File size: 11,770 Bytes
fc29ddb | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 | """
MHA CV Relational Test β Prototype
Train a minimal embedding + MHA + classifier on 10 noise patterns.
Measure CV on embedding weights, Q/K/V projections, and attention output
across different head counts per embedding dimension.
Hypothesis: head_dim (D / n_heads) determines CV of internal representations,
and the band-valid head_dims produce qualitatively different geometric behavior.
"""
import torch
import torch.nn as nn
import torch.nn.functional as F
import math
# ββ CM primitives ββ
def cayley_menger_vol2(points):
B, N, D = points.shape
gram = torch.bmm(points, points.transpose(1, 2))
norms = torch.diagonal(gram, dim1=1, dim2=2)
d2 = F.relu(norms.unsqueeze(2) + norms.unsqueeze(1) - 2 * gram)
cm = torch.zeros(B, N + 1, N + 1, device=points.device, dtype=points.dtype)
cm[:, 0, 1:] = 1.0
cm[:, 1:, 0] = 1.0
cm[:, 1:, 1:] = d2
k = N - 1
sign = (-1.0) ** (k + 1)
fact = math.factorial(k)
return sign * torch.linalg.det(cm.float()).to(points.dtype) / ((2 ** k) * (fact ** 2))
def cv_metric(weight, n_samples=300):
"""CV of pentachoron volumes. weight: (N, D)"""
V, D = weight.shape
if V < 5:
return None
pool = min(V, 512)
indices = torch.stack([
torch.randperm(pool, device=weight.device)[:5]
for _ in range(n_samples)
])
pts = weight[:pool][indices]
vol2 = cayley_menger_vol2(pts)
valid = vol2 > 1e-20
if valid.sum() < 10:
return None
vols = vol2[valid].sqrt()
return (vols.std() / (vols.mean() + 1e-8)).item()
# ββ Minimal model ββ
class MHAClassifier(nn.Module):
def __init__(self, vocab, dim, n_heads, seq_len, n_classes):
super().__init__()
self.emb = nn.Embedding(vocab, dim)
self.pos = nn.Parameter(torch.randn(1, seq_len, dim) * 0.02)
self.mha = nn.MultiheadAttention(dim, n_heads, batch_first=True)
self.norm = nn.LayerNorm(dim)
self.head = nn.Linear(dim, n_classes)
def forward(self, x):
# x: (B, seq_len) token indices
h = self.emb(x) + self.pos
attn_out, _ = self.mha(h, h, h)
h = self.norm(h + attn_out)
# pool over sequence
h = h.mean(dim=1)
return self.head(h)
@torch.no_grad()
def forward_activations(self, x, n_heads):
"""Forward pass returning per-head Q/K/V activations and post-attn output.
Returns dict of (B*seq, head_dim) tensors for CV measurement.
"""
h = self.emb(x) + self.pos # (B, seq, D)
B, S, D = h.shape
head_dim = D // n_heads
# Manually compute Q, K, V from in_proj
w = self.mha.in_proj_weight
b = self.mha.in_proj_bias
qkv = F.linear(h, w, b) # (B, seq, 3*D)
q, k, v = qkv.chunk(3, dim=-1) # each (B, seq, D)
# Reshape to per-head: (B, seq, n_heads, head_dim)
q = q.view(B, S, n_heads, head_dim)
k = k.view(B, S, n_heads, head_dim)
v = v.view(B, S, n_heads, head_dim)
# Compute attention output
attn_out, _ = self.mha(h, h, h)
post_attn = self.norm(h + attn_out) # (B, seq, D)
# Post-attn per head view
post_heads = post_attn.view(B, S, n_heads, head_dim)
acts = {}
for i in range(n_heads):
acts[f"act_Q_h{i}"] = q[:, :, i, :].reshape(-1, head_dim)
acts[f"act_K_h{i}"] = k[:, :, i, :].reshape(-1, head_dim)
acts[f"act_V_h{i}"] = v[:, :, i, :].reshape(-1, head_dim)
acts[f"act_post_h{i}"] = post_heads[:, :, i, :].reshape(-1, head_dim)
# Also full-dim activations
acts["act_emb"] = h.reshape(-1, D)
acts["act_post_full"] = post_attn.reshape(-1, D)
return acts
def get_qkv_weights(self):
"""Extract Q, K, V projection weight matrices."""
