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
|
#!/usr/bin/env python3
"""Assemble the anonymized supplementary archive for the paper.
The archive contains the experiment code, pinned dependency versions, every
row-level CSV referenced by the paper's artifact map, the generated LaTeX
table fragments, the figure files, a machine readable seed manifest, and a
SHA256 manifest over every packaged file. Research notes, draft papers, raw
datasets, and git history are excluded.
"""
from __future__ import annotations
import argparse
import csv
import glob
import hashlib
import json
import shutil
import zipfile
from pathlib import Path
CODE_FILES = ["scripts", "requirements.txt", "requirements-lock.txt"]
OUTPUT_GLOBS = [
"outputs/aaai_depth_experiments/*.csv",
"outputs/aaai_depth_experiments/summary.json",
"outputs/cnn_initialization_validation/*",
"outputs/real_data_validation_mnist/*.csv",
"outputs/real_data_validation_mnist/summary.json",
"outputs/soft_erosion_theory_vs_empirical/*.csv",
"outputs/learning_rate_compensation/*",
"outputs/parameter_matched_depth_control/*",
"outputs/teacher_test_gap_sgd_T8000/runs.csv",
"outputs/teacher_test_gap_sgd_T8000/width_summary.csv",
"outputs/teacher_test_gap_sgd_T8000/summary.json",
"outputs/operator_stress_grid_summary/*.csv",
"outputs/compressed_operator_s20_256traj_T50_width64_N*/*",
"outputs/compressed_operator_retangent_T50_width64_N*/*",
"outputs/compressed_operator_T50_width64_N*/*",
"outputs/early_directional_predictors_T50_width64_N*/*",
"outputs/early_kernel_predictors_T50_N128_32runs/*",
"outputs/finite_time_kernel_probe_T50_N128_8runs/*.csv",
"outputs/finite_time_kernel_probe_T50_N128_8runs/summary.json",
"outputs/fa_tangent_hierarchy_derivative_probe_eps005/*",
"outputs/aaai_main_figures/*",
"outputs/aaai_supplementary/*",
"outputs/tables/*.tex",
]
SEED_SOURCES = [
("depth_initialization", "outputs/aaai_depth_experiments/initialization_rows.csv"),
("depth_finite_time", "outputs/aaai_depth_experiments/finite_time_rows.csv"),
("cnn_initialization", "outputs/cnn_initialization_validation/cnn_initialization_rows.csv"),
("mnist_initialization", "outputs/real_data_validation_mnist/e0_mnist_rows.csv"),
("mnist_estimator", "outputs/real_data_validation_mnist/estimator_mnist_rows.csv"),
("learning_rate_evaluation", "outputs/learning_rate_compensation/evaluation_rows.csv"),
("learning_rate_tuning", "outputs/learning_rate_compensation/tuning_rows.csv"),
("parameter_matched", "outputs/parameter_matched_depth_control/parameter_matched_rows.csv"),
("teacher_task", "outputs/teacher_test_gap_sgd_T8000/runs.csv"),
("stress_grid", "outputs/operator_stress_grid_summary/stress_grid_rows.csv"),
]
README = """\
# Supplementary code and data
This archive accompanies the anonymous submission "The Optimization Cost of
Fixed Random Feedback". It contains the experiment code, the row-level CSV
outputs behind every reported number, the generated LaTeX table fragments,
the figure files, a seed manifest, and a SHA256 manifest.
## Environment
Python 3.13, CPU, double precision. Pinned versions in
`requirements-lock.txt`. Install with:
pip install -r requirements-lock.txt
## Reproduction
Appendix K of the paper lists the exact command for every experiment.
