File size: 16,225 Bytes
744d24d 9c2a7e1 744d24d 9c2a7e1 744d24d 9c2a7e1 744d24d 9c2a7e1 744d24d 9c2a7e1 744d24d 9c2a7e1 744d24d 9c2a7e1 744d24d 9c2a7e1 744d24d 9c2a7e1 744d24d 9c2a7e1 744d24d 9c2a7e1 744d24d 9c2a7e1 744d24d 9c2a7e1 744d24d 9c2a7e1 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 | """
Flow Ensemble β Expanded Test Suite.
Assumes geolip-core is installed (Colab with repo loaded).
Tests: smoke, linalg integration, multi-scale, ensemble fusion,
gradient health, ablation, compile compatibility, memory.
"""
import torch
import torch.nn as nn
import torch.nn.functional as F
import sys, time, gc
# ββ Verify geolip_core.linalg is available ββ
try:
import geolip_core.linalg as LA
HAS_GEOLIP_LINALG = True
print(f"geolip_core.linalg: available")
LA.backend.status()
except ImportError:
import torch.linalg as LA
HAS_GEOLIP_LINALG = False
print("geolip_core.linalg: NOT available, using torch.linalg fallback")
dev = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
def sync():
if dev.type == 'cuda':
torch.cuda.synchronize()
def time_fn(fn, warmup=5, runs=50):
for _ in range(warmup): fn()
sync()
t0 = time.perf_counter()
for _ in range(runs): fn()
sync()
return (time.perf_counter() - t0) / runs * 1000
def fmt(ms):
if ms < 1: return f"{ms*1000:.0f}us"
return f"{ms:.2f}ms"
def make_data(B, n, k, d):
anchors = F.normalize(torch.randn(B, k, d, device=dev), dim=-1)
queries = F.normalize(torch.randn(B, n, d, device=dev), dim=-1)
return anchors, queries
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
print("=" * 72)
print(" Flow Ensemble β Expanded Test Suite")
print("=" * 72)
print(f" device={dev} geolip_core.linalg={HAS_GEOLIP_LINALG}")
if dev.type == 'cuda':
print(f" GPU: {torch.cuda.get_device_name()}")
print()
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# 1. SMOKE TEST β all flows, all shapes
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
print(f"{'='*72}\n 1. SMOKE TEST\n{'='*72}")
B, n, k, d = 16, 64, 32, 128
anchors, queries = make_data(B, n, k, d)
flows_cfg = [
('QuaternionFlow', lambda d,k: QuaternionFlow(d, k, n_heads=4)),
('QuaternionLiteFlow', lambda d,k: QuaternionLiteFlow(d, k)),
('VelocityFlow', lambda d,k: VelocityFlow(d, k)),
('MagnitudeFlow', lambda d,k: MagnitudeFlow(d, k)),
('OrbitalFlow', lambda d,k: OrbitalFlow(d, k)),
('AlignmentFlow', lambda d,k: AlignmentFlow(d, k)),
]
print(f"\n {'Flow':<22} {'Params':>8} {'Shape':>14} {'Time':>10} {'Conf':>8} {'Res norm':>10}")
print(f" {'β'*22} {'β'*8} {'β'*14} {'β'*10} {'β'*8} {'β'*10}")
live_flows = []
flow_ctors = []
for name, ctor in flows_cfg:
try:
flow = ctor(d, k).to(dev)
params = sum(p.numel() for p in flow.parameters())
pred, conf = flow(anchors, queries)
ms = time_fn(lambda: flow(anchors, queries))
res = (pred - queries).norm(dim=-1).mean().item()
shape_str = str(tuple(pred.shape))
print(f" {name:<22} {params:>8,} {shape_str:>14} {fmt(ms):>10} {conf.mean().item():>8.3f} {res:>10.3f}")
live_flows.append(flow)
flow_ctors.append((name, ctor))
except Exception as e:
print(f" {name:<22} FAILED: {str(e)[:50]}")
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# 2. LINALG INTEGRATION
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
print(f"\n{'='*72}\n 2. LINALG INTEGRATION\n{'='*72}")
if HAS_GEOLIP_LINALG:
print(f"\n Testing eigh dispatch in MagnitudeFlow and OrbitalFlow...")
