def renderout(tmplfname, data, groupby, nperhit):
"""Renders mturk output"""
import web, web.template
from nkutils import partitionByFunc
from nkwebutils import NKStor, mystorify
# de-multiplex and simplify data
data = sum([demultiplex(row, nperhit) for row in data], [])
# group by common key
grouped, _ = partitionByFunc(data, lambda d: d[groupby])
results = []
Cls = NKStor
# build up list of results
for gkey, g in sorted(grouped.items()):
# build up a list of common keys for this result group
r = Cls(g[0])
for el in g:
for k, v in r.items():
if el[k] != v:
del r[k]
# now create each individual sub-output
r['outputs'] = [Cls(el) for el in g]
results.append(r)
#pprint(results)
# render results
renfunc = web.template.frender(tmplfname)
s = renfunc(results)
return s
def textaudiomainauto(txtfname, labelsfname, subfname):
"""A driver that takes a text and label file and creates subtitles.
This tries to do it automatically, but doesn't work too well.
The txt file should contain blank lines for major parts with no dialog.
Lines starting with '(' are for signs in the video (no speech).
The labels are as output from audacity's labeling feature:
start time in seconds \t end time in seconds \t optional label
(The labels are ignored.)
"""
# Read script and tokenize into chunks
import re
from nkutils import memoize, spark, partitionByFunc
import numpy as np
if 0:
DLM = '([".,;:?!\n][\n]?)'
DLMSPACE = '([ ".,;:?!\n][\n]?)'
else:
DLM = '([".,;:?!\n]+)'
DLMSPACE = '([ ".,;:?!\n]+)'
lines = [l.strip() for l in open(txtfname)]
full = " ".join([l.strip() for l in open(txtfname) if l.strip()])
ntotallines = len(lines)
# script = [l.strip() for l in open(txtfname) if not l.startswith('(')]
allseqs, indices = partitionByFunc(lines, lambda s: "comment" if s.startswith("(") else "script")
# indices is a dictionary of (outval, i) -> orig_i, which allows mapping results back.
comments, script = allseqs["comment"], allseqs["script"]
script = "\n".join(script)
while "\n\n" in script:
script = script.replace("\n\n", "\n")
nlines = len(script.split("\n"))
nchars = len(script)
nwords = len(list(re.finditer(DLMSPACE, script)))
tokens = list(re.finditer(DLM, script))
locs = set([0, len(script) - 1])
for t in tokens:
locs.add(t.end())
locs = sorted(locs)
toks = ["%s (%s)" % (t.group(), t.span()) for t in tokens]
print "Read %d non-comment script lines (%d words, %d tokens, %d chars, %d locs): %s %s" % (
nlines,
nwords,
len(tokens),
nchars,
len(locs),
toks[:4],
locs[:4],
)
# Read labels and compute speaking rates
labels = [map(float, l.strip().split("\t")[:2]) for l in open(labelsfname)]
llens = [b - a for a, b in labels]
totalsecs = sum(llens)
print "Read %d labels, %0.2f secs: %s" % (len(labels), totalsecs, zip(labels, llens)[:2])
wpm = nwords / (totalsecs / 60.0)
spc = totalsecs / nchars
print "Got %0.1f wpm, %0.4f secs per char" % (wpm, spc)
# Define cost function and memoize it
def costfunc(labelnum, start, end, zerocost=0.2, spc=spc):
"""Computes the cost (in secs) of assigning the given start and end locs to the label.
The locs are specified w.r.t. to the 'locs' array. They can be identical.
If the length is 0, the cost is 'zerocost'.
Else, the cost is (length of label) - (length of chunk)*spc
Notice that's signed: positive means label is longer than chunk, and vice versa.
"""
if start == end:
return zerocost
t = llens[labelnum]
try:
i, j = locs[start], locs[end]
nchars = j - i
nsecs = spc * nchars
# print t, i, j, nchars, nsecs
return t - nsecs
except:
return zerocost
C = memoize(costfunc)
# print C(0, 0, 0)
# print C(0, 0, 1)
# print C(0, 0, 2)
# print C(0, 1, 2)
# Initialize chunks
M = len(locs) - 1
fac = M / float(len(llens))
chunks = [[min(int(i * fac), M), min(int((i + 1) * fac), M)] for i in range(len(llens))]
print len(llens), len(chunks), llens[:5], chunks[:5] + chunks[-5:]
if 0:
print locs
for a, b in zip(locs, locs[1:]):
print "<%s>" % (script[a:b].strip())
sys.exit()
costs = [C(i, a, b) for i, (a, b) in enumerate(chunks)]
acosts = np.abs(np.array(costs))
best = [sum(acosts), chunks]
iter = 0
from random import randint
while iter < 10:
iter += 1
n = np.argmax(acosts)
mc = costs[n]
which = randint(0, 1)
print "On iter %d, total cost %0.3f, maxcost %0.3f at %d, shifting %d" % (iter, sum(acosts), mc, n, which)
print " %s" % (chunks2str(chunks))
if mc < 0: # label shorter than chunk
incr = 1 if which == 0 else -1
else: # label longer than chunk
incr = 1 if which == 1 else -1
newchunks = shiftchunk(chunks, n, which, incr)
costs = [C(i, a, b) for i, (a, b) in enumerate(newchunks)]
acosts = np.abs(np.array(costs))
if sum(acosts) < best[0]:
chunks = newchunks
print chunks
# now write output
sf = srtfile(subfname)
last = 0
# print full
for idx, ((i, j), (t0, t1)) in enumerate(zip(chunks, labels)):
if i == j:
continue
if i < 0 or j >= len(locs):
continue
s = script[locs[i] : locs[j]].strip()
try:
n = full.index(s.replace("\n", " "))
except Exception, e:
print " ERROR: |%s|: %s" % (s, full[:200])
# TODO this is because of comments inside the s
n = 1
# raise
if n > 0:
# we have some skipped stuff, so dump it all in a single line
dif = 0.05 * (t0 - last) # so we're not touching boundaries
sf(full[:n].strip(), last + dif, t0 - dif)
# print '%d: %s' % ((full.index(s) if s in full else -1), s)
full = full[n + len(s) :].strip()
# see if we have any skipped things todo
sf(s, t0, t1)
last = t1
def textaudiomainauto(txtfname, labelsfname, subfname):
"""A driver that takes a text and label file and creates subtitles.
