File size: 12,651 Bytes
5952424 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 | """Correctness properties for the elastic masked-diffusion noising."""
import sys
from pathlib import Path
import torch
sys.path.insert(0, str(Path(__file__).resolve().parents[1]))
from train.noising import NoiseConfig, noise_sequence, _pick_halfword_span
def _seq(n=64):
g = torch.Generator().manual_seed(0)
chars = torch.randint(0, 24, (n,), generator=g)
# random word ends every ~5 chars, last is sentence-final
boundary = torch.zeros(n, dtype=torch.uint8)
i = 4
while i < n - 1:
boundary[i] = 1
i += int(torch.randint(3, 7, (1,), generator=g).item())
boundary[-1] = 2
return chars, boundary
def test_labels_only_at_masked_positions_fixed_patterns():
cfg = NoiseConfig(w_span=1, w_word=0, w_elastic=0, w_iid=0, w_halfword=0, w_substitute=0)
g = torch.Generator().manual_seed(1)
chars, boundary = _seq()
for _ in range(50):
out = noise_sequence(chars, boundary, cfg, g)
assert not out["rebuilt"]
masked = out["input_ids"] == cfg.mask_id
supervised = out["labels"] != -100
# every supervised position is masked, and its label is the original char
assert torch.equal(masked, supervised)
assert torch.equal(out["labels"][supervised], chars[supervised])
# non-masked inputs are unchanged originals
assert torch.equal(out["input_ids"][~masked], chars[~masked])
def test_iid_rate_matches_t():
cfg = NoiseConfig(w_span=0, w_word=0, w_elastic=0, w_iid=1, w_halfword=0, w_substitute=0,
beta_a=1e6, beta_b=2e6) # t ~ 1/3 tightly
g = torch.Generator().manual_seed(2)
fracs = []
for _ in range(200):
chars, boundary = _seq(256)
out = noise_sequence(chars, boundary, cfg, g)
fracs.append((out["input_ids"] == cfg.mask_id).float().mean().item())
mean = sum(fracs) / len(fracs)
assert 0.28 < mean < 0.39, mean
def test_word_masking_respects_boundaries():
cfg = NoiseConfig(w_span=0, w_word=1, w_elastic=0, w_iid=0, w_halfword=0, w_substitute=0)
g = torch.Generator().manual_seed(3)
chars, boundary = _seq()
ends = set((boundary >= 1).nonzero(as_tuple=True)[0].tolist())
starts = {0} | {e + 1 for e in ends}
for _ in range(50):
out = noise_sequence(chars, boundary, cfg, g)
masked = (out["input_ids"] == cfg.mask_id).tolist()
# each maximal masked run must start at a word start and end at a word end
i = 0
while i < len(masked):
if masked[i]:
j = i
while j + 1 < len(masked) and masked[j + 1]:
j += 1
assert i in starts, f"masked run starts mid-word at {i}"
assert j in ends, f"masked run ends mid-word at {j}"
i = j + 1
else:
i += 1
def test_halfword_span_always_within_the_word_and_hits_all_anchors():
"""Unit-level: _pick_halfword_span must always return a sub-range of [s,e), and over many
draws should hit all three anchors (begin/end/middle) — weighted toward the end (Greek is
suffixal, so ending-restoration is the primary use case)."""
cfg = NoiseConfig()
g = torch.Generator().manual_seed(7)
s, e = 100, 112 # a 12-char synthetic word
touches_start = touches_end = interior_only = 0
for _ in range(2000):
a, b = _pick_halfword_span(s, e, cfg, g)
assert s <= a < b <= e, f"halfword span [{a},{b}) escapes word range [{s},{e})"
if a == s:
touches_start += 1
if b == e:
touches_end += 1
if a > s and b < e:
interior_only += 1
assert touches_start > 0 and touches_end > 0 and interior_only > 0
assert touches_end > touches_start # weighted toward the end (endings/inflection)
def test_halfword_pattern_only_masks_within_single_words():
"""Integration-level smoke test: halfword-only noising must not crash and every masked
position must fall inside SOME word's char range (never the inter-word separator itself,
which doesn't exist as a char anyway, but guards against off-by-one word-range bugs)."""
