def align(track1, track2, mat1, mat2):
""" Constrained search between a settled section and a new section.
Outputs location in mat2 and the number of rows used in the transition.
"""
# Get the average marker duration.
marker1 = average_duration(getattr(track1.analysis, track1.resampled['rate'])[track1.resampled['index']:track1.resampled['index']+rows(mat1)])
marker2 = average_duration(getattr(track2.analysis, track2.resampled['rate'])[track2.resampled['index']:track2.resampled['index']+rows(mat2)])
def get_adjustment(tr1, tr2):
"""Update tatum rate if necessary"""
dist = np.log2(tr1 / tr2)
if dist < -0.5: return (1, 2)
elif dist > 0.5: return (2, 1)
else: return (1, 1)
rate1, rate2 = get_adjustment(marker1, marker2)
if rate1 == 2: mat1 = upsample_matrix(mat1)
if rate2 == 2: mat2 = upsample_matrix(mat2)
# Update sizes.
rows2 = rows(mat2)
rows1 = min( rows(mat1), max(rows2 - MIN_SEARCH, MIN_MARKERS)) # at least the best of MIN_SEARCH choices
# Search for minimum.
def dist(i):
return evaluate_distance(mat1[0:rows1,:], mat2[i:i+rows1,:])
min_loc = min(xrange(rows2 - rows1), key=dist)
min_val = dist(min_loc)
# Let's make sure track2 ends its transition on a regular tatum.
if rate2 == 2 and (min_loc + rows1) & 1:
rows1 -= 1
return min_loc, rows1, rate1, rate2
def timbre_whiten(mat):
if rows(mat) < 2: return mat
m = np.zeros((rows(mat), 12), dtype=np.float32)
m[:, 0] = mat[:, 0] - np.mean(mat[:, 0], 0)
m[:, 0] = m[:, 0] / np.std(m[:, 0], 0)
m[:, 1:] = mat[:, 1:] - np.mean(mat[:, 1:].flatten(), 0)
m[:, 1:] = m[:, 1:] / np.std(m[:, 1:].flatten(), 0) # use this!
return m
def timbre_whiten(mat):
if rows(mat) < 2: return mat
m = np.zeros((rows(mat), 12), dtype=np.float32)
m[:,0] = mat[:,0] - np.mean(mat[:,0],0)
m[:,0] = m[:,0] / np.std(m[:,0],0)
m[:,1:] = mat[:,1:] - np.mean(mat[:,1:].flatten(),0)
m[:,1:] = m[:,1:] / np.std(m[:,1:].flatten(),0) # use this!
return m
def get_paths_slow(matrix, size=MIN_RANGE):
paths = []
for i in xrange(rows(matrix)-MIN_ALIGN+1):
vector = np.zeros((rows(matrix)-MIN_ALIGN+1,), dtype=np.float32)
for j in xrange(rows(matrix)-MIN_ALIGN+1):
vector[j] = evaluate_distance(matrix[i:i+MIN_ALIGN,:], matrix[j:j+MIN_ALIGN,:])
paths.append(get_loop_points(vector, size))
return paths
def get_paths_slow(matrix, size=MIN_RANGE):
paths = []
for i in xrange(rows(matrix)-MIN_ALIGN+1):
vector = np.zeros((rows(matrix)-MIN_ALIGN+1,), dtype=np.float32)
for j in xrange(rows(matrix)-MIN_ALIGN+1):
vector[j] = evaluate_distance(matrix[i:i+MIN_ALIGN,:], matrix[j:j+MIN_ALIGN,:])
paths.append(get_loop_points(vector, size))
return paths
def initialize(track, inter, transition):
"""find initial cursor location"""
mat = track.resampled['matrix']
markers = getattr(track.analysis, track.resampled['rate'])
try:
# compute duration of matrix
mat_dur = markers[track.resampled['index'] + rows(
mat)].start - markers[track.resampled['index']].start
start = (mat_dur - inter - transition - FADE_IN) / 2
dur = start + FADE_IN + inter
# move cursor to transition marker
duration, track.resampled['cursor'] = move_cursor(track, dur, 0)
# work backwards to find the exact locations of initial fade in and playback sections
fi = Fadein(
track,
markers[track.resampled['index'] + track.resampled['cursor']].start
- inter - FADE_IN, FADE_IN)
pb = Playback(
track,
markers[track.resampled['index'] + track.resampled['cursor']].start
- inter, inter)
except:
track.resampled['cursor'] = FADE_IN + inter
fi = Fadein(track, 0, FADE_IN)
pb = Playback(track, FADE_IN, inter)
return [fi, pb]
def get_mat_in(track, transition, inter):
""" Find and output the search matrix to use in the next alignment.
Assumes that track.resampled exists.
