File size: 4,191 Bytes
c399dd1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""Evaluate a policy pool on a goal set under nominal ROM and TR variants (Sim2Sim, ROM side).

Usage: python run_rom.py --set A --pool explore --variants all
Writes results/rom_<set>/<variant>.npz (success, time_s, terminal/min distance [policy, goal, trial]; sampled
parameters [goal, trial]) and results/rom_<set>/<variant>.json (manifest: spec, K, seeds, policies, hashes).
Common random numbers: the same sampled parameters and noise stream for every policy.
"""
import argparse
import json
import time
import zlib

import numpy as np
import torch

import rom_batch
import variants as V
from common import RESULTS, Actor, make_goal_set, policy_registry, rel_path, sha256

POOLS = {
    'explore': ['rel1', 'rel2', 'rel3', 'rel4'] + [f'{c}_s{s}' for c in ('nonoise', 'release', 'robust') for s in (1001, 1002, 1003, 1004)],
    'confirm': [f'{c}_s{s}' for c in ('nonoise', 'release', 'robust') for s in (2001, 2002, 2003, 2004)],
    'released': ['rel1', 'rel2', 'rel3', 'rel4'],
}
POOLS['confirm16'] = POOLS['confirm'] + POOLS['released']   # set B: primary = confirm (12 held-out), secondary = all 16
N_PER_RANGE = {'A': 50, 'B': 75}


def seed_of(*parts):
    return zlib.crc32(':'.join(map(str, parts)).encode())


def run_variant(name, K, spec, goals, policies, reg, out_dir, set_name, device):
    G = len(goals)
    rng = np.random.default_rng(seed_of(set_name, name, 'params'))
    params = V.sample(spec, G * K, rng)
    noise_seed = seed_of(set_name, name, 'noise')
    gxy = np.repeat(np.array([[g['x'], g['y']] for g in goals]), K, axis=0)
    shape = (len(policies), G, K)
    res = {k: np.zeros(shape, dtype=np.float32) for k in ('time_s', 'terminal_distance_m', 'min_distance_m')}
    res['success'] = np.zeros(shape, dtype=bool)
    t0 = time.time()
    for i, p in enumerate(policies):
        actor = Actor(reg[p]['path']).to_device(device)
        o = rom_batch.simulate(actor, gxy, params, noise_seed=noise_seed, device=device)
        for k in res:
            res[k][i] = o[k].reshape(G, K)
    np.savez_compressed(out_dir / f'{name}.npz', policies=np.array(policies), goal_ids=np.array([g['goal_id'] for g in goals]),
                        **res, **{f'param_{k}': v.reshape(G, K).astype(np.float32) for k, v in params.items()})
    man = dict(variant=name, trials_per_goal=K, spec={k: list(v) for k, v in spec.items()}, goal_set=set_name, goals=G,
               param_seed=seed_of(set_name, name, 'params'), noise_seed=noise_seed, device=device,
               policies={p: dict(path=rel_path(reg[p]['path']), sha256=sha256(reg[p]['path'])) for p in policies},
               wall_s=time.time() - t0)
    (out_dir / f'{name}.json').write_text(json.dumps(man, indent=1) + '\n')
    return res


def main():
    ap = argparse.ArgumentParser()
    ap.add_argument('--set', choices=['A', 'B'], required=True)
    ap.add_argument('--pool', choices=list(POOLS), required=True)
    ap.add_argument('--variants', default='all', help='comma list or all (EXPLORE table, or the fixed design with --table fixed)')
    ap.add_argument('--device', default='cpu')
    ap.add_argument('--tag', default='')
    ap.add_argument('--table', choices=['explore', 'fixed'], default='explore')
    a = ap.parse_args()
    torch.set_num_threads(2)
    if a.set == 'B' and a.table != 'fixed':
        raise SystemExit('goal set B is the held-out set: use --table fixed')
    if a.table == 'fixed':
        import fixed_design
        table = fixed_design.CONDITIONS
    else:
        table = V.EXPLORE
    names = list(table) if a.variants == 'all' else a.variants.split(',')
    goals = make_goal_set(a.set, N_PER_RANGE[a.set])
    reg = policy_registry()
    policies = POOLS[a.pool]
    out_dir = RESULTS / f'rom_{a.set}_{a.pool}{a.tag}'
    out_dir.mkdir(parents=True, exist_ok=True)
    (out_dir / 'goals.json').write_text(json.dumps(goals, indent=1) + '\n')
    for n in names:
        K, spec = table[n]
        r = run_variant(n, K, spec, goals, policies, reg, out_dir, a.set, a.device)
        sr = r['success'].mean((1, 2))
        print(f'{n:18s} K={K:2d} success ' + ' '.join(f'{p}:{s:.2f}' for p, s in zip(policies, sr)), flush=True)


if __name__ == '__main__':
    main()