File size: 8,642 Bytes
57dbe46
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
#!/usr/bin/env python3
"""
NQR-SNN Optimization Test Suite
================================
Tests each proposed optimization against the baseline and reports
accuracy + inference speed. Run locally:

    python optimization_test.py

Expected runtime: ~10-15 min on a modern CPU.
Results saved to outputs/results/optimization_results.json
"""
import os, sys, time, json, types
import numpy as np
import torch
sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
from nqr_snn import config
from nqr_snn.data.generator import generate_dataset_v2, generate_signal_at_snr, generate_noise_only_at_power, generate_dataset
from nqr_snn.data.dataset import NQRDatasetV2, get_balanced_loader_v2, extract_features_batch
from nqr_snn.snn.model import SpikingClassifier, SPIKINGJELLY_AVAILABLE
from nqr_snn.snn.encoder import DeterministicEncoder
from nqr_snn.snn.ensemble import SNNEnsemble
from nqr_snn.evaluation.metrics import full_report

TRAIN_SIZE, VAL_SIZE, TEST_PER_CLASS, MAX_EPOCHS = 500, 150, 150, 50
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
print(f"SpikingJelly={SPIKINGJELLY_AVAILABLE} Device={DEVICE}")

def make_test(snr, n=TEST_PER_CLASS, seed=99):
    rng = np.random.RandomState(seed)
    s, l = [], []
    for _ in range(n):
        noisy, _ = generate_signal_at_snr(snr, "white", rng); s.append(noisy); l.append(1)
    for _ in range(n):
        s.append(generate_noise_only_at_power("white", rng, 1.0)); l.append(0)
    return np.array(s, dtype=np.complex128), np.array(l)

def evaluate(ens, enc, tag=""):
    res, tt = {}, 0
    for snr in config.EVAL_SNR_LEVELS:
        s, l = make_test(snr, seed=abs(snr)*100)
        f = torch.from_numpy(extract_features_batch(s))
        t0 = time.time()
        with torch.no_grad():
            x = enc.encode(f).to(DEVICE) if isinstance(enc, torch.nn.Module) else enc.encode(f).to(DEVICE)
        mp, _ = ens.predict(x); tt += time.time()-t0
        r = full_report(l, mp.cpu().numpy()); res[snr] = r
        print(f"  {tag} SNR={snr:4d} | acc={r['accuracy']:.4f} auc={r['auc']:.4f}")
    a = [v['accuracy'] for v in res.values()]
    ma, mi = np.mean(a), np.min(a)
    print(f"  {tag} MEAN={ma:.4f} MIN={mi:.4f} T={tt:.3f}s")
    return {"mean_acc": float(ma), "min_acc": float(mi), "inference_time": float(tt)}

def train_ens(sz, tl, vl, mdir):
    od = config.MODELS_DIR; config.MODELS_DIR = mdir; os.makedirs(mdir, exist_ok=True)
    e = SNNEnsemble(ensemble_size=sz, device=DEVICE, heterogeneous=True)
    t0 = time.time(); e.train_all(tl, vl, max_epochs=MAX_EPOCHS); tt = time.time()-t0
    e.load_checkpoints(); config.MODELS_DIR = od
    return e, tt

# Data
print("\n" + "="*70 + "\nGENERATING DATA\n" + "="*70)
tr = generate_dataset_v2(TRAIN_SIZE, config.TRAIN_SNR_RANGE, "white", seed=42)
vl = generate_dataset_v2(VAL_SIZE, config.VAL_SNR_RANGE, "white", seed=100)
tr_ds = NQRDatasetV2(tr["signals"], tr["labels"])
vl_ds = NQRDatasetV2(vl["signals"], vl["labels"])
tl64 = get_balanced_loader_v2(tr_ds, batch_size=64)
vl64 = get_balanced_loader_v2(vl_ds, batch_size=64, shuffle=False)
tl128 = get_balanced_loader_v2(tr_ds, batch_size=128)
vl128 = get_balanced_loader_v2(vl_ds, batch_size=128, shuffle=False)
R = {}

