Skip to content

Commit

Permalink
add note to multif0 converter (#508)
Browse files Browse the repository at this point in the history
  • Loading branch information
@rabitt
rabitt authored May 3, 2021
1 parent 79d3954 commit b31def1
Show file tree
Hide file tree
Showing 2 changed files with 264 additions and 159 deletions.
375 changes: 216 additions & 159 deletions mirdata/annotations.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -2,6 +2,7 @@
"""
import logging
import re
from typing import List, Optional, Tuple

from jams.schema import namespace
import librosa
Expand Down Expand Up @@ -144,165 +145,6 @@ def __init__(self, intervals, interval_unit, labels=None, label_unit=None):
self.label_unit = label_unit


class NoteData(Annotation):
"""NoteData class
Attributes:
intervals (np.ndarray): (n x 2) array of intervals
in the form [start_time, end_time]. Times should be positive
and intervals should have non-negative duration
interval_unit (str): unit of the time values in intervals. One
of TIME_UNITS.
pitches (np.ndarray): array of pitches
pitch_unit (str): note unit, one of PITCH_UNITS
confidence (np.ndarray or None): array of confidence values
confidence_unit (str or None): confidence unit, one of AMPLITUDE_UNITS
"""

def __init__(
self,
intervals,
interval_unit,
pitches,
pitch_unit,
confidence=None,
confidence_unit=None,
):
validate_array_like(intervals, np.ndarray, float)
validate_array_like(pitches, np.ndarray, float)
validate_array_like(confidence, np.ndarray, float, none_allowed=True)
validate_lengths_equal([intervals, pitches, confidence])
validate_intervals(intervals, interval_unit)
validate_pitches(pitches, pitch_unit)
validate_confidence(confidence, confidence_unit)

self.intervals = intervals
self.interval_unit = interval_unit
self.pitches = pitches
self.pitch_unit = pitch_unit
self.confidence = confidence
self.confidence_unit = confidence_unit

@property
def notes(self):
logging.warning(
"Deprecation warning: NoteData.notes will be removed in a future version."
+ "Use NoteData.pitches"
)
return self.pitches

def to_sparse_index(
self,
time_scale,
time_scale_unit,
frequency_scale,
frequency_scale_unit,
amplitude_unit="binary",
onsets_only=False,
):
"""Convert note annotations to indexes of a sparse matrix (piano roll)
Args:
time_scale (np.ndarray): array of matrix time stamps in seconds
time_scale_unit (str): units for time scale values, one of TIME_UNITS
frequency_scale (np.ndarray): array of matrix frequency values in seconds
frequency_scale_unit (str): units for frequency scale values, one of PITCH_UNITS
amplitude_unit (str): units for amplitude values, one of AMPLITUDE_UNITS.
Defaults to "binary".
onsets_only (bool, optional): If True, returns an onset piano roll.
Defaults to False.
Returns:
* sparse_index (np.ndarray): Array of sparce indices [(time_index, frequency_index)]
* amplitude (np.ndarray): Array of amplitude values for each index
"""
intervals = convert_time_units(
self.intervals, self.interval_unit, time_scale_unit
)
freqs_hz = convert_pitch_units(
self.pitches, self.pitch_unit, frequency_scale_unit
)

if self.confidence is not None:
confidence = convert_amplitude_units(
self.confidence, self.confidence_unit, amplitude_unit
)
else:
confidence = convert_amplitude_units(
np.ones((freqs_hz.shape)), "binary", amplitude_unit
)

time_index_0 = closest_index(
intervals[:, 0, np.newaxis], time_scale[:, np.newaxis]
)
freq_indexes = closest_index(
np.log(freqs_hz)[:, np.newaxis], np.log(frequency_scale)[:, np.newaxis]
)
if onsets_only:
onset_index = []
confidences = []
for t0, f, c in zip(time_index_0, freq_indexes, confidence):
if t0 == -1 or f == -1:
continue
onset_index.append([t0, f])
confidences.append(c)
return np.array(onset_index), np.array(confidences)

time_index_1 = closest_index(
intervals[:, 1, np.newaxis], time_scale[:, np.newaxis]
)
max_idx = len(time_scale) - 1
sparse_index = []
confidences = []
for t0, t1, f, c in zip(time_index_0, time_index_1, freq_indexes, confidence):
if f == -1 or (t0 == -1 and t1 == -1):
continue

