.rgbd_stream_ttt_eval_v5 import (
CachedSealBarrier,
context_bounds,
eval_context,
)
from mcss.mechanism_pilot.rgbd_stream_ttt_evaluation_v4 import clamped
from mcss.mechanism_pilot.rgbd_stream_ttt_v4 import (
StreamRGBDLoader,
cached_episode,
episode_generator,
rebuild_with,
ttt_fast,
)
from mcss.mechanism_pilot.small_training import sha, write_json
EXPERIMENT = "EXP-3D-RGBD-DYNAMIC-TTT-CONTEXT-V6"
POLICIES = ("NO_STREAM", "CTX_TTT", "OFF", "OFF_CTX_TTT", "TTT")
CONTROLS = {"TTT_WRONG_SCENE": "TTT", "CTX_TTT_WRONG_SCENE": "CTX_TTT"}
WITH_STREAM = {"NO_STREAM": False, "CTX_TTT": False, "OFF": True, "OFF_CTX_TTT": True, "TTT": True}
def generator_for(seed, episode_id, policy):
"""TTT keeps Core B V4/V5's generator; the context policies get their own."""
if policy == "TTT":
return episode_generator(seed, episode_id)
return episode_generator(seed, f"{episode_id}:{policy}")
def unroll_v6(carrier, warmup, stream, bounds, policy, episode_id, settings, seed):
"""(final state, anchor, fast weights, episode, TTT losses) of one fixed-policy episode."""
if policy not in POLICIES:
raise ValueError(f"Unregistered policy {policy}")
episode = cached_episode(carrier, warmup, stream, bounds, episode_id, WITH_STREAM[policy])
if policy in ("NO_STREAM", "OFF"):
fast, losses = carrier.initial_fast(episode_id), []
else:
supervision = stream if policy == "TTT" else warmup
fast, losses = ttt_fast(
carrier,
episode,
supervision,
settings["steps"],
settings["learning_rate"],
generator_for(seed, episode_id, policy),
)
with torch.no_grad():
state = episode.run(fast, "materialize")
return state, episode.anchor.clone(), fast, episode, losses
def cohort_records(cohorts, expected_split):
"""{scene_id: (cohort, record, manifest, base)}: disjoint, unprotected, expected split."""
records = {}
for name, (manifest, base) in cohorts.items():
protected = set(manifest.get("protected_scene_ids", []))
for record in manifest["scenes"]:
sid = record["scene_id"]
if record["split"] != expected_split[name] or sid in protected:
raise PermissionError(f"{sid}: not an unprotected {name} record")
if sid in records:
raise PermissionError(f"{sid} appears in two cohorts")
records[sid] = (name, record, manifest, Path(base))
if len(records) < 2:
raise PermissionError("At least two scenes required for the wrong-scene controls")
return records
@torch.no_grad()
def evaluate_v6(carrier, settings, cohorts, expected_split, out, seed, device):
"""Every policy and control on every cohort scene, all sealed before any query."""
out = Path(out)
if out.exists():
raise FileExistsError("Refusing to overwrite prior Core B evaluation")
entries = cohort_records(cohorts, expected_split)
records = {sid: item[1] for sid, item in entries.items()}
names = (*POLICIES, *CONTROLS)
keys = [f"{sid}/{role}/{name}" for sid in entries for role in ("A", "B") for name in names]
barrier = CachedSealBarrier(keys)
access, metadata, episodes = [], {}, {}
before = hash_value(carrier.state_dict())
for sid, (cohort, record, manifest, base) in entries.items():
ids = tuple(r["scene_id"] for r in manifest["scenes"])
protected = tuple(manifest.get("protected_scene_ids", []))
contexts = eval_context(record, manifest, base, device, access)
streams = StreamRGBDLoader(
record,
manifest["image_size"],
base,
device,
access,
allowed_scene_ids=ids,
holdout_scene_ids=protected,
)
for role in ("A", "B"):
if streams.frame_ids(role) is None:
raise PermissionError(f"{cohort} episode {sid}/{role} lacks a full stream")
warmup, stream = contexts[role], streams.stream(role)
bounds = context_bounds(warmup, device)
episode_id = f"{cohort.lower()}:{sid}:{role}"
for policy in POLICIES:
tick = time.perf_counter()
state, anchor, fast, episode, losses = unroll_v6(
carrier, warmup, stream, bounds, policy, episode_id, settings, seed
)
key = f"{sid}/{role}/{policy}"
barrier.seal(key, sid, state.to(device="cpu"))
metadata[key] = {
