def __init__(self, vcost=0., hcost=0., dcost=0.,):
super(BaseTranscriptionAlignment, self).__init__()
self._dtw = DynamicTimeWarping(
vcost=vcost, hcost=hcost, dcost=dcost,
no_vertical=self.no_vertical, no_horizontal=self.no_horizontal)
class BaseTranscriptionAlignment(object):
"""Base transcriptions alignment algorithm
This algorithm will temporally align two transcriptions
(called `vertical` and `horizontal` following standard DTW representation).
* ────────────────> horizontal
│ *
│ *
│ * * *
│ *
│ *
V *
vertical
Whenever two drifting times are merged, it will keep the `vertical` one.
Parameters
----------
vattribute : str
hattribute : str
"""
@property
def no_vertical(self):
return False
@property
def no_horizontal(self):
return False
def __init__(self, vcost=0., hcost=0., dcost=0.,):
super(BaseTranscriptionAlignment, self).__init__()
self._dtw = DynamicTimeWarping(
vcost=vcost, hcost=hcost, dcost=dcost,
no_vertical=self.no_vertical, no_horizontal=self.no_horizontal)
def pairwise_distance(self, vsequence, hsequence):
"""Compute pairwise distance matrix
Parameters
----------
vsequence : (index, data) iterable
hsequence : (index, data) iterable
Returns
-------
distance : numpy array
Shape = len(vsequence) x len(hsequence)
"""
raise NotImplementedError('')
def merge(self, vtranscription, htranscription, alignment):
"""Merge transcriptions based on their alignment
Parameters
----------
vtranscription, htranscription : Transcription
alignment : dict
Returns
-------
merged : Transcription
"""
# pre-process alignment to prevent this kind of situations:
# `A` and `a` are merged into `a`
# `A` and `b` are merged into `b` -- but `A` no longer exists as it is
# now called `a`!!!
# starts by creating a graph where an edge between two nodes
# indicate that the corresponding times will be eventually merged
# (we also keep track of the origin of the the nodes for later use)
same = nx.Graph()
for (v, h), status in alignment.iteritems():
if (status & STARTS_WITH == STARTS_WITH):
same.add_edge(('v', v[0]), ('h', h[0]))
if (status & ENDS_WITH == ENDS_WITH):
same.add_edge(('v', v[1]), ('h', h[1]))
# create a mapping
mapping = {}
for component in nx.connected_components(same):
anchored = [t for _, t in component if t.anchored]
if anchored:
to = anchored[0]
else:
vdrifting = [t for origin, t in component if origin == 'v']
to = vdrifting[0]
for t in component:
mapping[t] = to
# initialize new "union" transcription
merged = vtranscription.copy()
merged.add_edges_from(htranscription.edges(data=True))
# starts by merging nodes
for (origin, t), to in mapping.iteritems():
if t != to:
# according to Transcription.align documentation,
# if both t and to are drifting, the resulting graph
# will only contain `to` (from vertical sequence)
merged.align(t, to)
for (v, h), status in alignment.iteritems():
# connect start times in correct order
if status & STARTS_WITH != STARTS_WITH:
if status & STARTS_BEFORE == STARTS_BEFORE:
merged.add_edge(mapping.get(('v', v[0]), v[0]),
mapping.get(('h', h[0]), h[0]))
if status & STARTS_AFTER == STARTS_AFTER:
merged.add_edge(mapping.get(('h', h[0]), h[0]),
mapping.get(('v', v[0]), v[0]))
# connect end times in correct order
if status & ENDS_WITH != ENDS_WITH:
if status & ENDS_BEFORE == ENDS_BEFORE:
merged.add_edge(mapping.get(('v', v[1]), v[1]),
mapping.get(('h', h[1]), h[1]))
if status & ENDS_AFTER == ENDS_AFTER:
merged.add_edge(mapping.get(('h', h[1]), h[1]),
mapping.get(('v', v[1]), v[1]))
# remove self loops which might have resulted from various merging
for t in merged:
if merged.has_edge(t, t):
merged.remove_edge(t, t)
return merged
def _get_sequence(self, transcription, attribute=None):
"""Get raw sequence of attribute values
Parameters
----------
transcription : `Transcription`
attribute : str, optional
When `attribute` is not provided and there are only one attribute
on the first edge encountered, will automatically select this one
attribute from all other edges.
Returns
-------
sequence : list
Chronologically sorted list of ((start_t, end_t), value)
where `value` is the value of attribute `attribute` on the edge
between `start_t` and `end_t`.
"""
sequence = []
for s, e, data in transcription.ordered_edges_iter(data=True):
# go to next edge if data is empty
if not data:
continue
# if attribute is not provided and data has only one attribute
# we choose to use this attribute and make sure the same
# attribute will be used for the rest of loop.
# if there is more than one, then raise an error
if attribute is None:
if len(data) > 1:
msg = 'Which attribute should I use for alignment: {%s}?'
raise ValueError(msg % ', '.join(data))
attribute, item = dict(data).popitem()
sequence.append(((s, e), item))
# if attribute is provided (or was chosen automatically above)
# and data contains it, append the item at the end of the sequence
elif attribute in data:
item = data[attribute]
sequence.append(((s, e), item))
return sequence
def __call__(self, vtranscription, htranscription,
vattribute=None, hattribute=None):
"""Align two transcriptions
Parameters
----------
vtranscription, htranscription : `Transcription`
vattribute, hattribute : str
Returns
-------
merged : `Transcription`
"""
# make sure transcriptions do not share any drifting labels
# and also keep track of `vertical` mapping so that we can
# retrieve original `vertical` drifting times at the end
T.reset()
vtranscription, vmapping = vtranscription.relabel_drifting_nodes()
htranscription, _ = htranscription.relabel_drifting_nodes()
# compute distance matrix
vsequence = self._get_sequence(vtranscription, attribute=vattribute)
hsequence = self._get_sequence(htranscription, attribute=hattribute)
distance = self.pairwise_distance(vsequence, hsequence)
# align and merge
vindex, _ = itertools.izip(*vsequence)
hindex, _ = itertools.izip(*hsequence)
alignment = self._dtw.get_alignment(vindex, hindex, distance=distance)
merged = self.merge(vtranscription, htranscription, alignment)
# retrieve original `vertical` drifting times
# in case they have not been anchored
relabeled, _ = merged.relabel_drifting_nodes(vmapping)
return relabeled