class TextSpeechDataset(Dataset):
def __init__(self, root_dir, annotations_file, parameters):
self.root_dir = root_dir
self.annotations = pd.read_csv(os.path.join(root_dir, annotations_file), sep="|", header=None)
self.cleaners = parameters["text_cleaner"]
self.outputs_per_step = parameters["outputs_per_step"]
self.sample_rate = parameters["sample_rate"]
self.ap = AudioProcessor(sample_rate=self.sample_rate, preemphasis=parameters["preemphasis"],
frequency=parameters["frequency"], frame_length=parameters["frame_length"],
frame_shift=parameters["frame_shift"], min_dbs=parameters["min_dbs"],
ref_dbs=parameters["ref_dbs"],
mels_size=parameters["mels_size"],
griff_lim_iters=parameters["griff_lim_iters"],
power=parameters["spectro_power"]
)
def __len__(self):
return self.annotations.shape[0]
def __getitem__(self, index):
text = self.annotations.iloc[index][1]
text = np.asarray(text_to_sequence(text=text, cleaner_names=[self.cleaners]), dtype=np.int32)
wav_path = os.path.join(self.root_dir, "wavs", f"{self.annotations.iloc[index][0]}.wav")
wav = load_wav(filename=wav_path, sample_rate=self.sample_rate)
return {"text": text, "wav": wav}
def collate_fn(self, batch):
wavs = list(map(lambda sample: sample["wav"], batch))
texts = list(map(lambda sample: sample["text"], batch))
max_text_len = np.max(list(map(lambda txt: len(txt), texts)))
texts = np.stack(list(map(lambda txt: pad_data(txt, max_text_len), texts))).astype(np.int32)
max_wav_len = np.max(list(map(lambda waveform: len(waveform), wavs)))
wavs = np.stack(list(map(lambda waveform: pad_data(x=waveform, length=max_wav_len), wavs))).astype(np.int32)
linears = list(
map(lambda waveform: self.ap.wav_to_linear_spectrogram(waveform=waveform).astype(np.float32), wavs))
mels = list(map(lambda waveform: self.ap.wav_to_mel_spectrogram(waveform=waveform).astype(np.float32), wavs))
mel_lengths = list(map(lambda mel_: mel_.shape[1] + 1, mels))
linears = prepare_tensor(inputs=linears, out_steps=self.outputs_per_step)
mels = prepare_tensor(inputs=mels, out_steps=self.outputs_per_step)
linears = linears.transpose(0, 2, 1)
mels = mels.transpose(0, 2, 1)
texts = torch.LongTensor(texts)
linears = torch.FloatTensor(linears)
mels = torch.FloatTensor(mels)
mel_lengths = torch.LongTensor(mel_lengths)
return texts, linears, mels, mel_lengths
def __init__(self, root_dir, annotations_file, parameters):
self.root_dir = root_dir
self.annotations = pd.read_csv(os.path.join(root_dir, annotations_file), sep="|", header=None)
self.cleaners = parameters["text_cleaner"]
self.outputs_per_step = parameters["outputs_per_step"]
self.sample_rate = parameters["sample_rate"]
self.ap = AudioProcessor(sample_rate=self.sample_rate, preemphasis=parameters["preemphasis"],
frequency=parameters["frequency"], frame_length=parameters["frame_length"],
frame_shift=parameters["frame_shift"], min_dbs=parameters["min_dbs"],
ref_dbs=parameters["ref_dbs"],
mels_size=parameters["mels_size"],
griff_lim_iters=parameters["griff_lim_iters"],
power=parameters["spectro_power"]
)
def run_train(c, args, model_params):
c = copy_config_dict(c)
ap = AudioProcessor(**c.audio)
if args.seed is None:
log_path = os.path.join(c.train_config['logs_path'], c.model_name)
else:
log_path = os.path.join(os.path.join(c.train_config['logs_path'], str(args.seed)), c.model_name)
c.train_config['seed'] = args.seed
os.makedirs(log_path, exist_ok=True)
tensorboard = TensorboardWriter(os.path.join(log_path,'tensorboard'))
trainloader = train_dataloader(copy_config_dict(c), ap, class_balancer_batch=c.dataset['class_balancer_batch'])
max_seq_len = trainloader.dataset.get_max_seq_lenght()
