scripts/evaluate_rgbd_long_train_eval.py
#!/usr/bin/env python3
"""V17 EVAL stage: every prediction is sealed to disk before any query depth is read.
Derived from the V14 EVAL stage. Checkpoints are labelled by step: T<step> is the L checkpoint
at that step and SEL the one the frozen DEV rule selected over all 18 checkpoints.
1. seal --cohort NAME: for each scene and role, the context-only RGB-D loader gives the context
and its CONTEXT_DEPTH bounds; every registered checkpoint builds states and renders depth and
opacity at every query camera (primary cohorts: the curve steps and SEL; EVAL-FAR and
FRESH-V2: T6000, T24000 and SEL); the context-depth reprojection (REPROJ_NN, hit, NEAR mask)
and INPAINT-V1's harmonic filling come from the same context. Every prediction file is hashed.
2. score: query depth is read only after every cohort's seal exists; AbsRel on ALL, NEAR and FAR.
3. analyze: per-scene values over the pooled EVAL-V3 + EVAL-V4 scenes and the preregistered
contrasts LENGTH_GAIN and HARMONIC_MINUS_AHFILL8_T24000 (primary) plus descriptive ones; every
cohort is also reported on its own, with the FAR/NEAR/ALL curve over training steps.
"""
from __future__ import annotations
import argparse
import functools
import hashlib
import importlib.util
import json
from pathlib import Path
import numpy as np
import torch
from mcss.dynamic.types import hash_scene_state
from mcss.geometry import transform_cameras
from mcss.measurements import FixedMeasurementRenderer
from mcss.mechanism_pilot.geometry_carrier_experiment import read, verify_lock
from mcss.mechanism_pilot.readout_reanalysis import depth_metrics
from mcss.mechanism_pilot.rgbd_depth_bounds import context_depth_bounds
from mcss.mechanism_pilot.rgbd_long_train_carrier import build_state, load_checkpoint
from mcss.mechanism_pilot.small_training import sha, write_json
from mcss.types import Cameras
REGIONS = ("ALL", "NEAR", "FAR")
PRIMARY_COHORTS = ("EVAL_V3", "EVAL_V4")
SCRIPTS = Path(__file__).resolve().parent
CONTRASTS = {
"LENGTH_GAIN": ("T6000", "FAR", "T24000", "FAR"),
"HARMONIC_MINUS_AHFILL8_T24000": ("REPROJ_HARMONIC", "ALL", "AHFILL8_T24000", "ALL"),
"HARMONIC_MINUS_HFILL8_T24000": ("REPROJ_HARMONIC", "ALL", "HFILL8_T24000", "ALL"),
"HARMONIC_MINUS_AHFILL8_T6000": ("REPROJ_HARMONIC", "ALL", "AHFILL8_T6000", "ALL"),
"HARMONIC_MINUS_HFILL8_T6000": ("REPROJ_HARMONIC", "ALL", "HFILL8_T6000", "ALL"),
"HARMONIC_MINUS_T24000_FAR": ("REPROJ_HARMONIC", "FAR", "T24000", "FAR"),
"HARMONIC_MINUS_T6000_FAR": ("REPROJ_HARMONIC", "FAR", "T6000", "FAR"),
"AHFILL8_LENGTH_GAIN": ("AHFILL8_T6000", "ALL", "AHFILL8_T24000", "ALL"),
"HFILL8_LENGTH_GAIN": ("HFILL8_T6000", "ALL", "HFILL8_T24000", "ALL"),
"T6000_MINUS_T24000_ALL": ("T6000", "ALL", "T24000", "ALL"),
"T6000_MINUS_T24000_NEAR": ("T6000", "NEAR", "T24000", "NEAR"),
"T6000_MINUS_SEL_FAR": ("T6000", "FAR", "SEL", "FAR"),
"HARMONIC_MINUS_AHFILL8_SEL": ("REPROJ_HARMONIC", "ALL", "AHFILL8_SEL", "ALL"),
"CONSTANT_MINUS_T24000": ("CONSTANT", "ALL", "T24000", "ALL"),
"NN_MINUS_T24000": ("REPROJ_NN", "ALL", "T24000", "ALL"),
}
DEFINITIONS = {
"LENGTH_GAIN": "FAR AbsRel(L at step 6000) - FAR AbsRel(L at step 24000), "
"pooled EVAL-V3 + EVAL-V4",
"HARMONIC_MINUS_AHFILL8_T24000": "AbsRel(REPROJ_HARMONIC) - AbsRel(AHFILL8 of L at step 24000)",
"FAR": "valid pixels farther than 8 px from any context-depth reprojection hit",
"HFILL8": "REPROJ_HARMONIC within 8 px of a hit, the carrier elsewhere",
"AHFILL8": "HFILL8 whose FAR pixels of rendered opacity < 0.5 also take REPROJ_HARMONIC",
"T<step>": "the L checkpoint at that step; T6000 equals V11 C1's last step (built-in control)",
"SEL": "the checkpoint the frozen V2 DEV rule selected over all 18 checkpoints",
}
def module(name, filename):
spec = importlib.util.spec_from_file_location(name, SCRIPTS / filename)
loaded = importlib.util.module_from_spec(spec)
spec.loader.exec_module(loaded)
return loaded
@functools.cache
def host():
"""V11's frozen EVAL-V3 module: context loader, aggregation, bootstrap and V2 gate."""
