def get_folds_data(corrections=None):
print("Start generate folds data")
print("Audio config", get_audio_config())
train_folds_df = pd.read_csv(config.train_folds_path)
audio_paths_lst = []
targets_lst = []
folds_lst = []
for i, row in train_folds_df.iterrows():
labels = row.labels
if corrections is not None:
if row.fname in corrections:
action = corrections[row.fname]
if action == 'remove':
print(f"Skip {row.fname}")
continue
else:
print(
f"Replace labels {row.fname} from {labels} to {action}"
)
labels = action
folds_lst.append(row.fold)
audio_paths_lst.append(row.file_path)
target = torch.zeros(len(config.classes))
for label in labels.split(','):
target[config.class2index[label]] = 1.
targets_lst.append(target)
with mp.Pool(N_WORKERS) as pool:
images_lst = pool.map(read_as_melspectrogram, audio_paths_lst)
return images_lst, targets_lst, folds_lst
def get_corrected_noisy_data():
print("Start generate corrected noisy data")
print("Audio config", get_audio_config())
train_noisy_df = pd.read_csv(config.train_noisy_csv_path)
with open(config.noisy_corrections_json_path) as file:
corrections = json.load(file)
audio_paths_lst = []
targets_lst = []
for i, row in train_noisy_df.iterrows():
labels = row.labels
if row.fname in corrections:
action = corrections[row.fname]
if action == 'remove':
continue
else:
labels = action
else:
continue
audio_paths_lst.append(config.train_noisy_dir / row.fname)
target = torch.zeros(len(config.classes))
for label in labels.split(','):
target[config.class2index[label]] = 1.
targets_lst.append(target)
with mp.Pool(N_WORKERS) as pool:
images_lst = pool.map(read_as_melspectrogram, audio_paths_lst)
return images_lst, targets_lst
def get_augment_folds_data_generator(time_stretch_lst, pitch_shift_lst):
print("Start generate augment folds data")
print("Audio config", get_audio_config())
print("time_stretch_lst:", time_stretch_lst)
print("pitch_shift_lst:", pitch_shift_lst)
train_folds_df = pd.read_csv(config.train_folds_path)
audio_paths_lst = []
targets_lst = []
folds_lst = []
for i, row in train_folds_df.iterrows():
folds_lst.append(row.fold)
audio_paths_lst.append(row.file_path)
target = torch.zeros(len(config.classes))
for label in row.labels.split(','):
target[config.class2index[label]] = 1.
targets_lst.append(target)
with mp.Pool(N_WORKERS) as pool:
images_lst = pool.map(read_as_melspectrogram, audio_paths_lst)
yield images_lst, targets_lst, folds_lst
images_lst = []
for pitch_shift in pitch_shift_lst:
pitch_shift_read = partial(read_as_melspectrogram,
pitch_shift=pitch_shift)
with mp.Pool(N_WORKERS) as pool:
images_lst = pool.map(pitch_shift_read, audio_paths_lst)
yield images_lst, targets_lst, folds_lst
images_lst = []
for time_stretch in time_stretch_lst:
time_stretch_read = partial(read_as_melspectrogram,
time_stretch=time_stretch)
with mp.Pool(N_WORKERS) as pool:
images_lst = pool.map(time_stretch_read, audio_paths_lst)
yield images_lst, targets_lst, folds_lst
images_lst = []