Create examples/reverb_ood.py
Browse files- examples/reverb_ood.py +77 -0
examples/reverb_ood.py
ADDED
|
@@ -0,0 +1,77 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/usr/bin/env python3
|
| 2 |
+
"""
|
| 3 |
+
Out-of-distribution reverb test.
|
| 4 |
+
|
| 5 |
+
The main demo's reverb test uses rt60_s=0.35 for both training
|
| 6 |
+
augmentation and evaluation. That measures whether the classifier
|
| 7 |
+
handles reverb from the family it saw during training. It does
|
| 8 |
+
not measure robustness to reverb it has never seen.
|
| 9 |
+
|
| 10 |
+
This script trains with rt60_s=0.35 and evaluates at several
|
| 11 |
+
decay times the classifier never saw. The resulting accuracy
|
| 12 |
+
curve is the honest robustness measurement.
|
| 13 |
+
|
| 14 |
+
Run: python examples/reverb_ood.py
|
| 15 |
+
"""
|
| 16 |
+
import numpy as np
|
| 17 |
+
from whisper_decoder import (
|
| 18 |
+
MLP, train_mlp, generate_corpus, apply_reverb,
|
| 19 |
+
synth_word, extract_features, FEATURE_DIM,
|
| 20 |
+
WORD_ORDER, SR,
|
| 21 |
+
)
|
| 22 |
+
|
| 23 |
+
|
| 24 |
+
def eval_at_rt60(model, mu, sigma, rt60, n_per_word=20,
|
| 25 |
+
seed_base=200):
|
| 26 |
+
import random
|
| 27 |
+
rng = random.Random(seed_base + int(rt60 * 100))
|
| 28 |
+
n_words = len(WORD_ORDER)
|
| 29 |
+
X, y = [], []
|
| 30 |
+
for wi, word in enumerate(WORD_ORDER):
|
| 31 |
+
for k in range(n_per_word):
|
| 32 |
+
s = rng.randint(1, 1 << 30)
|
| 33 |
+
sig = synth_word(word, 'whisper', s)
|
| 34 |
+
sig = apply_reverb(sig, rt60_s=rt60,
|
| 35 |
+
seed=rng.randint(1, 1 << 30))
|
| 36 |
+
X.append(extract_features(sig))
|
| 37 |
+
y.append(wi)
|
| 38 |
+
X = (np.stack(X) - mu) / sigma
|
| 39 |
+
y = np.array(y)
|
| 40 |
+
return float((model.predict(X) == y).mean())
|
| 41 |
+
|
| 42 |
+
|
| 43 |
+
def main():
|
| 44 |
+
print("training on whisper + 25% reverb (rt60=0.35)...")
|
| 45 |
+
train = generate_corpus('whisper', 200, 40, seed=0,
|
| 46 |
+
reverb_train_frac=0.25)
|
| 47 |
+
mu = train.X_train.mean(axis=0)
|
| 48 |
+
sigma = train.X_train.std(axis=0) + 1e-9
|
| 49 |
+
X_tr = (train.X_train - mu) / sigma
|
| 50 |
+
|
| 51 |
+
model = MLP(FEATURE_DIM, 64, 32, 10, seed=0)
|
| 52 |
+
train_mlp(model, X_tr, train.y_train, epochs=200,
|
| 53 |
+
batch=64, lr=3e-3, seed=0)
|
| 54 |
+
|
| 55 |
+
print()
|
| 56 |
+
print(f" {'rt60 (s)':>10} {'accuracy':>10} note")
|
| 57 |
+
print(" " + "-" * 50)
|
| 58 |
+
for rt60 in [0.0, 0.10, 0.20, 0.35, 0.50, 0.70, 1.00, 1.50]:
|
| 59 |
+
acc = eval_at_rt60(model, mu, sigma, rt60)
|
| 60 |
+
if rt60 == 0.0:
|
| 61 |
+
note = "dry"
|
| 62 |
+
elif rt60 <= 0.40:
|
| 63 |
+
note = "in-distribution"
|
| 64 |
+
elif rt60 <= 0.70:
|
| 65 |
+
note = "mildly OOD"
|
| 66 |
+
else:
|
| 67 |
+
note = "strongly OOD"
|
| 68 |
+
print(f" {rt60:>10.2f} {acc*100:>9.1f}% {note}")
|
| 69 |
+
|
| 70 |
+
print()
|
| 71 |
+
print(" The training reverb distribution is centered at "
|
| 72 |
+
"0.35 s.")
|
| 73 |
+
print(" Accuracy at rt60 > 0.5 is out-of-distribution.")
|
| 74 |
+
|
| 75 |
+
|
| 76 |
+
if __name__ == "__main__":
|
| 77 |
+
main()
|