summaryrefslogtreecommitdiff
path: root/experiments/oral_a_confirmation.py
blob: e8b0dba116ab20de4abff1cfdc7e4ca2a628ba6c (plain)
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
#!/usr/bin/env python3
"""Run a deterministic shard of the frozen A4 independent test panel."""
import argparse
import json
import os
import subprocess
import sys


def main():
    parser = argparse.ArgumentParser()
    parser.add_argument("--gate", default="results/oral_a_full_gate.json")
    parser.add_argument("--bp_selection", default="results/oral_a_bp_selection.json")
    parser.add_argument("--short_selection", default="results/oral_a_short_selection.json")
    parser.add_argument("--device", default="cuda")
    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()
    if not 0 <= args.shard_index < args.num_shards:
        raise ValueError("invalid shard index")
    with open(args.gate) as handle:
        gate = json.load(handle)
    with open(args.bp_selection) as handle:
        bp_selection = json.load(handle)
    with open(args.short_selection) as handle:
        short = json.load(handle)
    if gate["status"] != "passed":
        raise ValueError("A3 did not pass; A4 test confirmation is prohibited")
    if short["status"] != "selected" or not bp_selection["status"].startswith("passed_"):
        raise ValueError("development selections are incomplete")
    dfa = short["selected_dfa"]
    sdil = short["selected_sdil"]
    bp_variant = bp_selection["selected"]["variant"]
    jobs = []
    for depth in (20, 32, 56):
        for seed in range(10, 15):
            common = [
                sys.executable, "experiments/conv_run.py", "--device", args.device,
                "--depth", str(depth), "--width", "16", "--seed", str(seed),
                "--loader_seed", str(seed), "--perturb_seed", str(1000 + seed),
                "--batch_size", "128", "--epochs", "200", "--val_examples", "0",
                "--eval_split", "test", "--eval_every", "0", "--augment_train", "1",
                "--momentum", "0.9", "--normalization", "batchnorm",
            ]
            if bp_variant == "primary":
                bp_schedule = [
                    "--lr", "0.1", "--lr_schedule", "step",
                    "--lr_milestones", "100,150", "--lr_gamma", "0.1",
                    "--warmup_epochs", "0", "--weight_decay", "1e-4"]
            else:
                bp_schedule = [
                    "--lr", "0.1", "--lr_schedule", "cosine",
                    "--warmup_epochs", "5", "--weight_decay", "5e-4"]
            jobs.append((f"bp_d{depth}_s{seed}", common + [
                "--mode", "bp", *bp_schedule,
                "--out", f"results/oral_a_confirm/bp_d{depth}_s{seed}.json"]))
            local_schedule = [
                "--output_lr", "0.1", "--lr_schedule", "step",
                "--lr_milestones", "100,150", "--lr_gamma", "0.1",
                "--warmup_epochs", "0", "--weight_decay", "1e-4",
                "--alignment_probe", "32"]
            jobs.append((f"dfa_d{depth}_s{seed}", common + local_schedule + [
                "--mode", "dfa", "--lr", str(dfa["lr"]), "--a_scale", "1",
                "--vectorizer_mode", "spatial_template",
                "--out", f"results/oral_a_confirm/dfa_d{depth}_s{seed}.json"]))
            jobs.append((f"sdil_d{depth}_s{seed}", common + local_schedule + [
                "--mode", "sdil", "--lr", str(sdil["lr"]),
                "--vectorizer_mode", sdil["vectorizer_mode"],
                "--a_scale", str(sdil["a_scale"]), "--eta_A", str(sdil["eta_A"]),
                "--a_warmup_steps", "100", "--pert_sigma", "0.01",
                "--pert_directions", "1", "--pert_every", "4",
                "--out", f"results/oral_a_confirm/sdil_d{depth}_s{seed}.json"]))
    os.makedirs("results/oral_a_confirm", exist_ok=True)
    for index, (tag, command) in enumerate(jobs):
        if index % args.num_shards != args.shard_index:
            continue
        print(tag, " ".join(command), flush=True)
        if not args.dry_run:
            subprocess.run(command, check=True)


if __name__ == "__main__":
    main()