From d91cfe4d806f4c1e09c6cb75829a8625ff6506ec Mon Sep 17 00:00:00 2001 From: YurenHao0426 Date: Thu, 6 Aug 2026 12:12:41 -0500 Subject: experiment: add contrastive state-bias screen --- experiments/analyze_contrastive_bias_b1.py | 125 ++++++++++++ experiments/bootstrap_plain_cnn.sh | 4 +- experiments/contrastive_bias_b1.py | 293 +++++++++++++++++++++++++++++ experiments/contrastive_bias_smoke.py | 45 +++++ 4 files changed, 465 insertions(+), 2 deletions(-) create mode 100644 experiments/analyze_contrastive_bias_b1.py create mode 100644 experiments/contrastive_bias_b1.py create mode 100644 experiments/contrastive_bias_smoke.py (limited to 'experiments') diff --git a/experiments/analyze_contrastive_bias_b1.py b/experiments/analyze_contrastive_bias_b1.py new file mode 100644 index 0000000..527b2ba --- /dev/null +++ b/experiments/analyze_contrastive_bias_b1.py @@ -0,0 +1,125 @@ +#!/usr/bin/env python3 +"""Audit the frozen contrastive state-bias B1 gate.""" +import argparse +import json +import math +from pathlib import Path + +from contrastive_bias_b1 import RESULT_ROOT, bias_cells + + +def read_json(path): + with open(path, encoding="utf-8") as handle: + return json.load(handle) + + +def cell_path(cell_id): + return RESULT_ROOT / ("dp-bias-b1-" + cell_id + ".json") + + +def valid_record(record, cell): + history = record.get("history") or {} + return ( + record.get("status") == "completed" + and record.get("cell_id") == cell["cell_id"] + and record.get("kind") == cell["kind"] + and record.get("rule") == cell["rule"] + and record.get("ratio") == cell["ratio"] + and history.get("finite") is True + and history.get("epochs_completed") == 20 + and math.isnan(float(history.get("test_accuracy", float("nan")))) + ) + + +def metric_max(record, key): + values = record["history"]["curves"][key] + return max(abs(float(value)) for value in values) + + +def main(): + parser = argparse.ArgumentParser() + parser.add_argument( + "--out", type=Path, + default=RESULT_ROOT.parent / "b1_gate.json") + args = parser.parse_args() + cells = bias_cells() + missing = [cell["cell_id"] for cell in cells if not cell_path(cell["cell_id"]).is_file()] + if missing: + raise RuntimeError("missing B1 cells: " + ", ".join(missing)) + records = {} + for cell in cells: + record = read_json(cell_path(cell["cell_id"])) + if not valid_record(record, cell): + raise RuntimeError(f"invalid or incomplete B1 cell: {cell['cell_id']}") + records[cell["cell_id"]] = record + clean = records["clean"] + clean_acc = float(clean["history"]["final_validation_accuracy"]) + common = records["common-activity-r4-raw"] + common_acc = float(common["history"]["final_validation_accuracy"]) + common_error = metric_max( + common, "maximum_used_clean_difference_relative_error") + predictor_instruction_max = max( + metric_max(record, "instruction_observations_for_predictor") + for record in records.values() + ) + innovation_post_max = max( + metric_max(record, "post_bias_raw_bias_rms_ratio") + for cell_id, record in records.items() if "-innovation" in cell_id + ) + candidates = [] + table = [] + for ratio in (0.25, 1.0, 4.0): + tag = rate_tag = f"{ratio:g}".replace(".", "p") + row = {"ratio": ratio} + for rule in ("raw", "innovation", "oracle"): + record = records[f"activity-r{tag}-{rule}"] + row[rule] = float(record["history"]["final_validation_accuracy"]) + row["raw_degradation"] = clean_acc - row["raw"] + row["innovation_clean_gap"] = abs(row["innovation"] - clean_acc) + row["innovation_oracle_gap"] = abs(row["innovation"] - row["oracle"]) + row["innovation_post_bias_ratio_max"] = metric_max( + records[f"activity-r{tag}-innovation"], + "post_bias_raw_bias_rms_ratio") + row["passes"] = ( + row["raw_degradation"] >= 