experiment: freeze same-path bias confirmation

1 files changed, 265 insertions, 0 deletions

diff --git a/experiments/contrastive_bias_c1.py b/experiments/contrastive_bias_c1.py
new file mode 100644
index 0000000..0ef1276
--- /dev/null
+++ b/experiments/contrastive_bias_c1.py
@@ -0,0 +1,265 @@
+#!/usr/bin/env python3
+"""Five-seed same-path full-schedule contrastive-bias confirmation."""
+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_CONFIRMATION.md"
+RESULT_ROOT = ROOT / "results" / "contrastive_bias" / "c1"
+BIAS_PATCH = (
+    ROOT / "external" / "dualprop_patches" /
+    "0020-experiment-add-neuron-specific-bias-to-Dual-Prop.patch"
+)
+UPSTREAM = "7b2595b34421e1483a721dbfdeff8cdabda3a1ff"
+SEEDS = (1989, 1990, 1991, 1992, 1993)
+CONDITIONS = (
+    ("same_path_clean", "common", "raw"),
+    ("raw", "activity", "raw"),
+    ("innovation", "activity", "innovation"),
+    ("oracle", "activity", "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 command_for(seed, condition, kind, rule, author_python):
+    name = f"dp-bias-c1-s{seed}-{condition.replace('_', '-')}"
+    command = [
+        author_python, "train.py", "--model", "miniCNN", "--dataset", "cifar10",
+        "--num-epochs", "130", "--batch-size", "100",
+        "--learning-rate", "0.025", "--learning-rate-final", "2e-6",
+        "--warmup-learning-rate", "0.001", "--warmup-epochs", "10",
+        "--decay-epochs", "120", "--momentum", "0.9",
+        "--weight-decay", "5e-4", "--dtype", "float32",
+        "--param-dtype", "float32", "--percent-train", "90",
+        "--percent-val", "10", "--seeds", str(seed),
+        "--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", "--dp-bias-kind", kind,
+        "--dp-bias-rule", rule, "--dp-bias-ratio", "4.0",
+        "--dp-bias-seed", "6100", "--dp-bias-calibration-examples", "64",
+    ]
+    return name, command
+
+
+def jobs(author_python):
+    rows = []
+    for seed in SEEDS:
+        for condition, kind, rule in CONDITIONS:
+            name, command = command_for(
+                seed, condition, kind, rule, author_python)
+            rows.append({
+                "seed": seed, "condition": condition, "kind": kind,
+                "rule": rule, "ratio": 4.0, "experiment_name": name,
+                "command": command,
+                "output": str(RESULT_ROOT / (name + ".json")),
+                "timeout_seconds": 60 * 60,
+            })
+    if len(rows) != 20 or len({row["experiment_name"] for row in rows}) != 20:
+        raise AssertionError("C1 registry must contain 20 unique cells")
+    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("C1 requires clean tracked SDIL source")
+    if git_output(author_root, "status", "--porcelain", "--untracked-files=no"):
+        raise RuntimeError("C1 requires clean tracked author source")
+    tracked = [
+        PROTOCOL, Path(__file__).resolve(),
+        ROOT / "experiments" / "analyze_contrastive_bias_c1.py", BIAS_PATCH,
+    ]
+    for path in tracked:
+        subprocess.run([
+            "git", "ls-files", "--error-unmatch", str(path.relative_to(ROOT))
+        ], 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()
+    fields = [part.strip() for part in output.split(",")]
+    visible = os.environ.get("CUDA_VISIBLE_DEVICES")
+    if len(fields) != 4 or fields[0] != str(physical_index):
+        raise RuntimeError(f"could not resolve physical GPU {physical_index}: {output}")
+    if visible != str(physical_index):
+        raise RuntimeError(
+            f"CUDA_VISIBLE_DEVICES must equal physical GPU {physical_index}, got {visible}")
+    return {
+        "physical_index": int(fields[0]), "uuid": fields[1], "name": fields[2],
+        "memory_total_mib": int(fields[3]), "cuda_visible_devices": visible,
+    }
+
+
+def ensure_launch(source, rows):
+    path = RESULT_ROOT / "launch.json"
+    expected = {
+        "stage": "contrastive_bias_c1", "source": source,
+        "registry_sha256": registry_sha256(rows), "num_jobs": len(rows),
+        "seeds": list(SEEDS), "conditions": [row[0] for row in CONDITIONS],
+        "allowed_physical_gpus": [5, 7], "test_policy": "none",
+    }
+    if path.is_file():
+        with open(path, encoding="utf-8") as handle:
+            if json.load(handle) != expected:
+                raise RuntimeError("C1 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 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: [float(value) for value in np.asarray(hist[key])[:completed]]
+        for key in keys
+    }
+    finite = completed == 130 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):
+    output = Path(job["output"])
+    if output.exists():
+        print(f"preserving s{job['seed']} {job['condition']}", flush=True)
+        return
+    if git_output(ROOT, "rev-parse", "HEAD") != source["sdil_commit"]:
+        raise RuntimeError("SDIL commit changed after C1 launch")
+    if git_output(author_root, "rev-parse", "HEAD") != source["author_commit"]:
+        raise RuntimeError("author commit changed after C1 launch")
+    print("RUN", " ".join(job["command"]), flush=True)
+    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_c1", "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} s{job['seed']} {job['condition']}", 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("--seed-shard-index", type=int, choices=(0, 1), required=True)
+    parser.add_argument("--dry-run", action="store_true")
+    args = parser.parse_args()
+    args.author_root = args.author_root.resolve()
+    rows = jobs(args.author_python)
+    selected_seeds = SEEDS[args.seed_shard_index::2]
+    selected = [row for row in rows if row["seed"] in selected_seeds]
+    if args.dry_run:
+        for row in selected:
+            print(row["seed"], row["condition"], " ".join(row["command"]))
+        return
+    source = source_report(args.author_root)
+    gpu = gpu_report(args.physical_gpu)
+    launch = ensure_launch(source, rows)
+    print(f"C1 launch lock: {launch}", flush=True)
+    registry_hash = registry_sha256(rows)
+    for row in selected:
+        run_job(row, args.author_root, source, registry_hash, gpu)
+
+
+if __name__ == "__main__":
+    main()