# BASELINE
print("\n" + "="*70 + "\nBASELINE: 3-member, bs64\n" + "="*70)
eb, ttb = train_ens(3, tl64, vl64, "outputs/m_base")
enc = DeterministicEncoder()
R["baseline"] = {**evaluate(eb, enc, "[BASE]"), "train_time": ttb}

# OPT1: CNN ONCE
print("\n" + "="*70 + "\nOPT1: CNN SINGLE-PASS\n" + "="*70)
if SPIKINGJELLY_AVAILABLE:
    from spikingjelly.activation_based import functional
    def fwd1(self, x_seq, return_per_timestep=False):
        if not self.use_spikingjelly: return self.snn_head(self.cnn(x_seq[0]).unsqueeze(0))
        T = x_seq.shape[0]
        fs = self.cnn(x_seq[0])
        for sl in self._snn_layers: functional.reset_net(sl)
        with torch.no_grad():
            e = x_seq.pow(2).sum(dim=(-1,-2)); r = (e/(e[0:1]+1e-8)).unsqueeze(-1)
        fm = fs.unsqueeze(0).expand(T,-1,-1) * r
        fm = self.snn_norm(fm); x = self.fc1(fm); x = self.neuron1(x); x = self.drop1(x)
        x = self.fc2(x); x = self.neuron2(x); x = self.drop2(x); x = self.fc_out(x)
        if return_per_timestep: return x
        return torch.sigmoid(x.mean(dim=0))
    for m in eb.models: m.forward = types.MethodType(fwd1, m)
    R["cnn_once"] = evaluate(eb, enc, "[CNN1x]")
    from nqr_snn.snn.model import SpikingClassifier as SC
    for m in eb.models: m.forward = types.MethodType(SC.forward, m)
else:
    R["cnn_once"] = R["baseline"]

# OPT3: FUSED FEATURES
print("\n" + "="*70 + "\nOPT3: FUSED FEATURES (speed only)\n" + "="*70)
def fused(signals):
    N, L = signals.shape
    r = signals.real.astype(np.float32); i = signals.imag.astype(np.float32)
    m = np.abs(signals).astype(np.float32)
    fv = np.fft.fft(signals, axis=1)
    lf = np.log1p(np.abs(fv)).astype(np.float32); fp = np.angle(fv).astype(np.float32)
    p = np.zeros((N,2*L), dtype=signals.dtype); p[:,:L] = signals
    pp = np.abs(np.fft.fft(p, axis=1))**2
    ac = np.real(np.fft.ifft(pp, axis=1))[:,:L].astype(np.float32)
    lc = np.arange(L,0,-1,dtype=np.float32)[None,:]; ac /= lc
    an = ac[:,0:1]; ac = np.where(an>0, ac/an, ac)
    pu = np.unwrap(np.angle(signals), axis=1)
    inf = np.zeros((N,L), dtype=np.float32)
    inf[:,1:] = np.diff(pu, axis=1).astype(np.float32)/(2*np.pi*config.SAMPLING_INTERVAL)
    inf[:,0] = inf[:,1]; inf = np.clip(inf,-1000,1000)
    s = np.std(inf, axis=1, keepdims=True); inf = np.where(s>0, inf/(s+1e-8), inf)
    return np.stack([r,i,m,lf,fp,ac,inf], axis=1)
sb, _ = make_test(-35, 200, 77)
t0=time.time()
for _ in range(3): fo=extract_features_batch(sb)
to=(time.time()-t0)/3
t0=time.time()
for _ in range(3): ff=fused(sb)
tf=(time.time()-t0)/3
md = float(np.max(np.abs(fo-ff)))
print(f"  Orig={to:.4f}s Fused={tf:.4f}s Speed={to/max(tf,1e-9):.2f}x Diff={md:.1e}")
R["features"] = {"speedup": to/max(tf,1e-9), "max_diff": md}