5.0 + and row["innovation_clean_gap"] <= 2.0 + and row["innovation_oracle_gap"] <= 1.0 + and row["innovation_post_bias_ratio_max"] <= 1e-3 + ) + if row["passes"]: + candidates.append(ratio) + table.append(row) + checks = { + "clean_at_least_70": clean_acc >= 70.0, + "common_within_0p2": abs(common_acc - clean_acc) <= 0.2, + "common_difference_error_at_most_1e_6": common_error <= 1e-6, + "some_activity_ratio_passes": bool(candidates), + "all_innovation_post_bias_at_most_1e_3": innovation_post_max <= 1e-3, + "predictor_saw_zero_instruction_observations": predictor_instruction_max == 0.0, + } + source_values = {json.dumps(row["source"], sort_keys=True) for row in records.values()} + registry_values = {row["registry_sha256"] for row in records.values()} + checks["single_source_lock"] = len(source_values) == 1 + checks["single_registry_lock"] = len(registry_values) == 1 + report = { + "stage": "contrastive_bias_b1", "gate": ( + "pass" if all(checks.values()) else "fail"), + "checks": checks, "clean_final_validation_accuracy": clean_acc, + "common_final_validation_accuracy": common_acc, + "common_maximum_difference_relative_error": common_error, + "innovation_maximum_post_bias_ratio": innovation_post_max, + "predictor_maximum_instruction_observations": predictor_instruction_max, + "activity_table": table, + "selected_confirmation_ratio": max(candidates) if candidates else None, + "num_expected_records": 17, "num_audited_records": len(records), + "source": clean["source"], "registry_sha256": clean["registry_sha256"], + } + args.out.parent.mkdir(parents=True, exist_ok=True) + with open(args.out, "w", encoding="utf-8") as handle: + json.dump(report, handle, indent=2, sort_keys=True) + handle.write("\n") + print(json.dumps(report, indent=2, sort_keys=True)) + + +if __name__ == "__main__": + main() diff --git a/experiments/bootstrap_plain_cnn.sh b/experiments/bootstrap_plain_cnn.sh index 0d6145f..7a1142b 100755 --- a/experiments/bootstrap_plain_cnn.sh +++ b/experiments/bootstrap_plain_cnn.sh @@ -20,8 +20,8 @@ main_root="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)" patch_root="$main_root/external/dualprop_patches" base_revision="7b2595b34421e1483a721dbfdeff8cdabda3a1ff" -if [[ "$(find "$patch_root" -maxdepth 1 -name '*.patch' | wc -l)" -ne 19 ]]; then - echo "expected 19 frozen Dual Propagation patches" >&2 +if [[ "$(find "$patch_root" -maxdepth 1 -name '*.patch' | wc -l)" -ne 20 ]]; then + echo "expected 20 frozen Dual Propagation patches" >&2 exit 1 fi diff --git a/experiments/contrastive_bias_b1.py b/experiments/contrastive_bias_b1.py new file mode 100644 index 0000000..dce9426 --- /dev/null +++ b/experiments/contrastive_bias_b1.py @@ -0,0 +1,293 @@ +#!/usr/bin/env python3 +"""Frozen 17-cell B1 contrastive state-bias screen on author Dual Prop.""" +import argparse +import hashlib +import json +import math +import os +from pathlib import Path +import subprocess +import time + +import numpy as np + + +ROOT = Path(__file__).resolve().parents[1] +PROTOCOL = ROOT / "CONTRASTIVE_BIAS.md" +RESULT_ROOT = ROOT / "results" / "contrastive_bias" / "b1" +BIAS_PATCH = ( + ROOT / "external" / "dualprop_patches" / + "0020-experiment-add-neuron-specific-bias-to-Dual-Prop.patch" +) +UPSTREAM = "7b2595b34421e1483a721dbfdeff8cdabda3a1ff" +RULES = ("raw", "innovation", "oracle") + + +def sha256(path): + digest = hashlib.sha256() + with open(path, "rb") as handle: + for chunk in iter(lambda: handle.read(1024 * 1024), b""): + digest.update(chunk) + return digest.hexdigest() + + +def