for FlowCls in [MagnitudeFlow, OrbitalFlow]:
flow = FlowCls(d, k).to(dev)
pred, conf = flow(anchors, queries)
ok = torch.isfinite(pred).all().item() and torch.isfinite(conf).all().item()
print(f" {flow.name:<18} finite={ok} conf={conf.mean():.3f}")
oflow = OrbitalFlow(d, k).to(dev)
a_geom = oflow.anchor_proj(anchors)
G = torch.bmm(a_geom.transpose(-2, -1), a_geom)
vals, vecs = LA.eigh(G)
print(f"\n Gram eigenspectrum: shape={tuple(vals.shape)} "
f"range=[{vals.min().item():.4f}, {vals.max().item():.4f}]")
print(f" Eigenvector orth err: {(torch.bmm(vecs.mT, vecs) - torch.eye(oflow.geom_dim, device=dev)).abs().max().item():.2e}")
else:
print(" Skipped β geolip_core.linalg not available")
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# 3. MULTI-SCALE
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
print(f"\n{'='*72}\n 3. MULTI-SCALE\n{'='*72}")
configs = [
(4, 16, 8, 64, 'tiny'),
(16, 64, 32, 128, 'small'),
(32, 128, 64, 256, 'medium'),
(64, 256, 128, 256, 'large'),
(8, 512, 256, 512, 'wide'),
]
print(f"\n OrbitalFlow across scales:")
print(f" {'Config':<10} {'B':>4} {'n':>5} {'k':>5} {'d':>5} {'Time':>10} {'OK':>4}")
print(f" {'β'*10} {'β'*4} {'β'*5} {'β'*5} {'β'*5} {'β'*10} {'β'*4}")
for B_, n_, k_, d_, label in configs:
try:
of = OrbitalFlow(d_, k_).to(dev)
a, q = make_data(B_, n_, k_, d_)
pred, conf = of(a, q)
ms = time_fn(lambda: of(a, q), warmup=3, runs=20)
ok = torch.isfinite(pred).all().item()
print(f" {label:<10} {B_:>4} {n_:>5} {k_:>5} {d_:>5} {fmt(ms):>10} {'OK' if ok else 'NO':>4}")
del of, a, q
except Exception as e:
print(f" {label:<10} {B_:>4} {n_:>5} {k_:>5} {d_:>5} FAILED: {str(e)[:30]}")
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# 4. ENSEMBLE FUSION MODES
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
print(f"\n{'='*72}\n 4. ENSEMBLE FUSION\n{'='*72}")
B, n, k, d = 16, 64, 32, 128
anchors, queries = make_data(B, n, k, d)
for fusion in ['weighted', 'gated', 'residual']:
ens = FlowEnsemble(live_flows, d, fusion=fusion).to(dev)
out = ens(anchors, queries)
ms = time_fn(lambda: ens(anchors, queries), warmup=3, runs=20)
preds = [flow(anchors, queries)[0] for flow in ens.flows]
cos_sims = []
for i in range(len(preds)):
for j in range(i+1, len(preds)):
cs = F.cosine_similarity(preds[i].flatten(1), preds[j].flatten(1), dim=-1).mean().item()
cos_sims.append(cs)
avg_sim = sum(cos_sims) / max(len(cos_sims), 1)
print(f"\n {fusion}: time={fmt(ms)} norm={out.norm(dim=-1).mean():.3f} diversity={1-avg_sim:.3f}")
diag = ens.flow_diagnostics(anchors, queries)
for fname, stats in diag.items():
print(f" {fname:<18} conf={stats['confidence_mean']:.3f}Β±{stats['confidence_std']:.3f} "
f"res={stats['residual_norm']:.3f}")
del ens
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# 5. GRADIENT HEALTH
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
print(f"\n{'='*72}\n 5. GRADIENT HEALTH\n{'='*72}")
B, n, k, d = 16, 64, 32, 128
anchors, queries = make_data(B, n, k, d)
losses = {
'mse': (lambda o,q: (o - q).pow(2).mean()),