return module("v17_host_v11_eval", "evaluate_rgbd_completion_eval_v3.py")
@functools.cache
def inpainting():
"""INPAINT-V1's frozen module: ray cosine, harmonic filling, NEAR mask, label rule."""
return module("v17_inpaint", "run_rgbd_inpaint_baselines.py")
def cohorts(root):
"""The frozen contract and {cohort: (entry, manifest, base)}, manifests hash-checked."""
config = read(root / "training_contract.json")
frozen = {}
for name, entry in config["eval"]["cohorts"].items():
path = Path(entry["manifest"])
if sha(path) != entry["manifest_sha256"]:
raise PermissionError(f"{name} manifest changed after the freeze")
manifest = read(path)
if any(r["split"] != entry["split"] for r in manifest["scenes"]):
raise PermissionError(f"{name} records only")
frozen[name] = (entry, manifest, path.parent)
return config, frozen
def labels_for(name, config):
steps = (
config["eval"]["curve_steps_primary"]
if name in PRIMARY_COHORTS
else config["eval"]["steps_descriptive_cohorts"]
)
return [f"T{step}" for step in steps] + ["SEL"]
def checkpoint_path(root, seed, label):
"""The hash-checked checkpoint behind a label, from the run's own DEV curve and selection."""
run = root / "checkpoints" / f"L_{seed}"
curve = {c["step"]: c for c in read(run / "dev_curve.json")}
if label == "SEL":
selection = read(run / "selection.json")
step = selection["selected_step"]
if selection["checkpoint_sha256"] != curve[step]["checkpoint_sha256"]:
raise PermissionError(f"Selection and DEV curve disagree: seed {seed}")
else:
step = int(label[1:])
item = curve[step]
if sha(Path(item["checkpoint"])) != item["checkpoint_sha256"]:
raise PermissionError(f"Checkpoint changed: seed {seed} step {step}")
return Path(item["checkpoint"]), step
def carriers(root, config, labels):
"""{(label, seed): model} for the registered checkpoints (hash- and lock-checked)."""
models, steps = {}, {}
lock = sha(root / "preregistration.json")
for seed in config["seeds"]:
for label in labels:
path, step = checkpoint_path(root, seed, label)
model, payload = load_checkpoint(path, "cpu")
if payload["lock_sha256"] != lock or payload["step"] != step:
raise PermissionError("Checkpoint not from the frozen V17 training")
models[label, seed] = model.eval()
steps[f"{label}_{seed}"] = step
return models, steps
@torch.no_grad()
def seal(root, cohort):
root = Path(root).resolve()
verify_lock(root)
if read(root / "audit/v11_prefix_control.json")["status"] != "PASS":
raise PermissionError("The built-in V11 C1 control must PASS before any EVAL prediction")
config, frozen = cohorts(root)
entry, manifest, base = frozen[cohort]
out = root / "raw" / "eval" / cohort
if (out / "prediction_seal.json").exists():
raise FileExistsError("EVAL predictions are sealed once")
v11, inp = host(), inpainting()
geo = v11.geometry()
labels = labels_for(cohort, config)
models, steps = carriers(root, config, labels)
renderer = FixedMeasurementRenderer(n_samples=64, ray_chunk_size=2048)
constant = config["eval"]["reprojection_fill_constant_m"]
if config["eval"]["near_px"] != inp.NEAR_PX:
raise PermissionError("NEAR definition must be V11's")
access, entries, states = [], {}, {}
for record in manifest["scenes"]:
sid = record["scene_id"]
for role, context in v11.context_of(record, manifest, base, access).items():
anchor = context[0].camera.c2w
bounds = context_depth_bounds(context, anchor)
points = geo.context_points(context)
built = {}
for (label, seed), carrier in models.items():
episode = f"{cohort.lower()}:{sid}:{role}"
state, state_anchor = build_state(carrier, context, bounds, episode)
built[label, seed] = (state, state_anchor)
states[f"{cohort}:{label}:{seed}:{sid}/{role}"] = hash_scene_state(state)
for qid in record["roles"][entry["role"]]:
access.append(
{
"cohort": cohort,
"scene_id": sid,
"frame_id": qid,
"purpose": "QUERY_CAMERA_ONLY",
}
)
camera = geo.query_camera(record, qid, manifest["image_size"])
depth, hit = geo.reproject(points, camera)
reproj = geo.fill_nearest(depth, hit, constant) if hit.any() else None