# nn.MultiheadAttention packs Q, K, V into in_proj_weight: (3*dim, dim)
w = self.mha.in_proj_weight.detach()
d = w.shape[1]
q_w = w[:d] # (dim, dim)
k_w = w[d:2*d] # (dim, dim)
v_w = w[2*d:] # (dim, dim)
return q_w, k_w, v_w
def get_per_head_projections(self, n_heads):
"""Split Q/K/V weights into per-head chunks. Returns list of (head_dim, dim) per head."""
q_w, k_w, v_w = self.get_qkv_weights()
d = q_w.shape[0]
head_dim = d // n_heads
q_heads = [q_w[i*head_dim:(i+1)*head_dim] for i in range(n_heads)]
k_heads = [k_w[i*head_dim:(i+1)*head_dim] for i in range(n_heads)]
v_heads = [v_w[i*head_dim:(i+1)*head_dim] for i in range(n_heads)]
return q_heads, k_heads, v_heads
# ββ Data: 10 noise patterns with perturbations ββ
def make_data(n_classes=10, samples_per_class=50, seq_len=8, vocab=256):
"""Create simple classification data. Each class has a base token pattern with noise."""
torch.manual_seed(42)
# Base patterns: each class gets a fixed token sequence
base_patterns = torch.randint(0, vocab, (n_classes, seq_len))
all_x, all_y = [], []
for cls in range(n_classes):
for _ in range(samples_per_class):
pattern = base_patterns[cls].clone()
# Perturb ~25% of positions
mask = torch.rand(seq_len) < 0.25
pattern[mask] = torch.randint(0, vocab, (mask.sum(),))
all_x.append(pattern)
all_y.append(cls)
x = torch.stack(all_x)
y = torch.tensor(all_y)
perm = torch.randperm(len(x))
return x[perm], y[perm]
# ββ CV measurement suite ββ
def measure_all_cv(model, n_heads, x=None):
"""Measure CV on all relevant weight matrices and activations."""
results = {}
# Embedding weights
emb_w = model.emb.weight.detach()
results["emb"] = cv_metric(emb_w)
# Full Q, K, V projection matrices (dim Γ dim)
q_w, k_w, v_w = model.get_qkv_weights()
results["Q_full"] = cv_metric(q_w)
results["K_full"] = cv_metric(k_w)
results["V_full"] = cv_metric(v_w)
# Per-head projections (head_dim Γ dim) β CV measured on head_dim rows
q_heads, k_heads, v_heads = model.get_per_head_projections(n_heads)
for i in range(n_heads):
results[f"Q_h{i}"] = cv_metric(q_heads[i])
results[f"K_h{i}"] = cv_metric(k_heads[i])
results[f"V_h{i}"] = cv_metric(v_heads[i])
# Output projection
out_w = model.mha.out_proj.weight.detach()
results["out_proj"] = cv_metric(out_w)
# Classifier head
head_w = model.head.weight.detach()
results["cls_head"] = cv_metric(head_w)
# Activations β the space where attention actually operates
if x is not None:
model.eval()
acts = model.forward_activations(x, n_heads)
for name, tensor in acts.items():
results[name] = cv_metric(tensor)
return results
def fmt_cv(cv):
if cv is None:
return " N/A "
band = "*" if 0.13 < cv < 0.30 else " "
return f"{band}{cv:.4f}{band}"
# ββ Training + measurement loop ββ
def run_experiment(dim, n_heads, vocab=256, seq_len=8, n_classes=10, epochs=50, lr=1e-3):
head_dim = dim // n_heads
print(f"\n{'='*70}")
print(f"D={dim} heads={n_heads} head_dim={head_dim}")
print(f"{'='*70}")
x, y = make_data(n_classes=n_classes, seq_len=seq_len, vocab=vocab)
model = MHAClassifier(vocab, dim, n_heads, seq_len, n_classes)
opt = torch.optim.Adam(model.parameters(), lr=lr)
# Pre-training CV
print(f"\n [pre-train]")
pre_cv = measure_all_cv(model, n_heads, x)
for k, v in pre_cv.items():
print(f" {k:16s}: {fmt_cv(v)}")
# Training
mid_cv = None