"cohort": cohort,
"scene_id": sid,
"role": role,
"policy": policy,
"anchor_c2w": anchor.cpu().tolist(),
"state_hash": hash_scene_state(state),
"bounds": state.bounds.cpu().tolist(),
"warmup_frame_ids": [o.frame_id for o in warmup],
"stream_frame_ids": [o.frame_id for o in stream] if WITH_STREAM[policy] else [],
"supervised_frame_ids": (
[o.frame_id for o in stream]
if policy == "TTT"
else [o.frame_id for o in warmup]
if losses
else []
),
"fast_step": fast.step,
"fast_norm": float(fast.delta_fuse.norm() + fast.delta_complete.norm()),
"ttt_loss_first": losses[0] if losses else None,
"ttt_loss_last": losses[-1] if losses else None,
"construction_seconds": time.perf_counter() - tick,
}
if policy in CONTROLS.values():
episodes[sid, role, policy] = (episode, fast, anchor)
ids = sorted(entries)
for sid in ids:
donor = ids[(ids.index(sid) + 1) % len(ids)]
for role in ("A", "B"):
for control, policy in CONTROLS.items():
episode, _, anchor = episodes[sid, role, policy]
state = rebuild_with(episode, episodes[donor, role, policy][1])
key = f"{sid}/{role}/{control}"
barrier.seal(key, sid, state.to(device="cpu"))
metadata[key] = {
"cohort": entries[sid][0],
"scene_id": sid,
"role": role,
"policy": control,
"donor_scene_id": donor,
"donor_policy": policy,
"anchor_c2w": anchor.cpu().tolist(),
"state_hash": hash_scene_state(state),
"bounds": state.bounds.cpu().tolist(),
}
barrier.assert_ready()
preseal_depth_audit(access, records)
seal_boundary = len(access)
access.append({"event": "ALL_STREAM_STATES_SEALED", "state_count": len(keys)})
out.mkdir(parents=True)
torch.save({key: barrier.state(key) for key in keys}, out / "states.pt")
write_json(out / "state_hashes.json", metadata)
renderer = FixedMeasurementRenderer(n_samples=64, ray_chunk_size=2048).to(device)
(out / "predictions").mkdir()
rows = []
for sid, (cohort, record, manifest, base) in entries.items():
evaluator = CapacityEvaluator(
record,
manifest["image_size"],
base,
device,
barrier=barrier,
capacity_scene_ids=tuple(r["scene_id"] for r in manifest["scenes"]),
holdout_scene_ids=tuple(manifest.get("protected_scene_ids", [])),
access_log=access,
)
for role in ("A", "B"):
anchor = torch.tensor(
metadata[f"{sid}/{role}/NO_STREAM"]["anchor_c2w"],
dtype=torch.float32,
device=device,
)
for qid in record["roles"]["primary_query"]:
batch = evaluator.query(qid)
camera = transform_cameras(batch.cameras, torch.linalg.inv(anchor))
for name in names:
key = f"{sid}/{role}/{name}"
state = barrier.state(key).to(device=device)
pred, excess = clamped(renderer(state, camera, ("rgb", "depth", "visibility")))
path = out / "predictions" / f"{sid}_{role}_{qid}_{name}.npz"
np.savez_compressed(
path,
rgb=pred["rgb"][0, 0].cpu().numpy(),
depth=pred["depth"][0, 0, 0].cpu().numpy(),
opacity=pred["visibility"][0, 0, 0].cpu().numpy(),
)
rows.append(
{
"seed": seed,
"cohort": cohort,
"scene_id": sid,
"role": role,
"query_id": qid,
"policy": name,
"used_state_hash": hash_scene_state(state),
"prediction_path": str(path.relative_to(out)),
"prediction_file_sha256": sha(path),
"rgb_clamp_max_excess": excess,
**_metrics(pred, batch.rgb, batch.depth),
}
)
barrier.release()
if hash_value(carrier.state_dict()) != before:
raise AssertionError("Evaluation changed a carrier tensor")
for name, value in (
("query_results", rows),
("GT_access", access),
("seal_events", barrier.events),
):
write_json(out / f"{name}.json", value)
write_json(
out / "integrity.json",
{
"status": "PASS",
"seed": seed,
"settings": settings,
"cohorts": sorted(cohorts),
"states": len(keys),
"query_rows": len(rows),
"states_file_sha256": sha(out / "states.pt"),
"all_states_sealed_before_GT": True,
"seal_barrier": "CachedSealBarrier: full re-hash when the seal closes and at release",
"preseal_access_count": seal_boundary,
"slow_carrier_unchanged": True,
"test_time_input": "RGB+DEPTH+CAMERA of warmup and stream frames",
},
)
return rows