git_output(repo, *args): + return subprocess.run( + ["git", *args], cwd=repo, check=True, capture_output=True, text=True + ).stdout.strip() + + +def rate_tag(value): + return f"{value:g}".replace(".", "p") + + +def bias_cells(): + cells = [{ + "cell_id": "clean", "kind": "none", "rule": "none", "ratio": 0.0, + }, { + "cell_id": "common-activity-r4-raw", "kind": "common", "rule": "raw", + "ratio": 4.0, + }] + for kind, ratios in (("fixed", (1.0, 4.0)), + ("activity", (0.25, 1.0, 4.0))): + for ratio in ratios: + for rule in RULES: + cells.append({ + "cell_id": f"{kind}-r{rate_tag(ratio)}-{rule}", + "kind": kind, "rule": rule, "ratio": ratio, + }) + if len(cells) != 17 or len({row["cell_id"] for row in cells}) != 17: + raise AssertionError("B1 registry must contain 17 unique cells") + return cells + + +def author_command(cell, author_python): + name = "dp-bias-b1-" + cell["cell_id"] + command = [ + author_python, "train.py", + "--model", "miniCNN", "--dataset", "cifar10", + "--num-epochs", "20", "--batch-size", "100", + "--learning-rate", "0.025", "--learning-rate-final", "0.025", + "--warmup-learning-rate", "0.025", "--warmup-epochs", "0", + "--decay-epochs", "20", "--momentum", "0.9", + "--weight-decay", "5e-4", "--dtype", "float32", + "--param-dtype", "float32", "--percent-train", "90", + "--percent-val", "10", "--seeds", "1988", + "--feedback-seed", "1729", "--gradient-diagnostics", "none", + "--spectral-diagnostics", "none", "--test-policy", "none", + "--early-stop-policy", "none", "--learning-algorithm", + "dualprop-lagr-ff", "--experiment-name", name, + "--optimizer-schedule", "author", "--loss", "sce", + "--alpha", "0.0", "--beta", "0.1", + "--inference-sequence", "fwK", "--inference-passes-nudged", "16", + ] + if cell["rule"] != "none": + command.extend([ + "--dp-bias-kind", cell["kind"], "--dp-bias-rule", cell["rule"], + "--dp-bias-ratio", str(cell["ratio"]), "--dp-bias-seed", "6100", + "--dp-bias-calibration-examples", "64", + ]) + return name, command + + +def jobs(author_python): + rows = [] + for cell in bias_cells(): + name, command = author_command(cell, author_python) + rows.append({ + **cell, "experiment_name": name, "command": command, + "output": str(RESULT_ROOT / (name + ".json")), + "timeout_seconds": 2 * 60 * 60, + }) + return rows + + +def registry_sha256(rows): + payload = [ + {key: value for key, value in row.items() if key != "output"} + for row in rows + ] + return hashlib.sha256(json.dumps( + payload, sort_keys=True, separators=(",", ":") + ).encode()).hexdigest() + + +def source_report(author_root): + if git_output(ROOT, "status", "--porcelain", "--untracked-files=no"): + raise RuntimeError("B1 requires clean tracked SDIL source") + if git_output(author_root, "status", "--porcelain", "--untracked-files=no"): + raise RuntimeError("B1 requires clean tracked author source") + tracked = [ + PROTOCOL, Path(__file__).resolve(), + ROOT / "experiments" / "analyze_contrastive_bias_b1.py", + BIAS_PATCH, + ] + for path in tracked: + relative = path.relative_to(ROOT) + subprocess.run( + ["git", "ls-files", "--error-unmatch", str(relative)], cwd=ROOT, + check=True, capture_output=True, + ) + return { + "sdil_commit": git_output(ROOT, "rev-parse", "HEAD"), + "author_commit": git_output(author_root, "rev-parse", "HEAD"), + "author_upstream": UPSTREAM, + "tracked_files": { + str(path.relative_to(ROOT)): sha256(path) for path in tracked + }, + } + + +def gpu_report(physical_index): + output = subprocess.run([ + "nvidia-smi", f"--id={physical_index}", + "--query-gpu=index,uuid,name,memory.total", "--format=csv,noheader,nounits", + ], check=True, capture_output=True, text=True).stdout.strip() + rows = [part.strip() for part in output.split(",")] + if len(rows) != 4 or rows[0] != str(physical_index): + raise RuntimeError(f"could not resolve physical GPU {physical_index}: {output}") + visible = os.environ.get("CUDA_VISIBLE_DEVICES") + if visible != str(physical_index): + raise RuntimeError( + f"CUDA_VISIBLE_DEVICES must equal physical GPU {physical_index}, got {visible}") + return { + "physical_index": int(rows[0]), "uuid": rows[1], "name": rows[2], + "memory_total_mib": int(rows[3]), "cuda_visible_devices": visible, + } + + +def ensure_launch(source, rows): + path = RESULT_ROOT / "launch.json" + expected = { + "stage": "contrastive_bias_b1", "source": source, + "registry_sha256": registry_sha256(rows), "num_jobs": len(rows), + "allowed_physical_gpus": [5, 7], + } + if path.is_file(): + with open(path, encoding="utf-8") as handle: + if json.load(handle) != expected: + raise RuntimeError("B1 launch lock drift") + else: + path.parent.mkdir(parents=True, exist_ok=True) + with open(path, "w", encoding="utf-8") as handle: + json.dump(expected, handle, indent=2, sort_keys=True) + handle.write("\n") + return path + + +def to_float_list(value, count): + array = np.asarray(value)[:count] + return [float(item) for item in array] + + +def summarize_hist(path): + hist = np.load(path, allow_pickle=True).item() + completed = int(hist["epochs_completed"]) + keys = ( + "val_loss", "val_accuracy", "train_loss", "train_accuracy", + "train_time", "val_time", "raw_bias_clean_difference_rms_ratio", + "post_bias_raw_bias_rms_ratio", "used_clean_difference_rms_ratio", + "maximum_used_clean_difference_relative_error", "neutral_observations", + "instruction_observations_for_predictor", + ) + curves = {key: to_float_list(hist[key], completed) for key in keys} + finite = completed == 20 and all( + math.isfinite(value) + for key in ("val_loss", "val_accuracy", "train_loss") + for value in curves[key] + ) + return { + "epochs_completed": completed, "finite": finite, + "final_validation_accuracy": curves["val_accuracy"][-1], + "best_validation_accuracy": float(hist["best_validation_accuracy"]), + "best_epoch": int(hist["best_epoch"]), + "test_accuracy": float(hist["test_accuracy"]), + "dp_bias_initialization": hist.get("dp_bias_initialization"), + "curves": curves, + } + + +def find_hist(author_root, experiment_name): + paths = list((author_root / "runs" / experiment_name).glob("*/hist.npy")) + if len(paths) != 1: + raise RuntimeError( + f"expected one history for {experiment_name}, found {len(paths)}") + return paths[0] + + +def run_job(job, author_root, source, registry_hash, gpu, dry_run): + output = Path(job["output"]) + if output.exists(): + print(f"preserving {job['cell_id']}", flush=True) + return + print("RUN", " ".join(job["command"]), flush=True) + if dry_run: + return + if git_output(ROOT, "rev-parse", "HEAD") != source["sdil_commit"]: + raise RuntimeError("SDIL commit changed after B1 launch") + if git_output(author_root, "rev-parse", "HEAD") != source["author_commit"]: + raise RuntimeError("author commit changed after B1 launch") + started = time.time() + try: + result = subprocess.run( + job["command"], cwd=author_root, timeout=job["timeout_seconds"]) + return_code = result.returncode + status = "completed" if return_code == 0 else "nonzero_exit" + except subprocess.TimeoutExpired: + return_code, status = None, "timeout" + history = None + history_path = None + if status == "completed": + try: + resolved = find_hist(author_root, job["experiment_name"]) + history_path = str(resolved) + history = summarize_hist(resolved) + except