def _init(self, model_name):
if model_name is not None:
self.model_definition = model_name
else:
self.model_definition = self.config["model"]
if self.model_definition:
self.feature_mode = self.model_definition.split('_')[0]
else:
self.feature_mode = None
self.max_files_per_class = self.config.get(
"training/max_files_per_class", None)
self.dataset = None
if self.experiment_name:
self._model_dir = os.path.join(
os.path.expanduser(self.config["paths/model_dir"]),
self.experiment_name)
self._experiment_config_path = os.path.join(
self._model_dir, self.config['experiment/config_path'])
# if these don't exist, we're not actually running anything
if self.model_definition and self.feature_mode:
utils.create_directory(self._model_dir)
def _init(self, model_name):
if model_name is not None:
self.model_definition = model_name
else:
self.model_definition = self.config["model"]
if self.model_definition:
self.feature_mode = self.model_definition.split('_')[0]
else:
self.feature_mode = None
self.max_files_per_class = self.config.get(
"training/max_files_per_class", None)
self.dataset = None
if self.experiment_name:
self._model_dir = os.path.join(
os.path.expanduser(self.config["paths/model_dir"]),
self.experiment_name)
self._experiment_config_path = os.path.join(
self._model_dir, self.config['experiment/config_path'])
# if these don't exist, we're not actually running anything
if self.model_definition and self.feature_mode:
utils.create_directory(self._model_dir)
def _init_cross_validation(self, test_set):
self._cv_model_dir = os.path.join(self._model_dir, test_set)
self._params_dir = os.path.join(self._cv_model_dir,
self.config["experiment/params_dir"])
self._training_loss_path = os.path.join(
self._cv_model_dir, self.config['experiment/training_loss'])
if os.path.exists(self._cv_model_dir):
logger.warning("Cleaning old experiment: {}".format(
self._cv_model_dir))
utils.create_directory(
self._cv_model_dir,
# aka if DO the clean, recreate.
recreate=(not self.skip_cleaning))
utils.create_directory(self._params_dir)
def _init_cross_validation(self, test_set):
self._cv_model_dir = os.path.join(self._model_dir, test_set)
self._params_dir = os.path.join(
self._cv_model_dir,
self.config["experiment/params_dir"])
self._training_loss_path = os.path.join(
self._cv_model_dir, self.config['experiment/training_loss'])
if os.path.exists(self._cv_model_dir):
logger.warning("Cleaning old experiment: {}".format(
self._cv_model_dir))
utils.create_directory(self._cv_model_dir,
# aka if DO the clean, recreate.
recreate=(not self.skip_cleaning))
utils.create_directory(self._params_dir)
def test_create_directory(workspace):
dname = os.path.join(workspace, "oh_hello")
assert not os.path.exists(dname)
assert utils.create_directory(dname)
assert utils.create_directory(dname)
def test_create_directory(workspace):
dname = os.path.join(workspace, "oh_hello")
assert not os.path.exists(dname)
assert utils.create_directory(dname)
assert utils.create_directory(dname)
def cqt_from_dataset(dataset, write_dir,
cqt_params=None, audio_params=None, harmonic_params=None,
num_cpus=-1, verbose=50, skip_existing=True):
"""Compute CQT representation over audio files referenced by
a dataframe, and return a new dataframe also containing a column
referencing the cqt files.
Parameters
----------
dataset : hcnn.data.Dataset
Dataset containing references to the audio files.
write_dir : str
Directory to write to.
cqt_params : dict, default=None
Parameters to use for CQT computation.
audio_params : dict, default=None
Parameters to use for loading the audio file.
harmonic_params : dict, default=None
Parameters to use on top of `cqt_params` for the harmonic cqt.
num_cpus : int, default=-1
Number of parallel threads to use for computation.
verbose : int
Passed to cqt_many; for "Parallel"
skip_existing : bool
If files exist, don't try to extract them.
Returns
-------
updated_dataset : data.dataset.Dataset
Dataset updated with parameters to the outputed features.
"""
utils.create_directory(write_dir)
####
## TODO IF skip_existing, try to reload the dataset with features
## And modify it instead of replacing it.
def features_path_for_audio(audio_path):
return os.path.join(write_dir,
utils.filebase(audio_path) + ".npz")
audio_paths = dataset.to_df()["audio_file"].tolist()
cqt_paths = [features_path_for_audio(x) for x in audio_paths]
failed_files = cqt_many(audio_paths, cqt_paths, cqt_params, audio_params,
harmonic_params, num_cpus, verbose, skip_existing)
logger.warning("{} files failed to extract.".format(len(failed_files)))
feats_df = dataset.to_df()
feats_df['cqt'] = pd.Series([None] * len(feats_df), index=feats_df.index)
# Update the features field if the file was successfully created.
for i, path in enumerate(cqt_paths):
if os.path.exists(path):
feats_df.loc[feats_df.index[i], "cqt"] = path
else:
logger.warning("CQT Not successfully created: {}".format(path))
return DS.Dataset(feats_df, dataset.split)