for epoch in range(1, epochs + 1):
model.train()
opt.zero_grad()
logits = model(x)
loss = F.cross_entropy(logits, y)
loss.backward()
opt.step()
if epoch == epochs // 2:
model.eval()
with torch.no_grad():
acc = (model(x).argmax(-1) == y).float().mean().item()
mid_cv = measure_all_cv(model, n_heads, x)
print(f"\n [epoch {epoch}] loss={loss.item():.4f} acc={acc:.2%}")
for k, v in mid_cv.items():
print(f" {k:16s}: {fmt_cv(v)}")
# Post-training CV
model.eval()
with torch.no_grad():
acc = (model(x).argmax(-1) == y).float().mean().item()
print(f"\n [post-train] loss={loss.item():.4f} acc={acc:.2%}")
post_cv = measure_all_cv(model, n_heads, x)
for k, v in post_cv.items():
pre = pre_cv.get(k)
delta = ""
if v is not None and pre is not None:
d = v - pre
delta = f" Ξ={d:+.4f}"
print(f" {k:16s}: {fmt_cv(v)}{delta}")
return {
"dim": dim, "n_heads": n_heads, "head_dim": head_dim,
"pre": pre_cv, "mid": mid_cv, "post": post_cv, "acc": acc,
}
# ββ Main ββ
if __name__ == "__main__":
print("MHA CV Relational Test β Prototype")
print("Band: 0.13 < CV < 0.30")
configs = [
# D=64: head_dims 64, 32, 16, 8
(64, 1),
(64, 2),
(64, 4),
(64, 8),
# D=128: head_dims 128, 64, 32, 16
(128, 1),
(128, 2),
(128, 4),
(128, 8),
# D=256: head_dims 256, 128, 64, 32
(256, 1),
(256, 2),
(256, 4),
(256, 8),
]
all_results = []
for dim, n_heads in configs:
r = run_experiment(dim, n_heads)
all_results.append(r)
# Summary β Weights
print(f"\n\n{'='*70}")
print("SUMMARY: Post-training WEIGHT CV by head_dim")
print(f"{'='*70}")
print(f"{'D':>5} {'heads':>5} {'hdim':>5} | {'emb':>8} {'Q_full':>8} {'K_full':>8} {'V_full':>8} {'out':>8} | acc")
print("-" * 80)
for r in all_results:
p = r["post"]
print(f"{r['dim']:5d} {r['n_heads']:5d} {r['head_dim']:5d} | "
f"{fmt_cv(p.get('emb')):>8} {fmt_cv(p.get('Q_full')):>8} "
f"{fmt_cv(p.get('K_full')):>8} {fmt_cv(p.get('V_full')):>8} "
f"{fmt_cv(p.get('out_proj')):>8} | {r['acc']:.2%}")
# Summary β Activations (the real test)
print(f"\n\n{'='*70}")
print("SUMMARY: Post-training ACTIVATION CV by head_dim")
print("(These measure the space where attention actually operates)")
print(f"{'='*70}")
print(f"{'D':>5} {'heads':>5} {'hdim':>5} | {'act_emb':>8} {'aQ_h0':>8} {'aK_h0':>8} {'aV_h0':>8} {'aPost0':>8} {'act_full':>8} | acc")
print("-" * 90)
for r in all_results:
p = r["post"]
print(f"{r['dim']:5d} {r['n_heads']:5d} {r['head_dim']:5d} | "
f"{fmt_cv(p.get('act_emb')):>8} "
f"{fmt_cv(p.get('act_Q_h0')):>8} {fmt_cv(p.get('act_K_h0')):>8} "
f"{fmt_cv(p.get('act_V_h0')):>8} {fmt_cv(p.get('act_post_h0')):>8} "
f"{fmt_cv(p.get('act_post_full')):>8} | {r['acc']:.2%}")
# Summary β Activation CV delta (preβpost)
print(f"\n\n{'='*70}")
print("SUMMARY: ACTIVATION CV movement (post - pre)")
print(f"{'='*70}")
print(f"{'D':>5} {'heads':>5} {'hdim':>5} | {'act_emb':>8} {'aQ_h0':>8} {'aK_h0':>8} {'aV_h0':>8} {'aPost0':>8} {'act_full':>8}")
print("-" * 80)
for r in all_results:
pre, post = r["pre"], r["post"]
def delta(k):
a, b = pre.get(k), post.get(k)
if a is not None and b is not None:
d = b - a
return f"{d:+.4f}"
return " N/A "
print(f"{r['dim']:5d} {r['n_heads']:5d} {r['head_dim']:5d} | "
f"{delta('act_emb'):>8} "
f"{delta('act_Q_h0'):>8} {delta('act_K_h0'):>8} "
f"{delta('act_V_h0'):>8} {delta('act_post_h0'):>8} "
f"{delta('act_post_full'):>8}") |