The shared self tests are:
python scripts/test_feedback_rules.py
python scripts/cnn_initialization_validation.py --self-test
## Figure and table map
| Paper object | Generating script | Source data |
|---|---|---|
| Figures 1-2 | scripts/plot_aaai_main_figures.py | outputs/aaai_depth_experiments/, outputs/soft_erosion_theory_vs_empirical/, outputs/cnn_initialization_validation/, outputs/real_data_validation_mnist/ |
| Figure 3 | scripts/cnn_initialization_validation.py | outputs/cnn_initialization_validation/ |
| Figure 4 | scripts/finite_time_supplement.py | outputs/aaai_depth_experiments/finite_time_rows.csv |
| Figure 5 | scripts/learning_rate_compensation.py | outputs/learning_rate_compensation/ |
| Figure 6 | scripts/parameter_matched_depth_control.py | outputs/parameter_matched_depth_control/ |
| Figure 7 | scripts/downstream_capacity_sweep.py + scripts/plot_teacher_test_gap.py | outputs/teacher_test_gap_sgd_T8000/ |
| Generated appendix tables | scripts/generate_paper_tables.py, scripts/finite_time_supplement.py | outputs/tables/*.tex |
## Manifests
- `artifacts/seed_manifest.csv`: every (experiment, init seed, feedback seed)
pair present in the row-level outputs, with the shared data seed per
experiment where recorded.
- `artifacts/manifest.json`: byte size and SHA256 for every file in this
archive (excluding the manifest itself).
The CSV files are the authoritative source for all reported experimental
numbers; no table value in the paper is transcribed manually.
"""
def parse_args() -> argparse.Namespace:
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument("--outdir", type=Path, default=Path("supplement_build"))
return parser.parse_args()
def sha256(path: Path) -> str:
digest = hashlib.sha256()
with path.open("rb") as handle:
for chunk in iter(lambda: handle.read(1 << 20), b""):
digest.update(chunk)
return digest.hexdigest()
def seed_manifest(root: Path) -> list[dict[str, str]]:
rows: list[dict[str, str]] = []
for experiment, rel in SEED_SOURCES:
path = Path(rel)
if not path.exists():
continue
data_seed = ""
summary = path.parent / "summary.json"
if summary.exists():
config = json.load(summary.open()).get("config", {})
data_seed = str(config.get("data_seed", ""))
seen: set[tuple[str, str]] = set()
for row in csv.DictReader(path.open()):
init_seed = row.get("init_seed", "")
feedback_seed = row.get("feedback_seed", "")
key = (init_seed, feedback_seed)
if key in seen:
continue
seen.add(key)
rows.append(
{
"experiment": experiment,
"data_seed": data_seed,
"init_seed": init_seed,
"feedback_seed": feedback_seed,
}
)
return rows
def main() -> None:
args = parse_args()
stage = args.outdir / "supplement"
if stage.exists():
shutil.rmtree(stage)
stage.mkdir(parents=True)
for item in CODE_FILES:
src = Path(item)
if src.is_dir():
shutil.copytree(src, stage / src.name)
elif src.exists():
shutil.copy2(src, stage / src.name)
for pattern in OUTPUT_GLOBS:
for match in sorted(glob.glob(pattern)):
src = Path(match)
if not src.is_file():
continue
dest = stage / src
dest.parent.mkdir(parents=True, exist_ok=True)
shutil.copy2(src, dest)
(stage / "README.md").write_text(README)
artifacts = stage / "artifacts"
artifacts.mkdir(exist_ok=True)
seeds = seed_manifest(Path("."))
with (artifacts / "seed_manifest.csv").open("w", newline="") as handle:
writer = csv.DictWriter(
handle, fieldnames=["experiment", "data_seed", "init_seed", "feedback_seed"]
)
writer.writeheader()
writer.writerows(seeds)
manifest = {
str(path.relative_to(stage)): {"bytes": path.stat().st_size, "sha256": sha256(path)}
for path in sorted(stage.rglob("*"))
if path.is_file() and path.name != "manifest.json"
}
(artifacts / "manifest.json").write_text(json.dumps(manifest, indent=1))
zip_path = args.outdir / "supplement.zip"
with zipfile.ZipFile(zip_path, "w", zipfile.ZIP_DEFLATED) as archive:
for path in sorted(stage.rglob("*")):
if path.is_file():
archive.write(path, path.relative_to(args.outdir))
total = sum(p.stat().st_size for p in stage.rglob("*") if p.is_file())
print(f"files: {len(manifest) + 1}")
print(f"seed rows: {len(seeds)}")
print(f"staged bytes: {total/1e6:.1f} MB")
print(f"zip: {zip_path} ({zip_path.stat().st_size/1e6:.1f} MB)")
if __name__ == "__main__":
main()
|