"""
# search from the start
cursor = 0
track.resampled['cursor'] = cursor
mat = track.resampled['matrix']
# compute search zone by anticipating what's playing after the transition
marker_end = getattr(track.analysis,
track.resampled['rate'])[track.resampled['index'] +
rows(mat)].start
marker_start = getattr(
track.analysis,
track.resampled['rate'])[track.resampled['index']].start
search_dur = (marker_end - marker_start) - inter - 2 * transition
if search_dur < 0:
return mat[:MIN_MARKERS, :]
# find what the location is in rows
duration, cursor = move_cursor(track, search_dur, cursor)
return mat[:cursor, :]
def move_cursor(track, duration, cursor, buf=MIN_MARKERS):
dur = 0
while dur < duration and cursor < rows(track.resampled['matrix']) - buf:
markers = getattr(track.analysis, track.resampled['rate'])
dur += markers[track.resampled['index'] + cursor].duration
cursor += 1
return dur, cursor
def move_cursor(track, duration, cursor, buf=MIN_MARKERS):
dur = 0
while dur < duration and cursor < rows(track.resampled['matrix']) - buf:
markers = getattr(track.analysis, track.resampled['rate'])
dur += markers[track.resampled['index'] + cursor].duration
cursor += 1
return dur, cursor
def analyze(track):
timbre = resample_features(track, rate=RATE, feature='timbre')
timbre['matrix'] = timbre_whiten(timbre['matrix'])
pitch = resample_features(track, rate=RATE, feature='pitches')
#NOTE why not whiten pitch matrix?
# pick a tradeoff between speed and memory size
if rows(timbre['matrix']) < MAX_SIZE:
# faster but memory hungry. For euclidean distances only.
t_paths = get_paths(timbre['matrix'])
p_paths = get_paths(pitch['matrix'])
else:
# slower but memory efficient. Any distance possible.
t_paths = get_paths_slow(timbre['matrix'])
p_paths = get_paths_slow(pitch['matrix'])
# intersection of top timbre and pitch results
paths = path_intersect(t_paths, p_paths)
markers = getattr(track.analysis, timbre['rate'])[timbre['index']:timbre['index']+len(paths)]
graph = make_graph(paths, markers, timbre['matrix'], pitch['matrix'])
#NOTE remove last node because empty?
size = graph.number_of_nodes()
graph.remove_node(size-1)
return graph
def initialize(track, inter, transition):
"""find initial cursor location"""
mat = track.resampled['matrix']
markers = getattr(track.analysis, track.resampled['rate'])
try:
# compute duration of matrix
mat_dur = markers[track.resampled['index'] + rows(mat)].start - markers[track.resampled['index']].start
start = (mat_dur - inter - transition - FADE_IN) / 2
dur = start + FADE_IN + inter
# move cursor to transition marker
duration, track.resampled['cursor'] = move_cursor(track, dur, 0)
# work backwards to find the exact locations of initial fade in and playback sections
fi = Fadein(track, markers[track.resampled['index'] + track.resampled['cursor']].start - inter - FADE_IN, FADE_IN)
pb = Playback(track, markers[track.resampled['index'] + track.resampled['cursor']].start - inter, inter)
except:
track.resampled['cursor'] = FADE_IN + inter
fi = Fadein(track, 0, FADE_IN)
pb = Playback(track, FADE_IN, inter)
return [fi, pb]
def get_mat_in(track, transition, inter):
""" Find and output the search matrix to use in the next alignment.
Assumes that track.resampled exists.
"""
# search from the start
cursor = 0
track.resampled['cursor'] = cursor
mat = track.resampled['matrix']
# compute search zone by anticipating what's playing after the transition
marker_end = getattr(track.analysis, track.resampled['rate'])[track.resampled['index'] + rows(mat)].start
marker_start = getattr(track.analysis, track.resampled['rate'])[track.resampled['index']].start
search_dur = (marker_end - marker_start) - inter - 2 * transition
if search_dur < 0:
return mat[:MIN_MARKERS,:]
# find what the location is in rows
duration, cursor = move_cursor(track, search_dur, cursor)
return mat[:cursor,:]
def get_paths(matrix, size=MIN_RANGE):
mat = make_similarity_matrix(matrix, size=MIN_ALIGN)
paths = []
for i in xrange(rows(mat)):
paths.append(get_loop_points(mat[i,:], size))
return paths
def do_work(track, options):
dur = float(options.duration)
mlp = int(options.minimum)
lgh = bool(options.length)
inf = bool(options.infinite)
pkl = bool(options.pickle)
gml = bool(options.graph)
plt = bool(options.plot)
fce = bool(options.force)
sho = bool(options.shortest)
lon = bool(options.longest)
vbs = bool(options.verbose)
mp3 = track.filename
try:
if fce == True:
raise
graph = read_graph(mp3+".graph.gpkl")
except:
# compute resampled and normalized matrix
timbre = resample_features(track, rate=RATE, feature='timbre')
timbre['matrix'] = timbre_whiten(timbre['matrix'])
pitch = resample_features(track, rate=RATE, feature='pitches')
# pick a tradeoff between speed and memory size
if rows(timbre['matrix']) < MAX_SIZE:
