File size: 10,103 Bytes
2f4aa7f | 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 | #!/usr/bin/env python3
"""
Does inference-depth degradation come from step-embedding extrapolation (B)
or from genuine recurrent saturation (A)?
HYPOTHESIS
The core applies `step_emb[min(step, max_steps)]`. Beyond `max_steps` every
additional step receives the SAME embedding as the last trained step, so the
core gets an out-of-distribution "which step am I on" signal. If that is what
breaks deep inference, then removing the step embedding entirely (making the
core genuinely step-invariant) should change the shape of the depth curve.
DESIGN (weights are FIXED and IDENTICAL in every condition)
condition `clamped` : stock behaviour, step_emb[min(step, max_steps)]
condition `zeroed` : step_emb forced to 0 at every step -> step-invariant
condition `frozen_k` : every step uses step_emb[k] for a fixed in-distribution
k, isolating "OOD signal" from "no signal at all"
Each condition is evaluated at depths 1/2/4/8/16/32 against a FIXED opponent
pool (random-legal and one-ply material), with identical games, temperature,
seed and adjudication. Only the depth and the embedding policy vary.
DISCRIMINATION
A (saturation) -> all conditions degrade at the same depth
B (OOD step embeddings) -> `zeroed`/`frozen_k` keep improving where `clamped`
degrades
Anything else -> report as ambiguous
Also records a weight-free signal: policy agreement with the depth-1 policy and
mean entropy per depth, to show whether deep inference is drifting or collapsing.
"""
from __future__ import annotations
import argparse
import contextlib
import json
import os
import sys
import time
import torch
sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
from tinychess.arena import (MaterialAgent, RandomAgent, elo_from_score,
head_to_head, wilson_interval)
from tinychess.arena import policy_diagnostics
from tinychess.batching import encode_batch
from tinychess.checkpoint import load_model_only
@contextlib.contextmanager
def step_embedding_mode(model, mode: str, k: int = 0):
"""Temporarily replace the core's step embedding policy. Weights unchanged."""
core = model.core
original = core.step_emb.data.clone()
try:
if mode == "clamped":
pass # stock behaviour
elif mode == "zeroed":
core.step_emb.data.zero_() # step-invariant core
elif mode == "frozen":
row = original[min(k, original.shape[0] - 1)].clone()
core.step_emb.data[:] = row # same in-distribution signal
else:
raise ValueError(mode)
yield
finally:
core.step_emb.data.copy_(original)
def make_boards(n=48, seed=0):
"""Fixed, reproducible mid-game position set (same recipe as exp_compute_scaling)."""
import random
import chess
rng = random.Random(seed)
boards = []
while len(boards) < n:
b = chess.Board()
for _ in range(rng.randint(0, 40)):
ms = list(b.legal_moves)
if not ms:
break
b.push(rng.choice(ms))
if not b.is_game_over():
boards.append(b)
return boards
def policy_stats(model, boards, steps):
"""Top-1 move ids, entropy, top-1 prob and illegal mass at a given depth."""
pb = encode_batch(boards)
with torch.no_grad():
out = model(**pb.model_args(), steps=steps)
logits = out["logits"]
d = policy_diagnostics(model, boards, steps=steps)
return logits.argmax(-1), d["mean_entropy"], d["mean_top1_prob"], d["illegal_prob_mass"]
def main():
ap = argparse.ArgumentParser()
ap.add_argument("--ckpt", required=True)
ap.add_argument("--depths", default="1,2,4,8,16,32")
ap.add_argument("--modes", default="clamped,zeroed,frozen")
ap.add_argument("--frozen-k", type=int, default=-1,
help="step index for 'frozen' mode; -1 = last trained step")
ap.add_argument("--games", type=int, default=40)
ap.add_argument("--max-plies", type=int, default=140)
ap.add_argument("--temperature", type=float, default=0.3)
ap.add_argument("--adjudicate", default="material",
choices=["material", "none"])
ap.add_argument("--seed", type=int, default=0)
ap.add_argument("--out", default="runs/step_embedding.json")
a = ap.parse_args()
depths = [int(x) for x in a.depths.split(",")]
modes = [m.strip() for m in a.modes.split(",")]
model, cfg = load_model_only(a.ckpt)
rep = model.param_report()