# OPT5: DENOISING
print("\n" + "="*70 + "\nOPT5: DENOISING ABLATION\n" + "="*70)
from nqr_snn.denoising.selector import DenoisingSelector
from nqr_snn.denoising import denoise_batch
ov = config.VAL_SIZE; config.VAL_SIZE = 50
dd = generate_dataset("low", "white", "val", seed=42); config.VAL_SIZE = ov
sel = DenoisingSelector()
dn, dr, do = sel.select(dd["noisy"][:50], dd["clean"][:50])
print(f"  Denoiser: {dn} R2={dr:.1f}")
trd = NQRDatasetV2(tr["signals"], tr["labels"], denoiser=do)
vld = NQRDatasetV2(vl["signals"], vl["labels"], denoiser=do)
ed, _ = train_ens(3, get_balanced_loader_v2(trd,64), get_balanced_loader_v2(vld,64,shuffle=False), "outputs/m_den")
def eval_d(ens, enc, den, tag=""):
    tt=0; res={}
    for snr in config.EVAL_SNR_LEVELS:
        s,l = make_test(snr, seed=abs(snr)*100)
        t0=time.time(); sd=denoise_batch(den,s); f=extract_features_batch(sd)
        x=enc.encode(torch.from_numpy(f)).to(DEVICE); mp,_=ens.predict(x); tt+=time.time()-t0
        r=full_report(l,mp.cpu().numpy()); res[snr]=r
        print(f"  {tag} SNR={snr:4d} | acc={r['accuracy']:.4f}")
    a=[v['accuracy'] for v in res.values()]
    print(f"  {tag} MEAN={np.mean(a):.4f} MIN={np.min(a):.4f} T={tt:.3f}s")
    return {"mean_acc":float(np.mean(a)),"min_acc":float(np.min(a)),"inference_time":float(tt)}
R["denoised"] = eval_d(ed, enc, do, "[DEN]")

# OPT6: BS128
print("\n" + "="*70 + "\nOPT6: BS=128\n" + "="*70)
e128, tt128 = train_ens(3, tl128, vl128, "outputs/m_128")
R["bs128"] = {**evaluate(e128, enc, "[BS128]"), "train_time": tt128}

# OPT2: ENS5
print("\n" + "="*70 + "\nOPT2: ENSEMBLE 5\n" + "="*70)
e5, tt5 = train_ens(5, tl64, vl64, "outputs/m_ens5")
R["ens5"] = {**evaluate(e5, enc, "[ENS5]"), "train_time": tt5}

# SUMMARY
print("\n" + "="*70)
print("FINAL RESULTS")
print("="*70)
ba = R["baseline"]["mean_acc"]; bt = R["baseline"]["inference_time"]
print(f"{'Config':<15} {'MeanAcc':>8} {'MinAcc':>8} {'Time':>7} {'Delta':>7} {'Speed':>7} Verdict")
print("-"*70)
for n, r in R.items():
    if n == "features": continue
    a,mi,t = r["mean_acc"],r["min_acc"],r["inference_time"]
    d=a-ba; sp=bt/max(t,1e-9)
    v = "PASS" if a>=ba-0.005 else "WORSE"
    if a>ba+0.005: v="BETTER"
    print(f"{n:<15} {a:>7.4f} {mi:>8.4f} {t:>7.3f} {d:>+6.4f} {sp:>6.2f}x  {v}")
fe=R["features"]
print(f"\nFeatures: {fe['speedup']:.2f}x speedup (diff={fe['max_diff']:.1e})")

os.makedirs("outputs/results", exist_ok=True)
with open("outputs/results/optimization_results.json","w") as f:
    json.dump(R, f, indent=2, default=str)
print("\nSaved: outputs/results/optimization_results.json")
print("\nDONE. Share the output above and I will commit the winning optimizations.")