#!/usr/bin/env python
# -*- encoding: utf-8 -*-
'''Tag task transformers'''
import numpy as np
import jams
from librosa import time_to_frames
from librosa.sequence import transition_loop
from .base import BaseTaskTransformer
from ..exceptions import ParameterError
from ..labels import MultiLabelBinarizer
__all__ = ['DynamicLabelTransformer', 'StaticLabelTransformer']
class DynamicLabelTransformer(BaseTaskTransformer):
'''Time-series label transformer.
Attributes
----------
name : str
The name of this transformer object
namespace : str
The JAMS namespace for this task
labels : list of str [optional]
The list of labels for this task.
If not provided, it will attempt to infer the label set from the
namespace definition.
sr : number > 0
The audio sampling rate
hop_length : int > 0
The hop length for annotation frames
p_self : None, float in (0, 1), or np.ndarray [shape=(n_labels,)]
Optional self-loop probability(ies), used for Viterbi decoding
p_state : None or np.ndarray [shape=(n_labels,)]
Optional marginal probability for each class
p_init : None or np.ndarray [shape=(n_labels,)]
Optional initial probability for each class
See Also
--------
StaticLabelTransformer
'''
def __init__(self, name, namespace, labels=None, sr=22050, hop_length=512,
p_self=None, p_init=None, p_state=None):
super(DynamicLabelTransformer, self).__init__(name=name,
namespace=namespace,
sr=sr,
hop_length=hop_length)
if labels is None:
labels = jams.schema.values(namespace)
self.encoder = MultiLabelBinarizer()
self.encoder.fit([labels])
self._classes = set(self.encoder.classes_)
self.set_transition(p_self)
if p_init is not None:
if len(p_init) != len(self._classes):
raise ParameterError('Invalid p_init.shape={} for vocabulary size={}'.format(p_init.shape, len(self._classes)))
self.p_init = p_init
if p_state is not None:
if len(p_state) != len(self._classes):
raise ParameterError('Invalid p_state.shape={} for vocabulary size={}'.format(p_state.shape, len(self._classes)))
self.p_state = p_state
self.register('tags', [None, len(self._classes)], np.bool)
def set_transition(self, p_self):
'''Set the transition matrix according to self-loop probabilities.
Parameters
----------
p_self : None, float in (0, 1), or np.ndarray [shape=(n_labels,)]
Optional self-loop probability(ies), used for Viterbi decoding
'''
if p_self is None:
self.transition = None
else:
self.transition = np.empty((len(self._classes), 2, 2))
if np.isscalar(p_self):
self.transition = transition_loop(2, p_self)
elif len(p_self) != len(self._classes):
raise ParameterError('Invalid p_self.shape={} for vocabulary size={}'.format(p_self.shape, len(self._classes)))
else:
for i in range(len(self._classes)):
self.transition[i] = transition_loop(2, p_self[i])
def empty(self, duration):
'''Empty label annotations.
Constructs a single observation with an empty value (None).
Parameters
----------
duration : number > 0
The duration of the annotation
'''
ann = super(DynamicLabelTransformer, self).empty(duration)
ann.append(time=0, duration=duration, value=None)
return ann
def transform_annotation(self, ann, duration):
'''Transform an annotation to dynamic label encoding.
Parameters
----------
ann : jams.Annotation
The annotation to convert
duration : number > 0
The duration of the track
Returns
-------
data : dict
data['tags'] : np.ndarray, shape=(n, n_labels)
A time-varying binary encoding of the labels
'''
intervals, values = ann.to_interval_values()
# Suppress all intervals not in the encoder
tags = []
for v in values:
if v in self._classes:
tags.extend(self.encoder.transform([[v]]))
else:
tags.extend(self.encoder.transform([[]]))
tags = np.asarray(tags)
target = self.encode_intervals(duration, intervals, tags)
return {'tags': target}
def inverse(self, encoded, duration=None):
'''Inverse transformation'''
ann = jams.Annotation(namespace=self.namespace, duration=duration)
for start, end, value in self.decode_intervals(encoded,
duration=duration,
transition=self.transition,
p_init=self.p_init,
p_state=self.p_state):
# Map start:end to frames
f_start, f_end = time_to_frames([start, end],
sr=self.sr,
hop_length=self.hop_length)
confidence = np.mean(encoded[f_start:f_end+1, value])
value_dec = self.encoder.inverse_transform(np.atleast_2d(value))[0]
for vd in value_dec:
ann.append(time=start,
duration=end-start,
value=vd,
confidence=confidence)
return ann
class StaticLabelTransformer(BaseTaskTransformer):
'''Static label transformer.
Attributes
----------
name : str
The name of this transformer object
namespace : str
The JAMS namespace for this task
labels : list of str [optional]
The list of labels for this task.
If not provided, it will attempt to infer the label set from the
namespace definition.
See Also
--------
DynamicLabelTransformer
'''
def __init__(self, name, namespace, labels=None):
super(StaticLabelTransformer, self).__init__(name=name,
namespace=namespace,
sr=1, hop_length=1)
if labels is None:
labels = jams.schema.values(namespace)
self.encoder = MultiLabelBinarizer()
self.encoder.fit([labels])
self._classes = set(self.encoder.classes_)
self.register('tags', [len(self._classes)], np.bool)
def transform_annotation(self, ann, duration):
'''Transform an annotation to static label encoding.
Parameters
----------
ann : jams.Annotation
The annotation to convert
duration : number > 0
The duration of the track
Returns
-------
data : dict
data['tags'] : np.ndarray, shape=(n_labels,)
A static binary encoding of the labels
'''
intervals = np.asarray([[0, 1]])
values = list([obs.value for obs in ann])
intervals = np.tile(intervals, [len(values), 1])
# Suppress all intervals not in the encoder
tags = [v for v in values if v in self._classes]
if len(tags):
target = self.encoder.transform([tags]).astype(np.bool).max(axis=0)
else:
target = np.zeros(len(self._classes), dtype=np.bool)
return {'tags': target}
def inverse(self, encoded, duration=None):
'''Inverse static tag transformation'''
ann = jams.Annotation(namespace=self.namespace, duration=duration)
if np.isrealobj(encoded):
detected = (encoded >= 0.5)
else:
detected = encoded
for vd in self.encoder.inverse_transform(np.atleast_2d(detected))[0]:
vid = np.flatnonzero(self.encoder.transform(np.atleast_2d(vd)))
ann.append(time=0,
duration=duration,
value=vd,
confidence=encoded[vid])
return ann
