File size: 7,216 Bytes
47a719e ace30c6 47a719e ace30c6 47a719e | 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 | """ACCURACY, not likelihood: does the Δfloor rescue transfer to BLiMP?
The audit's sharpest point is that "recovery is not success" -- a likelihood rescue has not been
shown to transfer to benchmark accuracy. PolyPythia parents are English LMs, so BLiMP is directly
applicable to SET 1's merges. Scoring: sum log p over the sentence (all tokens after the first);
a paradigm item is correct when the grammatical sentence scores higher. Chance = 50%."""
import os, sys, json, time, glob, itertools, argparse, gc
sys.path.insert(0, "/root/compose-audit")
from common import *
from transformers import AutoModelForCausalLM, AutoTokenizer
import pyarrow.parquet as pq
ap = argparse.ArgumentParser()
ap.add_argument("--size", default="14m")
ap.add_argument("--seeds", default="1,2,3,4,5,6,7,8,9")
ap.add_argument("--n_per_paradigm", type=int, default=200)
ap.add_argument("--bs", type=int, default=128)
ap.add_argument("--acts_rows", type=int, default=2048)
ap.add_argument("--tag", default="blimp")
ap.add_argument("--blocks", type=int, default=48)
A = ap.parse_args()
SEEDS = [int(s) for s in A.seeds.split(",")]
OUT = f"/root/compose-audit/results/{A.tag}_{A.size}.jsonl"
DEV = "cuda"
BLIMP = glob.glob("/root/hf_cache_brainalign/hub/datasets--nyu-mll--blimp/snapshots/*/")[0]
def log(*a):
print(f"[{time.strftime('%H:%M:%S')}]", *a, flush=True)
# reuse SET 1's aligners
import importlib.util
spec = importlib.util.spec_from_file_location("s1", "/root/compose-audit/set1_polypythia.py")
def neox_head_match(sd_a, sd_b, d, nh, nl):
hd, perms = d // nh, {}
for L in range(nl):
qk = f"gpt_neox.layers.{L}.attention.query_key_value.weight"
de = f"gpt_neox.layers.{L}.attention.dense.weight"
if qk not in sd_a: continue
Aq = np.asarray(sd_a[qk], float).reshape(nh, 3 * hd, d)
Bq = np.asarray(sd_b[qk], float).reshape(nh, 3 * hd, d)
g = np.einsum("ixy,jxy->ij", Aq, Bq)
Ad = np.asarray(sd_a[de], float).reshape(d, nh, hd)
Bd = np.asarray(sd_b[de], float).reshape(d, nh, hd)
perms[L] = AL._assignment(g + np.einsum("xiy,xjy->ij", Ad, Bd))
return perms
def neox_apply_head(sd, perms, d, nh):
hd, out = d // nh, dict(sd)
for L, h in perms.items():
qk = f"gpt_neox.layers.{L}.attention.query_key_value.weight"
qb = f"gpt_neox.layers.{L}.attention.query_key_value.bias"
de = f"gpt_neox.layers.{L}.attention.dense.weight"
out[qk] = np.asarray(sd[qk], float).reshape(nh, 3 * hd, d)[h].reshape(3 * d, d)
if qb in sd: out[qb] = np.asarray(sd[qb], float).reshape(nh, 3 * hd)[h].reshape(3 * d)
out[de] = np.asarray(sd[de], float).reshape(d, nh, hd)[:, h].reshape(d, d)
return out
def align_full(sd_a, sd_b, d, aa, ab, nh, nl, method):
sd, info = AL.align_weights_full(sd_a, sd_b, d, acts_a=aa, acts_b=ab, n_heads=None,
method=method, strict=True, accept_each=True)
hp = neox_head_match(sd_a, sd, d, nh, nl)
if hp:
cand = neox_apply_head(sd, hp, d, nh)
if AL.block_normalised_distance(sd_a, cand) <= AL.block_normalised_distance(sd_a, sd):
sd = cand
return sd
# ------------------------------------------------------------------ BLiMP
def load_blimp(tok, n_per):
items = []
for d in sorted(glob.glob(BLIMP + "*/")):
name = os.path.basename(d.rstrip("/"))
f = glob.glob(d + "*.parquet")
if not f: continue
t = pq.read_table(f[0]).to_pydict()