# `cfg.max_steps` is the TRAINED depth. Capture it before any mutation:
# model.forward clamps n to cfg.max_steps, so without lifting the cap every
# depth > max_steps silently runs only max_steps iterations.
trained_max_steps = int(cfg.max_steps)
max_steps = trained_max_steps
if max(depths) > trained_max_steps:
import torch.nn as nn
old = model.core.step_emb.data
new = torch.zeros(max(depths) + 1, old.shape[1])
new[:old.shape[0]] = old
new[old.shape[0]:] = old[-1] # stock behaviour: repeat last row
model.core.step_emb = nn.Parameter(new)
model.cfg.max_steps = max(depths)
model.core.cfg.max_steps = max(depths)
print(f"[lifted] step cap {trained_max_steps} -> {max(depths)} "
f"(step_emb rows {old.shape[0]} -> {new.shape[0]})")
k = a.frozen_k if a.frozen_k >= 0 else trained_max_steps
print(f"ckpt={a.ckpt}")
print(f"params={rep['total']:,} trained max_steps={trained_max_steps} frozen_k={k}")
print(f"depths={depths} modes={modes} games={a.games} adj={a.adjudicate}\n")
boards = make_boards(48, seed=a.seed)
opponents = {"random_legal": RandomAgent(), "material_greedy": MaterialAgent()}
res = {"checkpoint": a.ckpt, "params": rep["total"],
"trained_max_steps": trained_max_steps,
"frozen_k": k, "args": vars(a), "conditions": {}}
ref_top1 = {}
for mode in modes:
res["conditions"][mode] = {}
with step_embedding_mode(model, mode, k):
for d in depths:
entry = {}
t0 = time.time()
top1, ent, top1p, illegal = policy_stats(model, boards, d)
entry["mean_entropy"] = round(ent, 4)
entry["mean_top1_prob"] = round(top1p, 4)
entry["illegal_prob_mass"] = float(illegal)
entry["ms_per_position"] = round(
1000 * (time.time() - t0) / len(boards), 3)
if d == depths[0]:
ref_top1[mode] = top1
entry["top1_agreement_vs_depth1"] = round(
float((top1 == ref_top1[mode]).float().mean()), 4)
for oname, opp in opponents.items():
h = head_to_head(model, opp, n_games=a.games, steps=d,
max_plies=a.max_plies,
temperature=a.temperature,
adjudicate=a.adjudicate, seed=100 + a.seed)
ci = wilson_interval(h["wins"], h["draws"], h["games"])
entry[oname] = {
"score": round(h["score"], 4),
"ci95": [round(ci[0], 3), round(ci[1], 3)],
"elo": round(elo_from_score(h["score"]), 1),
"W": h["wins"], "D": h["draws"], "L": h["losses"],
"mean_plies": round(h["mean_plies"], 1),
}
res["conditions"][mode][str(d)] = entry
r = entry["random_legal"]; mt = entry["material_greedy"]
print(f" [{mode:8s}] d={d:3d} rand={r['score']:.3f}"
f"{str(r['ci95']):16s} mat={mt['score']:.3f}"
f"{str(mt['ci95']):16s} ent={ent:.2f} "
f"agree_d1={entry['top1_agreement_vs_depth1']:.2f} "
f"{entry['ms_per_position']:.1f}ms")
# ---- verdict ----
def curve(mode, opp):
return [(d, res["conditions"][mode][str(d)][opp]["score"]) for d in depths]
verdict = {}
for opp in opponents:
v = {}
for mode in modes:
c = curve(mode, opp)
best_d, best_s = max(c, key=lambda t: t[1])
deep = [s for d, s in c if d > trained_max_steps]
shallow = [s for d, s in c if d <= trained_max_steps]
ms = (sum(shallow) / len(shallow)) if shallow else None
md = (sum(deep) / len(deep)) if deep else None
v[mode] = {"curve": c, "best_depth": best_d, "best_score": best_s,
"n_shallow": len(shallow), "n_deep": len(deep),
"mean_shallow": round(ms, 4) if ms is not None else None,
"mean_deep": round(md, 4) if md is not None else None,
"deep_minus_shallow": (round(md - ms, 4)
if (ms is not None and md is not None)
else None)}
verdict[opp] = v
res["verdict"] = verdict
print("\n=== deep-vs-shallow (mean score beyond trained max_steps minus within) ===")
for opp, v in verdict.items():
print(f" {opp}:")
for mode, d in v.items():
if d["deep_minus_shallow"] is None:
print(f" {mode:8s} best_depth={d['best_depth']:3d} "
f"shallow={d['mean_shallow']} deep=n/a "
f"(no depths beyond trained max_steps)")
else:
print(f" {mode:8s} best_depth={d['best_depth']:3d} "
f"shallow={d['mean_shallow']:.3f} deep={d['mean_deep']:.3f} "
f"delta={d['deep_minus_shallow']:+.3f}")
os.makedirs(os.path.dirname(os.path.abspath(a.out)), exist_ok=True)
json.dump(res, open(a.out, "w"), indent=2)
print(f"\n[saved] {a.out}")
if __name__ == "__main__":
main()
|