cfg = NoiseConfig(w_span=0, w_word=0, w_elastic=0, w_iid=0, w_halfword=1, w_substitute=0)
g = torch.Generator().manual_seed(9)
chars, boundary = _seq(256)
ends = sorted((boundary >= 1).nonzero(as_tuple=True)[0].tolist())
starts = [0] + [e + 1 for e in ends[:-1]]
in_word = torch.zeros(len(chars), dtype=torch.bool)
for ws, we in zip(starts, [e + 1 for e in ends]):
in_word[ws:we] = True
for _ in range(100):
out = noise_sequence(chars, boundary, cfg, g)
masked = out["input_ids"] == cfg.mask_id
assert bool((masked & ~in_word).any()) is False
def test_elastic_preserves_visible_chars_and_lengths():
cfg = NoiseConfig(w_span=0, w_word=0, w_elastic=1, w_iid=0, w_halfword=0, w_substitute=0)
g = torch.Generator().manual_seed(4)
chars, boundary = _seq()
for _ in range(80):
out = noise_sequence(chars, boundary, cfg, g)
assert out["rebuilt"]
inp, lab = out["input_ids"], out["labels"]
assert inp.numel() == lab.numel() == out["boundary"].numel()
assert inp.numel() >= chars.numel() # elastic only ever grows
# visible (non-mask) positions carry original chars, in order
visible = inp[inp != cfg.mask_id]
# reconstruct target: concatenation of visible + gap targets must recover the original
recon = []
for tok, l in zip(inp.tolist(), lab.tolist()):
if tok != cfg.mask_id:
recon.append(tok)
elif l != cfg.blank_id and l != -100:
recon.append(l)
assert recon == chars.tolist(), "elastic gap targets don't reconstruct the source"
# every gap has at least the true chars then >=elastic_extra_min blanks
blanks = (lab == cfg.blank_id).sum().item()
assert blanks >= cfg.elastic_extra_min
def test_substitute_never_uses_mask_token_and_label_is_true_char():
"""DENOISING pattern: corrupted positions show a WRONG letter (never MASK), and the label
is always the true original character, always different from what's shown."""
cfg = NoiseConfig(w_span=0, w_word=0, w_elastic=0, w_iid=0, w_halfword=0, w_substitute=1)
g = torch.Generator().manual_seed(11)
chars, boundary = _seq(256)
saw_any = False
for _ in range(80):
out = noise_sequence(chars, boundary, cfg, g)
assert not out["rebuilt"]
inp, lab = out["input_ids"], out["labels"]
assert not (inp == cfg.mask_id).any(), "substitute must never emit the MASK token"
corrupted = lab != -100
if corrupted.any():
saw_any = True
# every corrupted position actually differs from the true char (genuinely wrong)
assert bool((inp[corrupted] != chars[corrupted]).all())
# label is always the true original character
assert torch.equal(lab[corrupted], chars[corrupted])
# untouched positions are unchanged originals
assert torch.equal(inp[~corrupted], chars[~corrupted])
assert saw_any
def test_channel_dropout_frequencies():
"""Per-position keep masks should average out to roughly the configured marginal rates
(mixture of fully-known / fully-unknown / patchy-uniform-rate)."""
cfg = NoiseConfig()
g = torch.Generator().manual_seed(5)
chars, boundary = _seq(256)
kb_frac = kd_frac = kp_frac = 0.0
N = 400
for _ in range(N):
out = noise_sequence(chars, boundary, cfg, g)
kb_frac += out["keep_bnd_mask"].float().mean().item()
kd_frac += out["keep_dia_mask"].float().mean().item()
kp_frac += out["keep_punct_mask"].float().mean().item()
# E[rate] = p_full*1 + p_none*0 + p_patchy*0.5
exp_bnd = cfg.p_bnd_full + (1 - cfg.p_bnd_full - cfg.p_bnd_none) * 0.5
exp_dia = cfg.p_dia_full + (1 - cfg.p_dia_full - cfg.p_dia_none) * 0.5
exp_punct = cfg.p_punct_full + (1 - cfg.p_punct_full - cfg.p_punct_none) * 0.5
assert abs(kb_frac / N - exp_bnd) < 0.07, kb_frac / N
assert abs(kd_frac / N - exp_dia) < 0.07, kd_frac / N
assert abs(kp_frac / N - exp_punct) < 0.07, kp_frac / N
def test_channel_masks_are_genuinely_patchy_sometimes():
"""At least some draws must have a channel PARTIALLY known (not all-or-nothing) — this is
the actual point: some word-breaks/accents legible, others not, within one sequence."""