t_start = max([t0, 0])
t_end = (t1 if t1 != -1 else max_idx) + 1

sparse_index.extend([[t, f] for t in range(t_start, t_end)])
confidences.extend([c for _ in range(t_start, t_end)])

return np.array(sparse_index), np.array(confidences)

def to_matrix(
self,
time_scale,
time_scale_unit,
frequency_scale,
frequency_scale_unit,
amplitude_unit="binary",
onsets_only=False,
):
"""Convert f0 data to a matrix (piano roll) defined by a time and frequency scale
Args:
time_scale (np.ndarray): array of matrix time stamps in seconds
time_scale_unit (str): units for time scale values, one of TIME_UNITS
frequency_scale (np.ndarray): array of matrix frequency values in seconds
frequency_scale_unit (str): units for frequency scale values, one of PITCH_UNITS
onsets_only (bool, optional): If True, returns an onset piano roll.
Defaults to False.
Returns:
np.ndarray: 2D matrix of shape len(time_scale) x len(frequency_scale)
"""
index, voicing = self.to_sparse_index(
time_scale,
time_scale_unit,
frequency_scale,
frequency_scale_unit,
amplitude_unit,
onsets_only,
)
matrix = np.zeros((len(time_scale), len(frequency_scale)))
matrix[index[:, 0], index[:, 1]] = voicing
return matrix


class ChordData(Annotation):
"""ChordData class
Expand Down Expand Up @@ -765,6 +607,221 @@ def to_matrix(
return matrix


class NoteData(Annotation):
"""NoteData class
Attributes:
intervals (np.ndarray): (n x 2) array of intervals
in the form [start_time, end_time]. Times should be positive
and intervals should have non-negative duration
interval_unit (str): unit of the time values in intervals. One
of TIME_UNITS.
pitches (np.ndarray): array of pitches
pitch_unit (str): note unit, one of PITCH_UNITS
confidence (np.ndarray or None): array of confidence values
confidence_unit (str or None): confidence unit, one of AMPLITUDE_UNITS
"""

def __init__(
self,
intervals: np.ndarray,
interval_unit: str,
pitches: np.ndarray,
pitch_unit: str,
confidence: Optional[np.ndarray] = None,
confidence_unit: Optional[str] = None,
):
validate_array_like(intervals, np.ndarray, float)
validate_array_like(pitches, np.ndarray, float)
validate_array_like(confidence, np.ndarray, float, none_allowed=True)
validate_lengths_equal([intervals, pitches, confidence])
validate_intervals(intervals, interval_unit)
validate_pitches(pitches, pitch_unit)
validate_confidence(confidence, confidence_unit)

self.intervals = intervals
self.interval_unit = interval_unit
self.pitches = pitches
self.pitch_unit = pitch_unit
self.confidence = confidence
self.confidence_unit = confidence_unit

@property
def notes(self) -> np.ndarray:
logging.warning(
"Deprecation warning: NoteData.notes will be removed in a future version."
+ "Use NoteData.pitches"
)
return self.pitches

def to_sparse_index(
self,
time_scale: np.ndarray,
time_scale_unit: str,
frequency_scale: np.ndarray,
frequency_scale_unit: str,
amplitude_unit: str = "binary",
onsets_only: bool = False,
) -> Tuple[np.ndarray, np.ndarray]:
"""Convert note annotations to indexes of a sparse matrix (piano roll)
Args:
time_scale (np.array): array of matrix time stamps in seconds
time_scale_unit (str): units for time scale values, one of TIME_UNITS
frequency_scale (np.array): array of matrix frequency values in seconds
frequency_scale_unit (str): units for frequency scale values, one of PITCH_UNITS
amplitude_unit (str): units for amplitude values, one of AMPLITUDE_UNITS.
Defaults to "binary".
onsets_only (bool, optional): If True, returns an onset piano roll.
Defaults to False.
Returns:
* sparse_index (np.ndarray): Array of sparce indices [(time_index, frequency_index)]
* amplitude (np.ndarray): Array of amplitude values for each index
"""
intervals = convert_time_units(
self.intervals, self.interval_unit, time_scale_unit
)
freqs_hz = convert_pitch_units(
self.pitches, self.pitch_unit, frequency_scale_unit
)

if self.confidence is not None:
confidence = convert_amplitude_units(
self.confidence, self.confidence_unit, amplitude_unit
)
else:
confidence = convert_amplitude_units(
np.ones((freqs_hz.shape)), "binary", amplitude_unit
)

