def validation(v_trainset, model_end, X_test):
boxmin=0 #Minimum signal value (default 0).
boxmax=1 #Maximum signal value (default 1).
boxmul = (boxmax - boxmin) / 2.
boxplus = (boxmin + boxmax) / 2.
num_false_after_perturbation=0
snr_t=0
for s in range(X_test.shape[0]):
X_test_arced=np.arctanh((X_test[s] - boxplus) / boxmul * 0.999999)
audio_pert=np.add(X_test_arced,v_trainset)
audio_pert=np.tanh(audio_pert)* boxmul + boxplus
test_s=np.squeeze(X_test_arced)
test_s=np.tanh(test_s)*boxmul + boxplus
snr = 20 * np.log10(energy_np(np.squeeze(test_s))/energy_np(np.squeeze(audio_pert)-test_s))
snr_t=snr+snr_t
classified = np.argmax(model_end.predict([np.expand_dims(audio_pert, axis=0)]), axis=1)
correct=np.argmax(model_end.predict(np.expand_dims(X_test[s], axis=0)))
if (classified != correct):
num_false_after_perturbation = num_false_after_perturbation + 1
fooling_rate_test = float(num_false_after_perturbation)/float(X_test.shape[0])
SNR_avg = snr_t / float(X_test.shape[0])
return SNR_avg, fooling_rate_test
def validation(v, Target_class, model_end, X_test):
num_false_after_perturbation = 0
snr_t = 0
for s in range(X_test.shape[0]):
v_fin = np.expand_dims(X_test[s], axis=0) + v
snr = 20.0 * np.log10(
energy_np(np.squeeze(X_test[s])) / energy_np(v_fin - X_test[s]))
snr_t = snr + snr_t
classified = np.argmax(model_end.predict([v_fin]))
correct = np.argmax(
model_end.predict(np.expand_dims(X_test[s], axis=0)))
if (classified != correct):
num_false_after_perturbation = num_false_after_perturbation + 1
return snr_t / float(
X_test.shape[0]), float(num_false_after_perturbation) / float(
X_test.shape[0])
def validation(v, model_end, X_test):
num_false_after_perturbation=0
snr_t=0
for s in range(X_test.shape[0]):
wav_test_pert=np.expand_dims(X_test[s], axis=0)+v
wav_test_pert = np.clip(wav_test_pert, 0, 1)
snr = 20 * np.log10(energy_np(np.squeeze(X_test[s]))/energy_np(wav_test_pert-X_test[s]))
snr_t=snr+snr_t
if np.argmax(model_end.predict(np.expand_dims(X_test[s], axis=0)))!=np.argmax(model_end.predict(wav_test_pert)):
num_false_after_perturbation=num_false_after_perturbation+1
return snr_t / float(X_test.shape[0]), float(num_false_after_perturbation)/float(X_test.shape[0])