def predict_one_label(self, data):
self.query_num += 1
data = data.squeeze()
pred_real, pred_pro = Sincnet.sentence_test(
self.model,
torch.from_numpy(data).float().cuda())
return pred_real
if flag==False:
if a.split('-')[0]==str(start):
flag=True
else:
continue
if MODE=="Librispeech":
real_name=a.split('-')[2]
target_name = t.split('-')[2]
elif MODE=="TIMIT":
real_name = a.split('-')[-1].split('.')[0].lower()
target_name = t.split('-')[-1].split('.')[0].lower()
# read audio
real_data, fs = sf.read(os.path.join(attackdir, a))
target_data, fs = sf.read(os.path.join(targetdir, t))
# test is ok
pid, _ = Sincnet.sentence_test(model,torch.from_numpy(target_data).float().cuda())
target_index = speaker_label[target_name]
if pid != target_index:
print("error")
continue
# set file name
print("now attack audio:", a)
dl = a.split('-')[0]
# local and attenutation attack
save_pfn=os.path.join(abs_path,save_dir,"test-{}.txt".format(dl))
save_info_n=os.path.join(abs_path,save_dir,"info-{}.txt".format(dl))
laatk=LOCAL_ATT_HSJA_ATTACK.LAATTACK(os.path.join(attackdir, a), os.path.join(targetdir, t),
save_p_fname=save_pfn, save_info_fname=save_info_n, MODE=MODE, dct_field=0.65)
o2_audio, preturb,query_num,interval=laatk.targeted_attack()
# save audio
def predict_one_label(model, audio):
qq, _ = Sincnet.sentence_test(model, audio.float().cuda())
return qq
def show_max(wav):
qq, _ = Sincnet.sentence_test(model, wav.float().cuda())
print(qq)
def predict_label(self,data):
pred_real, pred_pro = Sincnet.sentence_test(self.model,
torch.from_numpy(data).float().cuda())
return pred_real