'cosine': (lambda o,q: (1 - F.cosine_similarity(o, q, dim=-1)).mean()),
'norm': (lambda o,q: o.norm(dim=-1).mean()),
}
print(f"\n {'Flow':<18} {'Loss':<10} {'Grad norm':>12} {'Status':>8}")
print(f" {'β'*18} {'β'*10} {'β'*12} {'β'*8}")
for loss_name, loss_fn in losses.items():
# Fresh flows for each loss β avoids in-place grad corruption across losses
try:
test_flows_grad = [ctor(d, k).to(dev) for _, ctor in flow_ctors]
ens_g = FlowEnsemble(test_flows_grad, d, fusion='residual').to(dev)
ens_g.zero_grad()
anchors_g = anchors.detach().clone().requires_grad_(True)
queries_g = queries.detach().clone().requires_grad_(True)
out = ens_g(anchors_g, queries_g)
loss = loss_fn(out, queries_g.detach())
loss.backward()
for flow in ens_g.flows:
grads = [p.grad for p in flow.parameters() if p.grad is not None]
if grads:
gn = torch.cat([g.flatten() for g in grads]).norm().item()
status = "OK" if 1e-8 < gn < 1e4 else "WARN"
print(f" {flow.name:<18} {loss_name:<10} {gn:>12.2e} {status:>8}")
else:
print(f" {flow.name:<18} {loss_name:<10} {'no grads':>12} {'WARN':>8}")
del ens_g, test_flows_grad
except RuntimeError as e:
if 'inplace' in str(e).lower() or 'in-place' in str(e).lower() or 'modified by' in str(e):
print(f" {'*':>18} {loss_name:<10} {'IN-PLACE ERR':>12} {'NOTE':>8}")
print(f" FL eigh deflation uses indexed assignment β needs .clone() fix")
else:
print(f" {'*':>18} {loss_name:<10} {'ERROR':>12}")
print(f" {str(e)[:60]}")
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# 6. ABLATION β solo vs pairs vs full ensemble
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
print(f"\n{'='*72}\n 6. ABLATION (100 training steps, rotation target)\n{'='*72}")
B, n, k, d = 32, 128, 64, 256
anchors, queries = make_data(B, n, k, d)
R = torch.linalg.qr(torch.randn(d, d, device=dev)).Q.unsqueeze(0)
target = torch.bmm(queries, R.expand(B, -1, -1))
def eval_quality(model, anchors, queries, target, steps=100, lr=1e-3):
opt = torch.optim.Adam(model.parameters(), lr=lr)
for _ in range(steps):
opt.zero_grad()
pred = model(anchors, queries) if isinstance(model, FlowEnsemble) else model(anchors, queries)[0]
loss = (pred - target).pow(2).mean()
loss.backward()
opt.step()
with torch.no_grad():
pred = model(anchors, queries) if isinstance(model, FlowEnsemble) else model(anchors, queries)[0]
return (pred - target).pow(2).mean().item()
print(f"\n {'Configuration':<35} {'MSE':>10} {'Params':>10}")
print(f" {'β'*35} {'β'*10} {'β'*10}")
for name, ctor in flow_ctors:
try:
flow = ctor(d, k).to(dev)
params = sum(p.numel() for p in flow.parameters())
mse = eval_quality(flow, anchors, queries, target)
print(f" {name:<35} {mse:>10.4f} {params:>10,}")
del flow
except Exception as e:
print(f" {name:<35} FAILED: {str(e)[:30]}")
pairs = [
('Quat + Orbital', [0, 4]),
('Velocity + Magnitude', [2, 3]),
('Orbital + Alignment', [4, 5]),
('Velocity + Orbital', [2, 4]),
]
for pair_name, indices in pairs:
try:
pair_flows = [flow_ctors[i][1](d, k).to(dev) for i in indices if i < len(flow_ctors)]
