def test_backfill_noise(noise_shape):
noise = np.random.random(noise_shape)
for t_len in [1, 2, 3, 8, 10, 43, 100]:
backfilled = utils.backfill_noise(noise, t_len)
assert backfilled.shape[-2] >= t_len
assert noise.shape[:-2] == backfilled.shape[:-2]
if noise.shape[0] > 1:
# Ideally we want to check that these
# are not all equal, but it should suffice to check
# that they're not equal with the first.
for i in range(1, noise.shape[0]):
assert not np.array_equal(backfilled[0], backfilled[i])
def test_backfill_noise(noise_shape):
noise = np.random.random(noise_shape)
for t_len in [1, 2, 3, 8, 10, 43, 100]:
backfilled = utils.backfill_noise(noise, t_len)
assert backfilled.shape[-2] >= t_len
assert noise.shape[:-2] == backfilled.shape[:-2]
if noise.shape[0] > 1:
# Ideally we want to check that these
# are not all equal, but it should suffice to check
# that they're not equal with the first.
for i in range(1, noise.shape[0]):
assert not np.array_equal(backfilled[0], backfilled[i])
def base_slicer(record,
t_len,
obs_slicer,
shuffle=True,
auto_restart=True,
add_noise=True,
random_seed=None,
npz_data_key='cqt',
slice_logger=None):
"""Base slicer function for yielding data from an .npz file
which contains the features for streaming.
Assumptions:
* Input file is about 1s in length (43 frames), but might be
shorter or longer. t_len should be 43 (~1s), however.
* It is possible that this file could get sampled more than
once consecutively, so for training, if shuffle is True,
and the number of frames in the file > t_len, it will
try to sample from frames other than the first t_len frames
randomly.
* For prediction (shuffle=False), it will always return
the first t_len frames.
To use this for training, use the following parameters:
(...shuffle=True, auto_restart=True, add_noise=True)
For prediction / eval, use the following parameters:
(...shuffle=False, auto_restart=False, add_noise=False)
Parameters
----------
record : pandas.Series
Single pandas record containing a 'cqt' record
which points to the cqt file in question.
Also must contain an "instrument" column
which contains the ground truth.
t_len : int
Length of the sliced array [in time/frames]
obs_slicer : function
Function which takes (cqt, idx, counter, t_len), and
returns a slice from the data, formatted correctly for
the desired data.
shuffle : bool
If True, shuffles the frames every time through the file.
If False, returns the first t_len frames.
auto_restart : bool
If True, yields infinitely.
If False, only goes through the file once.
add_noise : bool
If True, adds a small amount of noise to every sample.
If False, does nothing.
random_seed : int or None
If int, uses this number as the random seed. Otherwise,
makes it's own.
npz_data_key : str
The key in the npz file pointed to by record from which to
load the data. Choices = ['cqt', 'harmonic_cqt']
Yields
-------
sample : dict with fields {x_in, target}
The windowed observation.
"""
rng = np.random.RandomState(random_seed)
target = instrument_map.get_index(record['instrument'])
if slice_logger: slice_logger.start(record, target)
if not (('cqt' in record.index) and record['cqt'] is not None and
isinstance(record['cqt'], str) and os.path.exists(record['cqt'])):
logger.error(
'No valid feature file specified for record: {}'.format(record))
if slice_logger: slice_logger.error(record)
return
# Load the npz file with they key specified.
try:
cqt = np.load(record['cqt'])[npz_data_key]
except zipfile.BadZipFile:
logger.error('This zip file is bad! Sadness :( - {}'.format(
record['cqt']))
if slice_logger: slice_logger.error(record)
return
# Take the logmagnitude of the cqt
cqt = librosa.logamplitude(cqt**2, ref_power=np.max)
