Skip to content

Commit

Permalink
Various hacks to deal with the need to search very deep into raw_matc…
Browse files Browse the repository at this point in the history 
…h ordered list
  • Loading branch information
@dpwe
dpwe committed Oct 26, 2014
1 parent cc7f132 commit ec42a4b
Show file tree
Hide file tree
Showing 4 changed files with 89 additions and 70 deletions.
2 changes: 1 addition & 1 deletion audfprint.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -249,7 +249,7 @@ def setup_matcher(args):
"""Create a new matcher objects, set parameters from docopt structure"""
matcher = audfprint_match.Matcher()
matcher.window = int(args['--match-win'])
matcher.threshold = int(args['--min-count'])
matcher.threshcount = int(args['--min-count'])
matcher.max_returns = int(args['--max-matches'])
matcher.search_depth = int(args['--search-depth'])
matcher.sort_by_time = args['--sortbytime']
Expand Down
131 changes: 70 additions & 61 deletions audfprint_match.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -14,21 +14,15 @@
import audfprint_analyze
import matplotlib.pyplot as plt

from scipy import stats

def log(message):
""" log info with stats """
print time.ctime(), \
"physmem=", resource.getrusage(resource.RUSAGE_SELF).ru_maxrss, \
"utime=", resource.getrusage(resource.RUSAGE_SELF).ru_utime, \
message

def find_mode(data, window=0):
""" Find the (mode, count) of a set of data
including a tolerance window +/- window if > 0
"""
vals = np.unique(data)
counts = [len([x for x in data if abs(x-val) <= window]) for val in vals]
bestix = np.argmax(counts)
return (vals[bestix], counts[bestix])
#print time.ctime(), \
# "physmem=", resource.getrusage(resource.RUSAGE_SELF).ru_maxrss, \
# "utime=", resource.getrusage(resource.RUSAGE_SELF).ru_utime, \
# message
pass

def locmax(vec, indices=False):
""" Return a boolean vector of which points in vec are local maxima.
Expand All @@ -53,19 +47,14 @@ def find_modes(data, threshold=5, window=0):
pairs for every mode greater than or equal to threshold.
Only local maxima in counts are returned.
"""
vals = np.unique(data)
#counts = [len([x for x in data if abs(x-val) <= window]) for val in vals]
#counts = np.array([np.sum(np.abs(data - val) <= window) for val in vals])
counts = np.sum(np.abs(np.subtract.outer(vals, data)) <= window, axis=1)
# Put them into an actual vector
minval = min(vals)
fullvector = np.zeros(max(vals-minval)+1)
fullvector[vals-minval] = counts
# TODO: Ignores window at present
datamin = np.amin(data)
fullvector = np.bincount(data - datamin)
# Find local maxima
localmaxes = np.nonzero(locmax(fullvector) &
(fullvector >= threshold))[0].tolist()
return [(localmax+minval, fullvector[localmax]) for localmax in localmaxes]

localmaxes = np.nonzero(np.logical_and(locmax(fullvector),
np.greater_equal(fullvector,
threshold)))[0]
return localmaxes + datamin, fullvector[localmaxes]

class Matcher(object):
"""Provide matching for audfprint fingerprint queries to hash table"""
Expand All @@ -86,6 +75,8 @@ def __init__(self):
self.verbose = False
# Do illustration?
self.illustrate = False
# Careful counts?
self.exact_count = False

def match_hashes(self, ht, hashes, hashesfor=None):
""" Match audio against fingerprint hash table.
Expand All @@ -94,9 +85,9 @@ def match_hashes(self, ht, hashes, hashesfor=None):
hit (0=top hit).
"""
# find the implicated id, time pairs from hash table
#log("nhashes=%d" % np.shape(hashes)[0])
log("nhashes=%d" % np.shape(hashes)[0])
hits = ht.get_hits(hashes)
#log("nhits=%d" % np.shape(hits)[0])
log("nhits=%d" % np.shape(hits)[0])
## Sorted list of all the track ids that got hits
#idlist = np.r_[-1, sorted([id for id, time, hash, otime in hits]), -1]
## Counts of unique entries in the sorted list
Expand All @@ -123,48 +114,66 @@ def match_hashes(self, ht, hashes, hashesfor=None):
#counts = np.sum(np.equal.outer(ids, allids), axis=1)
# much faster, and doesn't explode memory
counts = np.bincount(allids)[ids]
#log("max(counts)=%d" % np.amax(counts))
log("max(counts)=%d" % np.amax(counts))