cfg = NoiseConfig()
g = torch.Generator().manual_seed(8)
chars, boundary = _seq(256)
patchy = 0
for _ in range(300):
out = noise_sequence(chars, boundary, cfg, g)
m = out["keep_bnd_mask"]
if 0 < m.float().mean().item() < 1:
patchy += 1
assert patchy > 0
def test_mixture_covers_all_patterns():
cfg = NoiseConfig()
g = torch.Generator().manual_seed(6)
chars, boundary = _seq()
rebuilt = fixed = 0
for _ in range(200):
out = noise_sequence(chars, boundary, cfg, g)
rebuilt += out["rebuilt"]; fixed += not out["rebuilt"]
assert rebuilt > 0 and fixed > 0
def _seq_with_lacuna(n=64, lac_start=20, lac_len=6):
"""A sequence with a real (whole-document) lacuna baked in: MASK_ID at [lac_start,
lac_start+lac_len), matching insc/data/iphi.py's text_to_full_planes() convention."""
chars, boundary = _seq(n)
MASK_ID = 24
chars = chars.clone()
chars[lac_start:lac_start + lac_len] = MASK_ID
boundary = boundary.clone()
boundary[lac_start:lac_start + lac_len] = 3 # UNK_BND
is_real_lacuna = torch.zeros(n, dtype=torch.bool)
is_real_lacuna[lac_start:lac_start + lac_len] = True
return chars, boundary, is_real_lacuna, lac_start, lac_len
def test_is_real_lacuna_none_is_fully_backward_compatible():
"""Passing is_real_lacuna=None (the default) must reproduce byte-identical output to the
pre-existing call signature, for every existing caller.
NOTE: _sample_t()'s Beta.sample() and the span pattern's Geometric.sample() don't actually
consume the passed `g` (a pre-existing bug in this file, unrelated to is_real_lacuna) --
they draw from torch's GLOBAL RNG state instead. So reproducibility across two calls
requires pinning torch.manual_seed() globally before each one, not just seeding two local
generator objects identically."""
cfg = NoiseConfig()
chars, boundary = _seq()
torch.manual_seed(42)
g1 = torch.Generator().manual_seed(42)
out1 = noise_sequence(chars, boundary, cfg, g1)
torch.manual_seed(42)
g2 = torch.Generator().manual_seed(42)
out2 = noise_sequence(chars, boundary, cfg, g2, is_real_lacuna=None)
assert torch.equal(out1["input_ids"], out2["input_ids"])
assert torch.equal(out1["labels"], out2["labels"])
def test_real_lacuna_never_selected_as_synthetic_target_fixed_patterns():
"""span/word/halfword/iid/substitute must never choose a real-lacuna position as an
ADDITIONAL synthetic-masking target -- its label must stay -100 and its input must stay
exactly MASK_ID (never substituted to a wrong letter, never re-labeled)."""
for w in ("w_span", "w_word", "w_halfword", "w_iid", "w_substitute"):
cfg = NoiseConfig(**{"w_span": 0, "w_word": 0, "w_elastic": 0, "w_iid": 0,
"w_halfword": 0, "w_substitute": 0, w: 1})
chars, boundary, is_real_lacuna, s, L = _seq_with_lacuna()
g = torch.Generator().manual_seed(0)
for trial in range(30):
out = noise_sequence(chars, boundary, cfg, g, is_real_lacuna=is_real_lacuna)
assert (out["labels"][s:s + L] == -100).all(), w
assert (out["input_ids"][s:s + L] == 24).all(), w # still exactly MASK_ID
def test_real_lacuna_excluded_from_elastic_rebuild_too():
cfg = NoiseConfig(w_span=0, w_word=0, w_elastic=1, w_iid=0, w_halfword=0, w_substitute=0)
chars, boundary, is_real_lacuna, s, L = _seq_with_lacuna()
g = torch.Generator().manual_seed(0)
for trial in range(30):
out = noise_sequence(chars, boundary, cfg, g, is_real_lacuna=is_real_lacuna)
# the lacuna's mask_id run must still be present somewhere, contiguous, with label -100
ids = out["input_ids"]; lab = out["labels"]
mask_positions = (ids == 24).nonzero(as_tuple=True)[0]
assert mask_positions.numel() >= L
# every genuinely-MASK_ID position with label -100 exists (the real lacuna survives
# copy-through unlabeled even though the sequence may have grown/shrunk elsewhere)
assert ((ids == 24) & (lab == -100)).sum() >= L
|