def get_variables(args):
model_pth = args.model_path
gt_fname, ext = osp.splitext(args.groundtruth_tsv)
median_win = args.median_window
meta_gt = args.meta_gt
gt_audio_pth = args.groundtruth_audio_dir
if meta_gt is None:
meta_gt = gt_fname + "_durations" + ext
if gt_audio_pth is None:
gt_audio_pth = meta_path_to_audio_dir(gt_fname)
# Useful because of the data format
if "validation" in gt_audio_pth:
gt_audio_pth = osp.dirname(gt_audio_pth)
groundtruth = pd.read_csv(args.groundtruth_tsv, sep="\t")
if osp.exists(meta_gt):
meta_dur_df = pd.read_csv(meta_gt, sep='\t')
if len(meta_dur_df) == 0:
meta_dur_df = generate_tsv_wav_durations(gt_audio_pth, meta_gt)
else:
meta_dur_df = generate_tsv_wav_durations(gt_audio_pth, meta_gt)
return model_pth, median_win, gt_audio_pth, groundtruth, meta_dur_df
def initialize_and_get_df(self,
tsv_path,
audio_dir=None,
audio_dir_ss=None,
pattern_ss=None,
ext_ss_feature_file="_ss",
nb_files=None,
download=True):
""" Initialize the dataset, extract the features dataframes
Args:
tsv_path: str, tsv path in the initial dataset
audio_dir: str, the path where to search the filename of the df
audio_dir_ss: str, the path where to search the separated_sources
pattern_ss: str, only when audio_dir_ss is not None, this should be defined. The pattern that's added
after normal filenames to get associated separated sources (have been done during source separation)
ext_ss_feature_file: str, only when audio_dir_ss is not None, what to add at the end of the feature files
nb_files: int, optional, the number of file to take in the dataframe if taking a small part of the dataset.
download: bool, optional, whether or not to download the data from the internet (youtube).
Returns:
pd.DataFrame
The dataframe containing the right features and labels
"""
# Parameters
if audio_dir_ss is not None and pattern_ss is None:
pattern_ss = "_events"
if audio_dir is None:
audio_dir = meta_path_to_audio_dir(tsv_path)
# Path to save features, subdir, otherwise could have duplicate paths for synthetic data
fdir = audio_dir if audio_dir_ss is None else audio_dir_ss
fdir = fdir[:-1] if fdir.endswith(osp.sep) else fdir
subdir = osp.sep.join(fdir.split(osp.sep)[-2:])
meta_feat_dir = osp.join(self.meta_feat_dir, subdir)
feature_dir = osp.join(self.feature_dir, subdir)
os.makedirs(meta_feat_dir, exist_ok=True)
os.makedirs(feature_dir, exist_ok=True)
df_meta = self.get_df_from_meta(tsv_path,
nb_files,
pattern_ss=pattern_ss)
logger.info(
f"{tsv_path} Total file number: {len(df_meta.filename.unique())}")
# Download real data
if download:
# Get only one filename once
filenames = df_meta.filename.drop_duplicates()
self.download(filenames, audio_dir)
# Meta filename
ext_tsv_feature = ""
if audio_dir_ss is not None:
ext_tsv_feature = ext_ss_feature_file
fname, ext = osp.splitext(osp.basename(tsv_path))
feat_fname = fname + ext_tsv_feature + ext
if nb_files is not None:
feat_fname = f"{nb_files}_{feat_fname}"
features_tsv = osp.join(meta_feat_dir, feat_fname)
# if not osp.exists(features_tsv):
t = time.time()
logger.info(f"Getting features ...")
df_features = self.extract_features_from_df(df_meta, audio_dir,
feature_dir, audio_dir_ss,
pattern_ss,
ext_ss_feature_file)
if len(df_features) != 0:
df_features.to_csv(features_tsv, sep="\t", index=False)
logger.info(
f"features created/retrieved in {time.time() - t:.2f}s, metadata: {features_tsv}"
)
else:
raise IndexError(f"Empty features DataFrames {features_tsv}")
return df_features