# Find all the actual hits for a the most popular ids
bestcountsids = sorted(zip(counts, ids), reverse=True)
maxdepth = np.minimum(np.count_nonzero(np.greater(counts, self.threshcount)), self.search_depth)
# Try the top N results
results = []
for rawcount, tid in bestcountsids[:self.search_depth]:
#modescounts = find_modes([time for (id, time, hash, otime) in hits
# if id == tid],
# window=window, threshold=threshcount)
modescounts = find_modes(alltimes[np.nonzero(allids == tid)[0]],
window=self.window,
threshold=self.threshcount)
for (mode, filtcount) in modescounts:
#matchhashes = [((otime), hash)
# for (id, time, hash, otime) in hits
# if id == tid and abs(time - mode) <= window]
## matchhashes may include repeats because multiple
## ref hashes may match a single query hash under window.
## Uniqify:
#matchhashes = sorted(list(set(matchhashes)))
matchix = np.nonzero((allids == tid) &
(np.abs(alltimes-mode) <= self.window))[0]
matchhasheshash = np.unique(allotimes[matchix]
+ maxotime*allhashes[matchix])
matchhashes = [(hash_ % maxotime, hash_ / maxotime)
for hash_ in matchhasheshash]
# much, much faster
filtcount = len(matchhashes)
results.append((tid, filtcount, mode, rawcount, matchhashes))
if not self.exact_count:
mintime = np.amin(alltimes)
alltimes -= mintime
for rawcount, tid in bestcountsids[:maxdepth]:
tidtimes = alltimes[allids==tid]
if np.amax(np.bincount(tidtimes)) <= self.threshcount:
continue
#mode, count = stats.mode(tidtimes)
#mode = int(mode[0])
modes, counts = find_modes(tidtimes, self.threshcount)
if len(modes):
mode = modes[np.nonzero(np.equal(counts,
np.amax(counts)))[0][0]]
count = np.count_nonzero(np.less_equal(
np.abs(tidtimes - mode), self.window))
log("tid %d raw %d count %d" % (tid, rawcount, count))
if count > self.threshcount:
results.append((tid, count, mode+mintime, rawcount))
if count > rawcount/4:
break
else:
for rawcount, tid in bestcountsids[:maxdepth]:
modes, counts = find_modes(alltimes[np.nonzero(allids == tid)[0]],
window=self.window,
threshold=self.threshcount)
for (mode, filtcount) in zip(modes, counts):
# matchhashes may include repeats because multiple
# ref hashes may match a single query hash under window.
# Uniqify:
#matchhashes = sorted(list(set(matchhashes)))
matchix = np.nonzero((allids == tid) &
(np.abs(alltimes-mode) <= self.window))[0]
matchhasheshash = np.unique(allotimes[matchix]
+ maxotime*allhashes[matchix])
matchhashes = [(hash_ % maxotime, hash_ / maxotime)
for hash_ in matchhasheshash]
# much, much faster
filtcount = len(matchhashes)
if filtcount >= self.threshcount:
if hashesfor is not None:
results.append((tid, filtcount, mode, rawcount, matchhashes))
else:
results.append((tid, filtcount, mode, rawcount))

results = sorted(results, key=lambda x: x[1], reverse=True)
# Make sure again to return only those meeting threshcount (needed??)
shortresults = [(tid, filtcnt, mode, rawcount)
for (tid, filtcnt, mode,
rawcount, matchhashes) in results
if filtcnt >= self.threshcount]

if hashesfor is not None:
return shortresults, results[hashesfor][4]
if hashesfor is None:
return results
else:
return shortresults
hashesforhashes = results[hashesfor][4]
results = [(tid, filtcnt, mode, rawcount)
for (tid, filtcnt, mode,
rawcount, matchhashes) in results]
return results, results[hashesfor][4]

def match_file(self, analyzer, ht, filename, number=None):
""" Read in an audio file, calculate its landmarks, query against
Expand Down
2 changes: 2 additions & 0 deletions config.txt
View file
Original file line number Diff line number Diff line change
Expand Up @@ -11,4 +11,6 @@ ncores: 4
[dpwe_matcher]
density: 70
fanout: 8
search_depth: 10000
min_count: 20
ncores: 4
24 changes: 16 additions & 8 deletions dpwe_matcher.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -46,8 +46,10 @@
resultFile = sys.argv[3]

# Default params
defaults = {'density': "70",
'fanout': "8",
defaults = {'density': "70",
'fanout': "8",
'search_depth': "10000",
'min_count': "20",
'ncores': "4"}

# Parse input file
Expand All @@ -62,18 +64,24 @@

density = config.getint(section, 'density')
fanout = config.getint(section, 'fanout')
search_depth = config.getint(section, 'search_depth')
min_count = config.getint(section, 'min_count')
ncores = config.getint(section, 'ncores')

print sys.argv[0], "density:", density, "fanout:", fanout, "ncores:", ncores
print sys.argv[0], "density:", density, "fanout:", fanout, \
"search_depth", search_depth, "min_count", min_count, \
"ncores:", ncores

# Run the command
argv = ["audfprint", "match",
"-d", os.path.join(dir4db, "data.fpdb"),
"--density", str(density),
argv = ["audfprint", "match",
"-d", os.path.join(dir4db, "data.fpdb"),
"--density", str(density),
"--fanout", str(fanout),
"--search-depth", str(search_depth),
"--min-count", str(min_count),
"--ncores", str(ncores),
"--verbose", 0,
"--opfile", resultFile,
"--verbose", 0,
"--opfile", resultFile,
"--list", fileList4query]

# Run audfprint
Expand Down

0 comments on commit ec42a4b

Please sign in to comment.