Exception as error: + status = "missing_or_invalid_history" + history = {"error": repr(error)} + record = { + **job, "stage": "contrastive_bias_b1", "source": source, + "registry_sha256": registry_hash, "hardware": gpu, "status": status, + "return_code": return_code, "driver_wall_seconds": time.time() - started, + "completed_unix_time": time.time(), "author_history": history_path, + "history": history, + } + output.parent.mkdir(parents=True, exist_ok=True) + with open(output, "w", encoding="utf-8") as handle: + json.dump(record, handle, indent=2, sort_keys=True) + handle.write("\n") + print(f"DONE status={status} {job['cell_id']}", flush=True) + + +def main(): + parser = argparse.ArgumentParser() + parser.add_argument("--author-root", type=Path, required=True) + parser.add_argument("--author-python", required=True) + parser.add_argument("--physical-gpu", type=int, choices=(5, 7), required=True) + parser.add_argument("--shard-index", type=int, default=0) + parser.add_argument("--num-shards", type=int, default=1) + parser.add_argument("--dry-run", action="store_true") + args = parser.parse_args() + args.author_root = args.author_root.resolve() + if not 0 <= args.shard_index < args.num_shards: + raise ValueError("invalid B1 shard") + rows = jobs(args.author_python) + selected = [ + row for index, row in enumerate(rows) + if index % args.num_shards == args.shard_index + ] + if args.dry_run: + for row in selected: + print(row["cell_id"], " ".join(row["command"])) + return + source = source_report(args.author_root) + gpu = gpu_report(args.physical_gpu) + launch = ensure_launch(source, rows) + print(f"B1 launch lock: {launch}", flush=True) + registry_hash = registry_sha256(rows) + for row in selected: + run_job(row, args.author_root, source, registry_hash, gpu, False) + + +if __name__ == "__main__": + main() diff --git a/experiments/contrastive_bias_smoke.py b/experiments/contrastive_bias_smoke.py new file mode 100644 index 0000000..026eb01 --- /dev/null +++ b/experiments/contrastive_bias_smoke.py @@ -0,0 +1,45 @@ +#!/usr/bin/env python3 +"""Static smoke checks for the frozen B1 registry and author patch.""" +import json +from pathlib import Path +import sys + + +ROOT = Path(__file__).resolve().parents[1] +sys.path.insert(0, str(ROOT / "experiments")) +from contrastive_bias_b1 import BIAS_PATCH, bias_cells, jobs, registry_sha256 + + +def main(): + rows = jobs("/frozen/author/python") + cells = bias_cells() + assert len(rows) == len(cells) == 17 + assert sum(row["kind"] == "common" for row in rows) == 1 + assert sum(row["kind"] == "fixed" for row in rows) == 6 + assert sum(row["kind"] == "activity" for row in rows) == 9 + assert sum(row["rule"] == "innovation" for row in rows) == 5 + for row in rows: + command = row["command"] + assert command[0] == "/frozen/author/python" + assert command[command.index("--seeds") + 1] == "1988" + assert command[command.index("--test-policy") + 1] == "none" + assert command[command.index("--num-epochs") + 1] == "20" + assert command[command.index("--model") + 1] == "miniCNN" + if row["rule"] == "none": + assert "--dp-bias-rule" not in command + else: + assert command[command.index("--dp-bias-rule") + 1] == row["rule"] + patch = BIAS_PATCH.read_text(encoding="utf-8") + for token in ( + "create_dp_bias_auxiliary", "dp_bias_differences", + "train_dp_bias_epoch", "instruction_observations_for_predictor", + ): + assert token in patch + print(json.dumps({ + "status": "passed", "num_cells": len(rows), + "registry_sha256": registry_sha256(rows), + }, indent=2, sort_keys=True)) + + +if __name__ == "__main__": + main() -- cgit v1.2.3