if reproj is None:
reproj = np.full(camera.image_size, constant, dtype=np.float64)
cosine = inp.ray_cosine(camera)
arrays = {
"REPROJ_NN": reproj,
"hit": hit,
"near": inp.near_mask(hit),
"REPROJ_HARMONIC": inp.harmonic_fill(depth, hit, cosine, constant),
}
batched = Cameras(
camera.intrinsics[None, None], camera.c2w[None, None], camera.image_size
)
for (label, seed), (state, state_anchor) in built.items():
local = transform_cameras(batched, torch.linalg.inv(state_anchor))
pred = renderer(state, local, ("depth", "visibility"))
arrays[f"{label}_{seed}"] = pred["depth"][0, 0, 0].numpy().astype(np.float64)
arrays[f"OPACITY_{label}_{seed}"] = pred["visibility"][0, 0, 0].numpy()
path = out / "predictions" / f"{sid}_{role}_{qid}.npz"
path.parent.mkdir(parents=True, exist_ok=True)
np.savez_compressed(path, **arrays)
entries[f"{cohort}/{sid}/{role}/{qid}"] = {"path": str(path), "sha256": sha(path)}
if any(
"depth" in e.get("channels", []) and e.get("purpose") != "CONTEXT_ONLY_RGBD" for e in access
):
raise AssertionError("Only context depth may be read before the seal")
write_json(out / "state_hashes.json", states)
write_json(out / "GT_access_before_seal.json", access)
write_json(
out / "prediction_seal.json",
{
"cohort": cohort,
"labels": labels,
"checkpoint_steps": steps,
"predictions": entries,
"state_hashes_sha256": sha(out / "state_hashes.json"),
"query_depth_read": False,
"near_px": config["eval"]["near_px"],
"reprojection_fill_constant_m": constant,
},
)
print("EVAL_PREDICTIONS_SEALED", cohort, len(entries), flush=True)
def absrel(value, gt, mask):
return float(np.mean(np.abs(value[mask] - gt[mask]) / gt[mask])) if mask.any() else None
def score(root):
root = Path(root).resolve()
config, frozen = cohorts(root)
out = root / "raw" / "eval"
seals = {name: read(out / name / "prediction_seal.json") for name in frozen} # all sealed first
threshold = config["eval"]["abstain_opacity_threshold"]
rows, access = [], []
for name, (entry, manifest, _) in frozen.items():
seal_file = seals[name]
constant = seal_file["reprojection_fill_constant_m"]
for record in manifest["scenes"]:
sid = record["scene_id"]
frames = {f["frame_id"]: f for f in record["frames"]}
for role in ("A", "B"):
for qid in record["roles"][entry["role"]]:
item = seal_file["predictions"][f"{name}/{sid}/{role}/{qid}"]
if sha(Path(item["path"])) != item["sha256"]:
raise PermissionError("A sealed prediction changed")
arrays = dict(np.load(item["path"]))
access.append(
{
"cohort": name,
"scene_id": sid,
"frame_id": qid,
"purpose": "POST_SEAL_QUERY_DEPTH",
}
)
gt = np.load(frames[qid]["depth"]).squeeze().astype(np.float64)
valid = np.isfinite(gt) & (gt > 0)
near = arrays["near"].astype(bool)
masks = {"ALL": valid, "NEAR": valid & near, "FAR": valid & ~near}
harmonic = arrays["REPROJ_HARMONIC"]
predictions = {
("CONSTANT", 0): np.full(gt.shape, constant, dtype=np.float64),
("REPROJ_NN", 0): arrays["REPROJ_NN"],
("REPROJ_HARMONIC", 0): harmonic,
}
for label in seal_file["labels"]:
for seed in config["seeds"]:
carrier = arrays[f"{label}_{seed}"]
low = arrays[f"OPACITY_{label}_{seed}"].astype(np.float64) < threshold
predictions[label, seed] = carrier
predictions[f"HFILL8_{label}", seed] = np.where(near, harmonic, carrier)
predictions[f"AHFILL8_{label}", seed] = np.where(
near | low, harmonic, carrier
)
common = {"cohort": name, "scene_id": sid, "role": role, "query_id": qid}
for (method, seed), value in predictions.items():
row = {**common, "method": method, "seed": seed}
for region, mask in masks.items():
row[f"absrel_{region}"] = absrel(value, gt, mask)
row["depth_absrel"] = depth_metrics(value, gt)["depth_absrel"]
row["far_fraction"] = float(masks["FAR"].sum() / max(valid.sum(), 1))
rows.append(row)
write_json(out / "query_rows.json", rows)
write_json(out / "GT_access_after_seal.json", access)
print("EVAL_SCORED", len(rows), flush=True)
def block(rows):
"""Absolute errors and every contrast over one set of scenes (V11 aggregation, labels)."""