if len(pair_flows) >= 2:
ens = FlowEnsemble(pair_flows, d, fusion='weighted').to(dev)
params = sum(p.numel() for p in ens.parameters())
mse = eval_quality(ens, anchors, queries, target)
print(f" {pair_name:<35} {mse:>10.4f} {params:>10,}")
del ens, pair_flows
except Exception as e:
print(f" {pair_name:<35} FAILED: {str(e)[:30]}")
for fusion in ['weighted', 'residual']:
try:
all_flows = [ctor(d, k).to(dev) for _, ctor in flow_ctors]
ens = FlowEnsemble(all_flows, d, fusion=fusion).to(dev)
params = sum(p.numel() for p in ens.parameters())
mse = eval_quality(ens, anchors, queries, target)
print(f" {'Full (' + fusion + ')':<35} {mse:>10.4f} {params:>10,}")
del ens, all_flows
except Exception as e:
print(f" {'Full (' + fusion + ')':<35} FAILED: {str(e)[:30]}")
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# 7. COMPILE COMPATIBILITY
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
print(f"\n{'='*72}\n 7. COMPILE COMPATIBILITY\n{'='*72}")
B, n, k, d = 8, 32, 16, 64
anchors, queries = make_data(B, n, k, d)
print(f"\n {'Flow':<22} {'fullgraph':>12} {'Raw':>10} {'Compiled':>12}")
print(f" {'β'*22} {'β'*12} {'β'*10} {'β'*12}")
for name, ctor in flow_ctors:
try:
flow = ctor(d, k).to(dev)
t_raw = time_fn(lambda: flow(anchors, queries), warmup=3, runs=30)
try:
compiled = torch.compile(flow, fullgraph=True)
compiled(anchors, queries); sync()
t_comp = time_fn(lambda: compiled(anchors, queries), warmup=3, runs=30)
status = "OK"
except Exception as e:
t_comp = -1
status = str(e)[:12]
t_str = fmt(t_comp) if t_comp > 0 else "N/A"
print(f" {name:<22} {status:>12} {fmt(t_raw):>10} {t_str:>12}")
del flow
except Exception as e:
print(f" {name:<22} FAILED: {str(e)[:40]}")
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# 8. MEMORY
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
if dev.type == 'cuda':
print(f"\n{'='*72}\n 8. MEMORY (B=32, n=128, k=64, d=256)\n{'='*72}")
B, n, k, d = 32, 128, 64, 256
anchors, queries = make_data(B, n, k, d)
print(f"\n {'Flow':<22} {'Peak MB':>10}")
print(f" {'β'*22} {'β'*10}")
for name, ctor in flow_ctors:
try:
flow = ctor(d, k).to(dev)
torch.cuda.empty_cache(); gc.collect()
torch.cuda.reset_peak_memory_stats()
base = torch.cuda.memory_allocated()
pred, conf = flow(anchors, queries); sync()
peak = (torch.cuda.max_memory_allocated() - base) / 1024**2
print(f" {name:<22} {peak:>9.1f}")
del flow, pred, conf
except Exception as e:
print(f" {name:<22} FAILED: {str(e)[:30]}")
try:
all_flows = [ctor(d, k).to(dev) for _, ctor in flow_ctors]
ens = FlowEnsemble(all_flows, d, fusion='weighted').to(dev)
torch.cuda.empty_cache(); gc.collect()
torch.cuda.reset_peak_memory_stats()
base = torch.cuda.memory_allocated()
out = ens(anchors, queries); sync()
peak = (torch.cuda.max_memory_allocated() - base) / 1024**2
print(f" {'Full ensemble':<22} {peak:>9.1f}")
del ens, all_flows
except Exception as e:
print(f" {'Full ensemble':<22} FAILED: {str(e)[:30]}")
print(f"\n{'='*72}")
print(f" Done.")
print(f"{'='*72}") |