# faster but memory hungry. For euclidean distances only.
t_paths = get_paths(timbre['matrix'])
p_paths = get_paths(pitch['matrix'])
else:
# slower but memory efficient. Any distance possible.
t_paths = get_paths_slow(timbre['matrix'])
p_paths = get_paths_slow(pitch['matrix'])
# intersection of top timbre and pitch results
paths = path_intersect(t_paths, p_paths)
# TEMPORARY -- check that the data looks good
if vbs == True:
print_screen(paths)
# make graph
markers = getattr(track.analysis, timbre['rate'])[timbre['index']:timbre['index']+len(paths)]
graph = make_graph(paths, markers)
# remove smaller loops for quality results
if 0 < mlp:
remove_short_loops(graph, mlp)
# plot graph
if plt == True:
save_plot(graph, mp3+".graph.png")
# save graph
if pkl == True:
save_graph(graph, mp3+".graph.gpkl")
if gml == True:
save_graph(graph, mp3+".graph.gml")
# single loops
if sho == True:
return one_loop(graph, track, mode='shortest')
if lon == True:
return one_loop(graph, track, mode='longest')
# other infinite loops
if inf == True:
if vbs == True:
print "\nInput Duration:", track.analysis.duration
# get the optimal path for a given duration
return infinite(graph, track, dur)
dur_intro = min(graph.nodes())
dur_outro = track.analysis.duration - max(graph.nodes())
if vbs == True:
print "Input Duration:", track.analysis.duration
# get the optimal path for a given duration
path = compute_path(graph, max(dur-dur_intro-dur_outro, 0))
# build actions
middle = make_jumps(path, track)
# complete list of actions
actions = terminate(dur_intro, middle, dur_outro, dur, lgh)
return actions
def do_work(track, options):
dur = float(options.duration)
mlp = int(options.minimum)
lgh = bool(options.length)
inf = bool(options.infinite)
pkl = bool(options.pickle)
gml = bool(options.graph)
plt = bool(options.plot)
fce = bool(options.force)
sho = bool(options.shortest)
lon = bool(options.longest)
vbs = bool(options.verbose)
mp3 = track.filename
try:
if fce == True:
raise
graph = read_graph(mp3+".graph.gpkl")
except:
# compute resampled and normalized matrix
timbre = resample_features(track, rate=RATE, feature='timbre')
timbre['matrix'] = timbre_whiten(timbre['matrix'])
pitch = resample_features(track, rate=RATE, feature='pitches')
# pick a tradeoff between speed and memory size
if rows(timbre['matrix']) < MAX_SIZE:
# faster but memory hungry. For euclidean distances only.
t_paths = get_paths(timbre['matrix'])
p_paths = get_paths(pitch['matrix'])
else:
# slower but memory efficient. Any distance possible.
t_paths = get_paths_slow(timbre['matrix'])
p_paths = get_paths_slow(pitch['matrix'])
# intersection of top timbre and pitch results
paths = path_intersect(t_paths, p_paths)
# TEMPORARY -- check that the data looks good
if vbs == True:
print_screen(paths)
# make graph
markers = getattr(track.analysis, timbre['rate'])[timbre['index']:timbre['index']+len(paths)]
graph = make_graph(paths, markers)
# remove smaller loops for quality results
if 0 < mlp:
remove_short_loops(graph, mlp)
# plot graph
if plt == True:
save_plot(graph, mp3+".graph.png")
# save graph
if pkl == True:
save_graph(graph, mp3+".graph.gpkl")
if gml == True:
save_graph(graph, mp3+".graph.gml")
# single loops
if sho == True:
return one_loop(graph, track, mode='shortest')
if lon == True:
return one_loop(graph, track, mode='longest')
# other infinite loops
if inf == True:
if vbs == True:
print "\nInput Duration:", track.analysis.duration
# get the optimal path for a given duration
return infinite(graph, track, dur)
dur_intro = min(graph.nodes())
dur_outro = track.analysis.duration - max(graph.nodes())
if vbs == True:
print "Input Duration:", track.analysis.duration
# get the optimal path for a given duration
path = compute_path(graph, max(dur-dur_intro-dur_outro, 0))
# build actions
middle = make_jumps(path, track)
# complete list of actions
actions = terminate(dur_intro, middle, dur_outro, dur, lgh)
return actions
def get_paths(matrix, size=MIN_RANGE):
mat = make_similarity_matrix(matrix, size=MIN_ALIGN)
paths = []
for i in xrange(rows(mat)):
paths.append(get_loop_points(mat[i,:], size))
return paths