v11, label = host(), inpainting().geometry().label
def values(method, region):
return v11.scene_values(rows, method, f"absrel_{region}")
methods = sorted({r["method"] for r in rows})
contrasts = {}
for name, (a, ra, b, rb) in CONTRASTS.items():
if a not in methods or b not in methods:
continue
stats = v11.contrast(values(a, ra), values(b, rb))
contrasts[name] = {"gain": stats, "definition": f"AbsRel {ra}({a}) - AbsRel {rb}({b})"}
if stats is not None:
contrasts[name]["label"] = label(stats)
contrasts[name]["gate"], contrasts[name]["gate_checks"] = v11.gate(stats)
curve = {}
for m in methods:
if m.startswith("T") and m[1:].isdigit():
curve[int(m[1:])] = {region: v11.summary(values(m, region)) for region in REGIONS}
return {
"absolute_absrel": {
m: {region: v11.summary(values(m, region)) for region in REGIONS} for m in methods
},
"contrasts": contrasts,
"curve": {str(k): curve[k] for k in sorted(curve)},
"n_scenes": len({r["scene_id"] for r in rows}),
"far_pixel_fraction": float(
np.mean([r["far_fraction"] for r in rows if r["method"] == "REPROJ_NN"])
),
}
def analyze_rows(rows):
primary = block([r for r in rows if r["cohort"] in PRIMARY_COHORTS])
contrasts = primary["contrasts"]
statuses = {
"LENGTH_STATUS": contrasts.get("LENGTH_GAIN", {}).get("gate", "UNDEFINED"),
"AHFILL8_LABEL": contrasts.get("HARMONIC_MINUS_AHFILL8_T24000", {}).get(
"label", "UNDEFINED"
),
}
first = statuses["LENGTH_STATUS"] == "SUPPORTED"
second = statuses["AHFILL8_LABEL"] == "ABOVE"
names = sorted({r["cohort"] for r in rows})
return {
"primary": primary,
"cohorts": {name: block([r for r in rows if r["cohort"] == name]) for name in names},
"statuses": statuses,
"INTERPRETATION_BRANCH": "A" if first and second else "B" if first or second else "C",
"definitions": DEFINITIONS,
}
def analyze(root, output=None):
root = Path(root).resolve()
rows_path = root / "raw" / "eval" / "query_rows.json"
result = analyze_rows(read(rows_path))
result["control"] = read(root / "audit/v11_prefix_control.json")["status"]
result["input_sha256"] = {str(rows_path): sha(rows_path)}
result["source_sha256"] = {
str(Path(__file__).resolve()): hashlib.sha256(Path(__file__).read_bytes()).hexdigest()
}
target = Path(output) if output else root / "eval_results.json"
write_json(target, result)
print(
"EVAL_BRANCH", result["INTERPRETATION_BRANCH"], json.dumps(result["statuses"]), flush=True
)
return result
if __name__ == "__main__":
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument("--root", type=Path, required=True)
parser.add_argument("--stage", choices=("seal", "score", "analyze"), required=True)
parser.add_argument("--cohort")
parser.add_argument("--output", type=Path)
args = parser.parse_args()
torch.set_num_threads(1)
if args.stage == "seal":
seal(args.root, args.cohort)
elif args.stage == "score":
score(args.root)
else:
analyze(args.root, args.output)