time_index_0 = closest_index(
intervals[:, 0, np.newaxis], time_scale[:, np.newaxis]
)
freq_indexes = closest_index(
np.log(freqs_hz)[:, np.newaxis], np.log(frequency_scale)[:, np.newaxis]
)
if onsets_only:
onset_index = []
confidences = []
for t0, f, c in zip(time_index_0, freq_indexes, confidence):
if t0 == -1 or f == -1:
continue
onset_index.append([t0, f])
confidences.append(c)
return np.array(onset_index), np.array(confidences)

time_index_1 = closest_index(
intervals[:, 1, np.newaxis], time_scale[:, np.newaxis]
)
max_idx = len(time_scale) - 1
sparse_index = []
confidences = []
for t0, t1, f, c in zip(time_index_0, time_index_1, freq_indexes, confidence):
if f == -1 or (t0 == -1 and t1 == -1):
continue

t_start = max([t0, 0])
t_end = (t1 if t1 != -1 else max_idx) + 1

sparse_index.extend([[t, f] for t in range(t_start, t_end)])
confidences.extend([c for _ in range(t_start, t_end)])

return np.array(sparse_index), np.array(confidences)

def to_matrix(
self,
time_scale: np.ndarray,
time_scale_unit: str,
frequency_scale: np.ndarray,
frequency_scale_unit: str,
amplitude_unit: str = "binary",
onsets_only: bool = False,
) -> np.ndarray:
"""Convert f0 data to a matrix (piano roll) defined by a time and frequency scale
Args:
time_scale (np.ndarray): array of matrix time stamps in seconds
time_scale_unit (str): units for time scale values, one of TIME_UNITS
frequency_scale (np.ndarray): array of matrix frequency values in seconds
frequency_scale_unit (str): units for frequency scale values, one of PITCH_UNITS
onsets_only (bool, optional): If True, returns an onset piano roll.
Defaults to False.
Returns:
np.ndarray: 2D matrix of shape len(time_scale) x len(frequency_scale)
"""
index, voicing = self.to_sparse_index(
time_scale,
time_scale_unit,
frequency_scale,
frequency_scale_unit,
amplitude_unit,
onsets_only,
)
matrix = np.zeros((len(time_scale), len(frequency_scale)))
matrix[index[:, 0], index[:, 1]] = voicing
return matrix

def to_multif0(
self, time_hop: float, time_hop_unit: str, max_time: Optional[float] = None
) -> MultiF0Data:
"""Convert note annotation to multiple f0 format.
Args:
time_hop (float): time between time stamps in multif0 annotation
time_hop_unit (str): unit for time_hop, and resulting multif0 data.
One of TIME_UNITS
max_time (float, optional): Maximum time stamp in time_hop units.
Defaults to None, in which case the maximum note interval
time is used.
Returns:
MultiF0Data: multif0 annotation
"""
intervals = convert_time_units(
self.intervals, self.interval_unit, time_hop_unit
)
note_time_max = np.max(intervals[:, 1])
max_time = note_time_max if not max_time else max_time
if max_time < note_time_max:
raise ValueError(
"max_time = {} cannot be smaller than the last note interval = {}".format(
max_time, note_time_max
)
)
times = np.arange(0, max_time + time_hop, time_hop)
frequency_list: List[List[float]] = [[] for _ in times]
confidence_list: List[List[float]] = [[] for _ in times]
if self.confidence is not None:

for t0, t1, pch, conf in zip(
intervals[:, 0], intervals[:, 1], self.pitches, self.confidence
):
for i in range(
int(np.round(t0 / time_hop)), int(np.round(t1 / time_hop)) + 1
):
frequency_list[i].append(pch)
confidence_list[i].append(conf)
else:
for t0, t1, pch in zip(intervals[:, 0], intervals[:, 1], self.pitches):
for i in range(
int(np.round(t0 / time_hop)), int(np.round(t1 / time_hop)) + 1
):
frequency_list[i].append(pch)

return MultiF0Data(
times,
time_hop_unit,
frequency_list,
self.pitch_unit,
None if self.confidence is None else confidence_list,
self.confidence_unit,
)


class KeyData(Annotation):
"""KeyData class
Expand Down
Loading

0 comments on commit b31def1

Please sign in to comment.