# Make sure the data is long enough.
# In practice this should no longer be necessary.
cqt = utils.backfill_noise(cqt, t_len + 1)
num_possible_obs = cqt.shape[-2] - t_len
if shuffle:
idx = np.arange(num_possible_obs)
rng.shuffle(idx)
else:
idx = np.arange(1)
counter = 0
while True:
if slice_logger: slice_logger.sample(record)
obs = obs_slicer(cqt, idx, counter, t_len)
if add_noise:
obs = obs + utils.same_shape_noise(obs, 1, rng)
data = dict(x_in=obs, target=np.atleast_1d((target, )))
yield data
counter += 1
if counter >= len(idx):
if not auto_restart:
break
if shuffle:
rng.shuffle(idx)
counter = 0
if slice_logger: slice_logger.close(record)
def base_slicer(
record,
t_len,
obs_slicer,
shuffle=True,
auto_restart=True,
add_noise=True,
random_seed=None,
npz_data_key="cqt",
slice_logger=None,
):
"""Base slicer function for yielding data from an .npz file
which contains the features for streaming.
Assumptions:
* Input file is about 1s in length (43 frames), but might be
shorter or longer. t_len should be 43 (~1s), however.
* It is possible that this file could get sampled more than
once consecutively, so for training, if shuffle is True,
and the number of frames in the file > t_len, it will
try to sample from frames other than the first t_len frames
randomly.
* For prediction (shuffle=False), it will always return
the first t_len frames.
To use this for training, use the following parameters:
(...shuffle=True, auto_restart=True, add_noise=True)
For prediction / eval, use the following parameters:
(...shuffle=False, auto_restart=False, add_noise=False)
Parameters
----------
record : pandas.Series
Single pandas record containing a 'cqt' record
which points to the cqt file in question.
Also must contain an "instrument" column
which contains the ground truth.
t_len : int
Length of the sliced array [in time/frames]
obs_slicer : function
Function which takes (cqt, idx, counter, t_len), and
returns a slice from the data, formatted correctly for
the desired data.
shuffle : bool
If True, shuffles the frames every time through the file.
If False, returns the first t_len frames.
auto_restart : bool
If True, yields infinitely.
If False, only goes through the file once.
add_noise : bool
If True, adds a small amount of noise to every sample.
If False, does nothing.
random_seed : int or None
If int, uses this number as the random seed. Otherwise,
makes it's own.
npz_data_key : str
The key in the npz file pointed to by record from which to
load the data. Choices = ['cqt', 'harmonic_cqt']
Yields
-------
sample : dict with fields {x_in, target}
The windowed observation.
"""
rng = np.random.RandomState(random_seed)
target = instrument_map.get_index(record["instrument"])
if slice_logger:
slice_logger.start(record, target)
if not (
("cqt" in record.index)
and record["cqt"] is not None
and isinstance(record["cqt"], str)
and os.path.exists(record["cqt"])
):
logger.error("No valid feature file specified for record: {}".format(record))
if slice_logger:
slice_logger.error(record)
return
# Load the npz file with they key specified.
try:
cqt = np.load(record["cqt"])[npz_data_key]
except zipfile.BadZipFile:
logger.error("This zip file is bad! Sadness :( - {}".format(record["cqt"]))
if slice_logger:
slice_logger.error(record)
return
# Take the logmagnitude of the cqt
cqt = librosa.logamplitude(cqt ** 2, ref_power=np.max)
# Make sure the data is long enough.
# In practice this should no longer be necessary.
cqt = utils.backfill_noise(cqt, t_len + 1)
num_possible_obs = cqt.shape[-2] - t_len
if shuffle:
idx = np.arange(num_possible_obs)
rng.shuffle(idx)
else:
idx = np.arange(1)
counter = 0
while True:
if slice_logger:
slice_logger.sample(record)
obs = obs_slicer(cqt, idx, counter, t_len)
if add_noise:
obs = obs + utils.same_shape_noise(obs, 1, rng)
data = dict(x_in=obs, target=np.atleast_1d((target,)))
yield data
counter += 1
if counter >= len(idx):
if not auto_restart:
break
if shuffle:
rng.shuffle(idx)
counter = 0
if slice_logger:
slice_logger.close(record)