c.dataset['max_seq_len'] = max_seq_len
model_params['config'] = copy_config_dict(c)
# save config in train dir, its necessary for test before train and reproducity
save_config_file(c, os.path.join(log_path,'config.json'))
# one_window in eval use overlapping
if c.dataset['temporal_control'] == 'one_window':
c.dataset['temporal_control'] = 'overlapping'
evaloader = eval_dataloader(c, ap, max_seq_len=max_seq_len)
# enablePrint()
# print(max_seq_len, trainloader.dataset.get_max_seq_lenght(), c.dataset['temporal_control'], c.dataset['max_seq_len'])
return train(args, log_path, args.checkpoint_path, trainloader, evaloader, tensorboard, c, c.model_name, ap, cuda=True, model_params=model_params)
def run_train(c, args, model_params=None):
ap = AudioProcessor(**c.audio)
log_path = os.path.join(c.train_config['logs_path'], c.model_name)
os.makedirs(log_path, exist_ok=True)
tensorboard = TensorboardWriter(os.path.join(log_path, 'tensorboard'))
print(c.dataset['train_csv'], c.dataset['eval_csv'])
trainloader = train_dataloader(
c, ap, class_balancer_batch=c.dataset['class_balancer_batch'])
max_seq_len = trainloader.dataset.get_max_seq_lenght()
c.dataset['max_seq_len'] = max_seq_len
print(c.dataset['train_csv'], c.dataset['eval_csv'])
# save config in train dir, its necessary for test before train and reproducity
save_config_file(c, os.path.join(log_path, 'config.json'))
evaloader = eval_dataloader(c, ap, max_seq_len=max_seq_len)
return train(args,
log_path,
args.checkpoint_path,
trainloader,
evaloader,
tensorboard,
c,
c.model_name,
ap,
cuda=True,
model_params=model_params)
type=str,
required=False,
default=False,
help="Librispeech format, if true load with librispeech format")
args = parser.parse_args()
os.makedirs(args.out_dir, exist_ok=True)
if args.train_data_csv:
os.makedirs(os.path.join(args.out_dir, 'train'), exist_ok=True)
if args.test_data_csv:
os.makedirs(os.path.join(args.out_dir, 'test'), exist_ok=True)
cpu_num = cpu_count() # num threads = num cpu cores
config = load_config(args.config)
ap = AudioProcessor(config.audio)
sample_rate = config.audio[config.audio['backend']]['sample_rate']
audio_len = config.audio['audio_len']
form = config.dataset['format']
output_dir_train = os.path.join(args.out_dir, 'train')
output_dir_test = os.path.join(args.out_dir, 'test')
dataset_root_dir = args.dataset_root_dir
train_data_csv = None
test_data_csv = None
noise_files = open(args.noise_csv).readlines()
if args.train_data_csv:
train_data_csv = pd.read_csv(args.train_data_csv, sep=',').values
outpath = os.path.join(
output_dir,
os.path.basename(mixed_path).replace(".wav", "_predict.wav"))
soundfile.write(outpath, est_wav, 16000)
return est_wav, target_wav, mixed_wav, emb_wav
print("... Load GE2E encoder model ...")
encoder.load_model(Path('encoder/saved_models/pretrained_en.pt'))
checkpoint_path = 'models/demo5.pt'
checkpoint = torch.load(checkpoint_path, map_location='cpu')
model_c = load_config_from_str(checkpoint['config_str'])
ap = AudioProcessor(model_c.audio) # create AudioProcessor for model
model_name = model_c.model_name
cuda = False
if model_name == 'voicefilter':
print('inicializado com voicefilter')
model = VoiceFilter(model_c)
elif model_name == 'voicesplit':
model = VoiceSplit(model_c)
else:
raise Exception(" The model '" + model_name + "' is not suported")
if model_c.train_config['optimizer'] == 'adam':
optimizer = torch.optim.Adam(model.parameters(),
lr=model_c.train_config['learning_rate'])
else:
raise Exception("The %s not is a optimizer supported" %
model_c.train['optimizer'])
help="Root directory of run.")
