File size: 5,410 Bytes
862587b | 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 | """Known-answer tests for the prior-shift statistics, on synthetic logits (CPU, about a minute).
Data are sampled from a known "true" model; old and new are constructed so the
share of the loss gap that a context-free correction can recover is known:
pure bias new = old + global bias -> S ~ 1 for bias and aff
pure temp new = old * 2 -> S ~ 1 for aff only, T* ~ 2
context damage new = old + context-dependent -> bias shares ~ 0; aff ~ 0.3 (temperature shrinkage)
proxy data = true + bias, old = true, -> S_card inflated; S_sym bias ~ 0
new = true + context damage
"""
import argparse
import math
import torch
import prior_shift as ps
V, D, NDOC, LEN = 400, 48, 600, 64
def make(seed=0):
g = torch.Generator().manual_seed(seed)
W = torch.randn(V, D, generator=g) * 0.5
H = torch.randn(NDOC * LEN, D, generator=g)
freq = torch.randn(V, generator=g) * 1.5 # a Zipf-ish unigram prior carried in an extra column
return g, W, H, freq
def sample(H, W, extra_bias, g):
z = H @ W.T + extra_bias
return torch.multinomial(torch.softmax(z, -1), 1, generator=g).squeeze(1)
def aug(H, W, bias):
"""Fold a global bias into an extra hidden coordinate fixed at 1."""
return torch.cat([H, torch.ones(H.shape[0], 1)], 1), torch.cat([W, bias[:, None]], 1)
def run_case(H_old, W_old, H_new, W_new, y, betas=(1e3,)):
N = y.numel()
doc = torch.arange(N) // LEN
pos = torch.arange(N) % LEN + 1
nF = NDOC // 2
F = doc < nF
args = argparse.Namespace(chunk=4096, betas=list(betas), max_iter=20, beta_frac=0.1, topk=128)
allowed = torch.ones(V, dtype=torch.bool)
res = {}
for role, H, W in (("old", H_old, W_old), ("new", H_new, W_new)):
res[role] = ps.analyse(H[F], y[F], pos[F], doc[F], H[~F], y[~F], pos[~F], doc[~F] - nF,
W, allowed, nF, NDOC - nF, args)
est, ci = ps.stats_from_docsums(res["old"][0], res["new"][0], res["old"][1], B=200)
return est, ci, res
def check(cond, msg):
print(("PASS " if cond else "FAIL ") + msg)
return cond
def main():
torch.manual_seed(0)
g, W, H, freq = make()
zero = torch.zeros(V)
y = sample(H, W, freq, g)
Ho, Wo = aug(H, W, freq) # old = true model
ok = True
# 1. pure global bias shift
shift = torch.randn(V, generator=g) * 1.0
Hn, Wn = aug(H, W, freq + shift)
est, ci, res = run_case(Ho, Wo, Hn, Wn, y)
print(f"[bias] Delta={est['Delta']:.4f} S_card_bias={est['S_card_bias']:.3f} S_sym_aff={est['S_sym_aff']:.3f} S_card_cf={est['S_card_cf_beta1000']:.3f}")
ok &= check(est["Delta"] > 0.1, "bias: forgetting is created")
ok &= check(est["S_card_bias"] > 0.9 and est["S_sym_aff"] > 0.9, "bias: fitted corrections recover >90%")
ok &= check(est["S_card_cf_beta1000"] > 0.6, "bias: closed form recovers most of it")
ok &= check(res["new"][4]["jensen_in_sample_E"]["holds"], "bias: Jensen R_cf >= KL on E")
# 2. pure temperature
Hn, Wn = aug(H, W / 0.5, freq / 0.5) # T = 0.5: overconfident new model
est, ci, res = run_case(Ho, Wo, Hn, Wn, y)
print(f"[temp] Delta={est['Delta']:.4f} S_card_bias={est['S_card_bias']:.3f} S_card_aff={est['S_card_aff']:.3f} T={res['new'][4]['T_aff']:.3f}")
ok &= check(est["S_card_aff"] > 0.9, "temp: affine recovers >90%")
ok &= check(abs(res["new"][4]["T_aff"] - 2.0) < 0.1, "temp: fitted T ~ 2 undoes the doubled logits")
ok &= check(est["S_card_bias"] < est["S_card_aff"] - 0.2, "temp: bias alone recovers clearly less")
# 3. context-dependent damage only
Wd = W + torch.randn(V, D, generator=g) * 0.25
Hn, Wn = aug(H, Wd, freq)
est, ci, res = run_case(Ho, Wo, Hn, Wn, y)
print(f"[ctx] Delta={est['Delta']:.4f} S_sym_aff={est['S_sym_aff']:.3f} S_sym_bias={est['S_sym_bias']:.3f} S_sym_cf={est['S_sym_cf_beta1000']:.3f}")
ok &= check(est["Delta"] > 0.05, "ctx: forgetting is created")
ok &= check(est["S_sym_bias"] < 0.1 and est["S_sym_cf_beta1000"] < 0.1, "ctx: bias shares small")
# Random context damage inflates the logit scale, and a
# temperature recovers part of that loss (shrinkage). This is the noise floor of S_aff
# discussed in README.md.
ok &= check(0.15 < est["S_sym_aff"] < 0.45, "ctx: affine share sits at the temperature noise floor (~0.3)")
# 4. proxy mismatch: data drawn from true+bias; old = true (miscalibrated on proxy); new = ctx-damaged
y2 = sample(H, W, freq + shift * 0.7, g)
est, ci, res = run_case(Ho, Wo, Hn, Wn, y2)
print(f"[proxy] Delta={est['Delta']:.4f} S_card_aff={est['S_card_aff']:.3f} S_sym_aff={est['S_sym_aff']:.3f} S_card_bias={est['S_card_bias']:.3f} S_sym_bias={est['S_sym_bias']:.3f}")
ok &= check(est["R_aff_old"] > 0.05, "proxy: base model itself gains from the correction")
ok &= check(est["S_card_aff"] > est["S_sym_aff"] + 0.1, "proxy: card statistic is inflated relative to symmetric")
ok &= check(est["S_sym_bias"] < 0.15, "proxy: symmetric bias share stays small")
# 5. no change: Delta == 0 exactly (ratios are then undefined; summarize.py marks such a pair invalid)
est, ci, res = run_case(Ho, Wo, Ho, Wo, y)
ok &= check(abs(est["Delta"]) < 1e-9, "identity: Delta == 0 exactly")
print("ALL PASS" if ok else "SOME FAILED")
return 0 if ok else 1
if __name__ == "__main__":
raise SystemExit(main())
|