good, bad = t["sentence_good"][:n_per], t["sentence_bad"][:n_per]
items.append((name, good, bad))
return items
def encode(tok, sents, maxlen=48):
enc = tok(sents, return_tensors="pt", padding=True, truncation=True, max_length=maxlen)
return enc["input_ids"], enc["attention_mask"]
@torch.no_grad()
def score(model, ids, am, dev, bs=128):
out = []
for i in range(0, ids.shape[0], bs):
x, m = ids[i:i + bs].to(dev), am[i:i + bs].to(dev)
lp = torch.log_softmax(model(x, attention_mask=m).logits.float()[:, :-1], -1)
tgt = x[:, 1:]
tokl = lp.gather(-1, tgt.unsqueeze(-1)).squeeze(-1) * m[:, 1:].float()
out.append(tokl.sum(1).cpu())
return torch.cat(out).numpy()
tok = AutoTokenizer.from_pretrained(f"EleutherAI/pythia-{A.size}")
if tok.pad_token is None: tok.pad_token = tok.eos_token
PARA = load_blimp(tok, A.n_per_paradigm)
log(f"BLiMP paradigms={len(PARA)} items/paradigm={len(PARA[0][1])}")
ENC = [(n, encode(tok, g), encode(tok, b)) for n, g, b in PARA]
lines = flores_lines("eng_Latn")
blocks = make_blocks(tok, lines, block=512, max_blocks=A.blocks)
shell = AutoModelForCausalLM.from_pretrained(f"EleutherAI/pythia-{A.size}-seed{SEEDS[0]}",
dtype=torch.float32).to(DEV).eval()
cfg = shell.config
D, NH, NL = cfg.hidden_size, cfg.num_attention_heads, cfg.num_hidden_layers
def blimp_acc(sd):
sd_load(shell, sd, DEV)
per, tot, cor = {}, 0, 0
for name, (gi, gm), (bi, bm) in ENC:
sg = score(shell, gi, gm, DEV, bs=A.bs)
sb = score(shell, bi, bm, DEV, bs=A.bs)
c = int((sg > sb).sum()); per[name] = c / len(sg); cor += c; tot += len(sg)
return cor / tot, per
SDS, ACTS, PACC = {}, {}, {}
for s in SEEDS:
m = AutoModelForCausalLM.from_pretrained(f"EleutherAI/pythia-{A.size}-seed{s}", dtype=torch.float32).to(DEV).eval()
SDS[s] = sd_np(m); ACTS[s] = capture_acts(m, blocks, DEV, n_rows=A.acts_rows, bs=16)
del m; torch.cuda.empty_cache()
a, _ = blimp_acc(SDS[s]); PACC[s] = a
log(f" seed{s} BLiMP={a:.4f}")
done = set()
if os.path.exists(OUT):
for l in open(OUT):
try: done.add(tuple(json.loads(l)["pair"]))
except Exception: pass
fh = open(OUT, "a")
for a, b in itertools.combinations(SEEDS, 2):
if (a, b) in done: continue
t0 = time.time()
sa, sb = SDS[a], SDS[b]
sbp = align_full(sa, sb, D, ACTS[a], ACTS[b], NH, NL, "permutation")
sbo = align_full(sa, sb, D, ACTS[a], ACTS[b], NH, NL, "orthogonal")
rungs = {"M0_naive_avg": MG.average([sa, sb]), "M1_perm_avg": MG.average([sa, sbp]),
"M1_orth_avg": MG.average([sa, sbo])}
res = {}
for k, sd in rungs.items():
acc, per = blimp_acc(sd)
res[k] = {"blimp_acc": acc, "per_paradigm": per}
r = {"set": "set1_blimp", "size": A.size, "pair": [a, b],
"metric": "BLiMP accuracy (chance=0.5) -- ACCURACY, not likelihood",
"n_per_paradigm": A.n_per_paradigm, "n_paradigms": len(ENC),
"parent_acc": {"a": PACC[a], "b": PACC[b]}, "ceiling": max(PACC[a], PACC[b]),
"rungs": {k: {"blimp_acc": v["blimp_acc"],
"delta_vs_best_parent": v["blimp_acc"] - max(PACC[a], PACC[b])} for k, v in res.items()},
"per_paradigm": {k: v["per_paradigm"] for k, v in res.items()},
"secs": time.time() - t0}
fh.write(json.dumps(r) + "\n"); fh.flush()
log(f"pair {a},{b} ceil={r['ceiling']:.4f} M0={res['M0_naive_avg']['blimp_acc']:.4f} "
f"M1p={res['M1_perm_avg']['blimp_acc']:.4f} M1o={res['M1_orth_avg']['blimp_acc']:.4f} ({r['secs']:.0f}s)")
del rungs, sbp, sbo; gc.collect()
fh.close()
log("DONE blimp", A.size)
|