parser.add_argument('-c', '--config_path', type=str, required=False, default=None,
help="json file with configurations")
parser.add_argument('--checkpoints_path', type=str, required=True,
help="path of checkpoint pt file, for continue training")
args = parser.parse_args()
all_checkpoints = sorted(glob(os.path.join(args.checkpoints_path, '*.pt')))
#print(all_checkpoints, os.listdir(args.checkpoints_path))
if args.config_path:
c = load_config(args.config_path)
else: #load config in checkpoint
checkpoint = torch.load(all_checkpoints[0], map_location='cpu')
c = load_config_from_str(checkpoint['config_str'])
ap = AudioProcessor(c.audio)
log_path = os.path.join(c.train_config['logs_path'], c.model_name)
audio_config = c.audio[c.audio['backend']]
tensorboard = TensorboardWriter(log_path, audio_config)
# set test dataset dir
c.dataset['test_dir'] = args.dataset_dir
# set batchsize = 32
c.test_config['batch_size'] = 5
test_dataloader = test_dataloader(c, ap)
best_loss = 999999999
best_loss_checkpoint = ''
sdrs_checkpoint = []
for i in tqdm.tqdm(range(len(all_checkpoints))):
checkpoint = all_checkpoints[i]
mean_loss= test(args, log_path, checkpoint, test_dataloader, tensorboard, c, c.model_name, ap, cuda=True)
help="path of checkpoint pt file, for continue training")
parser.add_argument('--batch_size', type=int, default=20,
help="Batch size for test")
parser.add_argument('--num_workers', type=int, default=10,
help="Number of Workers for test data load")
parser.add_argument('--no_insert_noise', type=bool, default=False,
help=" No insert noise in test ?")
parser.add_argument('--num_noise_control', type=int, default=1,
help="Number of Noise for insert in control")
parser.add_argument('--num_noise_patient', type=int, default=0,
help="Number of Noise for insert in patient")
args = parser.parse_args()
c = load_config(args.config_path)
ap = AudioProcessor(**c.audio)
if not args.no_insert_noise:
c.data_aumentation['insert_noise'] = True
else:
c.data_aumentation['insert_noise'] = False
# ste values for noisy insertion in test
c.data_aumentation["num_noise_control"] = args.num_noise_control
c.data_aumentation["num_noise_patient"] = args.num_noise_patient
print("Insert noise ?", c.data_aumentation['insert_noise'])
c.dataset['test_csv'] = args.test_csv
c.dataset['test_data_root_path'] = args.test_root_dir
if __name__ == "__main__":
# Get defaults so it can work with no Sacred
parser = argparse.ArgumentParser()
parser.add_argument('-f', "--wavfile_path", required=True)
parser.add_argument('-c', '--config', type=str, required=True,
help='JSON file for configuration')
args = parser.parse_args()
config = load_config(args.config)
filepath = args.wavfile_path
# extract spectrogram
config.audio['feature'] = 'spectrogram'
ap = AudioProcessor(**config.audio)
spectrogram = ap.get_feature_from_audio_path(filepath)
print("Spectogram with shape:",spectrogram.shape)
# extract spectrogram
config.audio['feature'] = 'melspectrogram'
ap = AudioProcessor(**config.audio)
melspectrogram = ap.get_feature_from_audio_path(filepath)
print("MelSpectogram with shape:",melspectrogram.shape)
# extract spectrogram
config.audio['feature'] = 'mfcc'
ap = AudioProcessor(**config.audio)
mfcc = ap.get_feature_from_audio_path(filepath)
print("MFCC with shape:",mfcc.shape)
def run_test_all_seeds(args, cuda=True, debug=False, return_potential=False):
runs_list = os.listdir(args.experiment_dir)
runs_list.sort()
num_runs = len(runs_list)
votes = []
wav_files = []
targets = []
# define loss function
criterion = nn.BCELoss(reduction='sum')
for run in runs_list:
blockPrint()
run_dir = os.path.join(args.experiment_dir, run)
if os.path.isfile(run_dir):
continue
model_name = os.listdir(run_dir)[0]
checkpoint_path = os.path.join(run_dir, model_name,
'best_checkpoint.pt')
config_path = os.path.join(run_dir, model_name, 'config.json')
c = load_config(config_path)
ap = AudioProcessor(**c.audio)
c.dataset['test_csv'] = args.test_csv
c.dataset['test_data_root_path'] = args.test_root_dir
c.test_config['batch_size'] = args.batch_size
c.test_config['num_workers'] = args.num_workers
max_seq_len = c.dataset['max_seq_len']
c.train_config['seed'] = 0
testdataloader = test_dataloader(c, ap, max_seq_len=max_seq_len)
# load model
model = return_model(c)
enablePrint()
if checkpoint_path is not None:
print("Loading checkpoint: %s" % checkpoint_path)
try:
checkpoint = torch.load(checkpoint_path, map_location='cpu')
model.load_state_dict(checkpoint['model'])
print("Model Sucessful Load !")
