def main(models, experiments, dataset_dir, phi, low, high):
# create kaldi instance
model_dir = select_model(models, phi, low, high)
kaldi = Kaldi.from_trained_model(
model_dir=model_dir,
base_dir=experiments.joinpath(
f'{time.strftime("%Y-%m-%d")}_{pydng.generate_name()}'))
# prepare dataset
dataset = Dataset(dataset_dir)
kaldi_dataset_dir = kaldi.base_dir.joinpath("data", dataset.name)
dataset.dump_as_kaldi_dataset(kaldi_dataset_dir,
wavs_prefix=f'data/{dataset.name}')
# decode
wer, meta = kaldi.decode_wavs(data_dir=dataset.data_dir, text=dataset.text)
print(f'\n[+] WER {dataset}: {wer:03.2f}%')
for utt in meta:
print(f"\n[+] {utt['wav_name']}")
print(f" REF: {utt['ref']}")
print(f" HYP: {utt['hyp']}")
print(f" WER: {utt['wer']*100:5.2f}%")
def decode_wavs(self, data_dir, text, phi=None):
data_dir = Path(data_dir)
# create decode dir
decode_name = f'decode_job_{data_dir.name}_{int(time.time())}'
decode_dir = self.base_dir.joinpath(
f"exp/nnet5d_gpu_time/{decode_name}")
# create kaldi dataset
dataset = Dataset(data_dir, name=decode_name)
dataset.text = text
dataset.dump_as_kaldi_dataset(
decode_dir, wavs_prefix=f"exp/nnet5d_gpu_time/{decode_name}")
# invoke decode script in container
self.run_in_container(
f'./decode_wavs.sh exp/nnet5d_gpu_time/{decode_name}',
additional_cmds=[
f'-e NUMJOBS={min(len(dataset), cpu_count())}',
f'-e PHI={phi} -e LOG_CLAMP=0'
])
# get best_wer
best_wer = open(f'{decode_dir}/scoring_kaldi/best_wer').read().strip()
best_wer = float(re.findall(r'%WER (.*) \[', best_wer)[0])
# get decoding meta
decoding_meta = parse_per_utt_file(decode_dir)
return best_wer, decoding_meta
def main(models, experiments, dataset_dir, inner_itr, max_itr, learning_rate, psycho_hiding_thresh, phi, low, high, attacker):
# create kaldi instance
model_dir = select_model(models, phi, low, high)
kaldi = Kaldi.from_trained_model(model_dir=model_dir,
base_dir=experiments.joinpath(f'{time.strftime("%Y-%m-%d")}_{pydng.generate_name()}'))
# prepare dataset
dataset = Dataset(dataset_dir)
kaldi_dataset_dir = kaldi.base_dir.joinpath("data", dataset.name)
dataset.dump_as_kaldi_dataset(kaldi_dataset_dir, wavs_prefix=f'data/{dataset.name}')
# optimization steps
max_outer_itr = max_itr // inner_itr
# dump config
print(f'\n[+] Compute adversarial examples')
print(f' -> attacker "{attacker}"')
print(f' -> dataset "{dataset}"')
print(f' -> psycho_hiding_thresh "{psycho_hiding_thresh}dB"')
print(f' -> {max_outer_itr} * {inner_itr} = {inner_itr*max_outer_itr} itr')
print(f' -> phi={phi} bandpass={low}-{high}')
print(f' -> learning_rate={learning_rate}')
kaldi.results['ae_config'] = {
'phi' : phi,
'low' : low,
'high': high,
'learning_rate' : learning_rate,
'attacker' : attacker,
'dataset' : dataset.name,
'psycho_hiding_thresh' : psycho_hiding_thresh,
'inner_itr' : inner_itr,
'max_outer_itr' : max_outer_itr
}
# decode
wer, _ = kaldi.decode_wavs(data_dir=dataset.data_dir, text=dataset.target)
print(f' -> inital WER: {wer:03.2f}%')
kaldi.results['inital_wer'] = f'{wer:03.2f}%'
# psychoacoustic filter causes an unstable gradient for backpropping to raw audio
# => caused by small input values to the log component
# => thus, for backprop we clamp input before the log component (@ nnet-component.cc)
log_clamp = 1 if phi != "None" else 0
# invoke adversarial examples script
logger = KaldiLogger(kaldi.base_dir)
logger.log_ae(inner_itr, max_outer_itr)
try:
running_time = kaldi.run_in_container(
f'./compute_adversarial_examples.sh {dataset} {psycho_hiding_thresh} {inner_itr} '
f'{max_outer_itr} {len(dataset)} {len(dataset)} {attacker}',
additional_cmds=[f'-v {kaldi_dataset_dir.joinpath("target_utterances")}:/root/kaldi/wsj_recipe/targets',
f'-e NUMJOBS={len(dataset)}',
f'-e PHI={phi} -e LOG_CLAMP={log_clamp} -e LEARNING_RATE={learning_rate}']
)
logger.stop()
kaldi.results['running_time'] = f'{running_time // 3600}h {(running_time % 3600) // 60}m {(running_time % 60)}s'
print(f" completed in {kaldi.results['running_time']}")
except KeyboardInterrupt:
logger.stop()
print(" terminated prematurely")
# score AEs
kaldi.results['history'] = parse_results_file(kaldi.base_dir)
score_AEs(kaldi=kaldi,
ae_dir=kaldi.base_dir.joinpath(f"adversarial_examples/wavs"),
ref_dir=dataset.data_dir,
stats_dir=kaldi.base_dir.joinpath(f"adversarial_examples/stats"),
original_text=dataset.text,
target_text=dataset.target,
phi=phi)