except Exception as e:
raise ValueError(
"You need pass a valid checkpoint, may be you need check your config.json because de the of this checkpoint cause the error: "
+ e)
blockPrint()
# convert model from cuda
if cuda:
model = model.cuda()
model.train(False)
vote, targets, wav_path = test(criterion,
ap,
model,
c,
testdataloader,
cuda=cuda,
confusion_matrix=True,
debug=debug,
simples_vote=args.simples_vote)
# print(vote)
wav_files.append(wav_path)
votes.append(vote)
if len(wav_files):
if wav_files[-1] != wav_files[0]:
raise ValueError(
"Diferents files or order for the test in diferrents seeds or folds"
)
# mean vote, and round is necessary if use composite vote
preds = np.mean(np.array(votes), axis=0)
# print(preds)
if not return_potential:
preds = preds.round()
file_names = wav_files[0]
if debug and not return_potential:
enablePrint()
targets = np.array(targets)
preds = np.array(preds)
names = np.array(file_names)
idxs = np.nonzero(targets == c.dataset['control_class'])
control_target = targets[idxs]
control_preds = preds[idxs]
names_control = names[idxs]
idxs = np.nonzero(targets == c.dataset['patient_class'])
patient_target = targets[idxs]
patient_preds = preds[idxs]
names_patient = names[idxs]
if debug:
print('+' * 40)
print("Control Files Classified incorrectly:")
incorrect_ids = np.nonzero(
control_preds != c.dataset['control_class'])
inc_names = names_control[incorrect_ids]
print("Num. Files:", len(inc_names))
print(inc_names)
print('+' * 40)
print('-' * 40)
print("Patient Files Classified incorrectly:")
incorrect_ids = np.nonzero(
patient_preds != c.dataset['patient_class'])
inc_names = names_patient[incorrect_ids]
print("Num. Files:", len(inc_names))
print(inc_names)
print('-' * 40)
acc_control = (control_preds == control_target).mean()
acc_patient = (patient_preds == patient_target).mean()
acc_balanced = (acc_control + acc_patient) / 2
f1 = f1_score(targets.tolist(), preds.tolist())
uar = recall_score(targets.tolist(), preds.tolist(), average='macro')
print("======== Confusion Matrix ==========")
y_target = pd.Series(targets, name='Target')
y_pred = pd.Series(preds, name='Predicted')
df_confusion = pd.crosstab(y_target,
y_pred,
rownames=['Target'],
colnames=['Predicted'],
margins=True)
print(df_confusion)
print("Test\n ", "Acurracy Control: ", acc_control,
"Acurracy Patient: ", acc_patient, "Acurracy Balanced",
acc_balanced)
print("F1:", f1, "UAR:", uar)
if return_potential:
return preds, file_names
else:
df = pd.DataFrame({
'filename': file_names,
'prediction': preds.astype(int)
})
df['prediction'] = df['prediction'].replace(
int(c.dataset['control_class']), 'negative',
regex=True).replace(int(c.dataset['patient_class']),
'positive',
regex=True)
if args.output_csv:
out_csv_path = args.output_csv
else:
out_csv_path = os.path.join(
args.experiment_dir, os.path.basename(c.dataset['test_csv']))
df.to_csv(out_csv_path, index=False)
"audio_len": 3,
"voicefilter": {
"n_fft": 1200,
"num_mels": 40,
"num_freq": 601,
"sample_rate": 16000,
"hop_length": 160,
"win_length": 400,
"min_level_db": -100.0,
"ref_level_db": 20.0,
"preemphasis": 0.97,
"power": 1.5,
"griffin_lim_iters": 60
}
}
ap = AudioProcessor(config)
# In[9]:
os.listdir(TEST_DATA)
# In[12]:
#Preprocess dataset
train_files = sorted(glob(os.path.join(TRAIN_DATA, glob_re_wav_emb)))
test_files = sorted(glob(os.path.join(TEST_DATA, glob_re_wav_emb)))
if len(train_files) == 0 or len(test_files):
print("check train and test path files not in directory")
files = train_files + test_files
]
if __name__ == "__main__":
# Get defaults so it can work with no Sacred
parser = argparse.ArgumentParser()
parser.add_argument('-c',
'--config',
type=str,
required=True,
help='JSON file for configuration')
args = parser.parse_args()
config = load_config(args.config)
# extract spectrogram
config.audio['feature'] = 'spectrogram'
ap = AudioProcessor(**config.audio)
max_amp = 0
min_amp = 99999999999999999
for samples_file in samples:
for sample in samples_file:
loc_max_amp = 0
wav = ap.load_wav(sample[0])
inicio = sample[1]
fim = sample[2]
slice_start = int(inicio * ap.sample_rate)
slice_end = int(fim * ap.sample_rate)
wav = wav[:, slice_start:slice_end]
wav_max_amp = wav.max().numpy()
if wav_max_amp > loc_max_amp:
loc_max_amp = wav_max_amp