This tries to do it automatically, but doesn't work too well.
The txt file should contain blank lines for major parts with no dialog.
Lines starting with '(' are for signs in the video (no speech).
The labels are as output from audacity's labeling feature:
start time in seconds \t end time in seconds \t optional label
(The labels are ignored.)
"""
# Read script and tokenize into chunks
import re
from nkutils import memoize, spark, partitionByFunc
import numpy as np
if 0:
DLM = '([".,;:?!\n][\n]?)'
DLMSPACE = '([ ".,;:?!\n][\n]?)'
else:
DLM = '([".,;:?!\n]+)'
DLMSPACE = '([ ".,;:?!\n]+)'
lines = [l.strip() for l in open(txtfname)]
full = ' '.join([l.strip() for l in open(txtfname) if l.strip()])
ntotallines = len(lines)
#script = [l.strip() for l in open(txtfname) if not l.startswith('(')]
allseqs, indices = partitionByFunc(
lines, lambda s: 'comment' if s.startswith('(') else 'script')
#indices is a dictionary of (outval, i) -> orig_i, which allows mapping results back.
comments, script = allseqs['comment'], allseqs['script']
script = '\n'.join(script)
while '\n\n' in script:
script = script.replace('\n\n', '\n')
nlines = len(script.split('\n'))
nchars = len(script)
nwords = len(list(re.finditer(DLMSPACE, script)))
tokens = list(re.finditer(DLM, script))
locs = set([0, len(script) - 1])
for t in tokens:
locs.add(t.end())
locs = sorted(locs)
toks = ['%s (%s)' % (t.group(), t.span()) for t in tokens]
print 'Read %d non-comment script lines (%d words, %d tokens, %d chars, %d locs): %s %s' % (
nlines, nwords, len(tokens), nchars, len(locs), toks[:4], locs[:4])
# Read labels and compute speaking rates
labels = [map(float, l.strip().split('\t')[:2]) for l in open(labelsfname)]
llens = [b - a for a, b in labels]
totalsecs = sum(llens)
print 'Read %d labels, %0.2f secs: %s' % (len(labels), totalsecs,
zip(labels, llens)[:2])
wpm = nwords / (totalsecs / 60.0)
spc = totalsecs / nchars
print 'Got %0.1f wpm, %0.4f secs per char' % (wpm, spc)
# Define cost function and memoize it
def costfunc(labelnum, start, end, zerocost=0.2, spc=spc):
"""Computes the cost (in secs) of assigning the given start and end locs to the label.
The locs are specified w.r.t. to the 'locs' array. They can be identical.
If the length is 0, the cost is 'zerocost'.
Else, the cost is (length of label) - (length of chunk)*spc
Notice that's signed: positive means label is longer than chunk, and vice versa.
"""
if start == end: return zerocost
t = llens[labelnum]
try:
i, j = locs[start], locs[end]
nchars = j - i
nsecs = spc * nchars
#print t, i, j, nchars, nsecs
return t - nsecs
except:
return zerocost
C = memoize(costfunc)
#print C(0, 0, 0)
#print C(0, 0, 1)
#print C(0, 0, 2)
#print C(0, 1, 2)
# Initialize chunks
M = len(locs) - 1
fac = M / float(len(llens))
chunks = [[min(int(i * fac), M),
min(int((i + 1) * fac), M)] for i in range(len(llens))]
print len(llens), len(chunks), llens[:5], chunks[:5] + chunks[-5:]
if 0:
print locs
for a, b in zip(locs, locs[1:]):
print '<%s>' % (script[a:b].strip())
sys.exit()
costs = [C(i, a, b) for i, (a, b) in enumerate(chunks)]
acosts = np.abs(np.array(costs))
best = [sum(acosts), chunks]
iter = 0
from random import randint
while iter < 10:
iter += 1
n = np.argmax(acosts)
mc = costs[n]
which = randint(0, 1)
print 'On iter %d, total cost %0.3f, maxcost %0.3f at %d, shifting %d' % (
iter, sum(acosts), mc, n, which)
print ' %s' % (chunks2str(chunks))
if mc < 0: # label shorter than chunk
incr = 1 if which == 0 else -1
else: # label longer than chunk
incr = 1 if which == 1 else -1
newchunks = shiftchunk(chunks, n, which, incr)
costs = [C(i, a, b) for i, (a, b) in enumerate(newchunks)]
acosts = np.abs(np.array(costs))
if sum(acosts) < best[0]:
chunks = newchunks
print chunks
# now write output
sf = srtfile(subfname)
last = 0
#print full
for idx, ((i, j), (t0, t1)) in enumerate(zip(chunks, labels)):
if i == j: continue
if i < 0 or j >= len(locs): continue
s = script[locs[i]:locs[j]].strip()
try:
n = full.index(s.replace('\n', ' '))
except Exception, e:
print ' ERROR: |%s|: %s' % (s, full[:200])
#TODO this is because of comments inside the s
n = 1
#raise
if n > 0:
# we have some skipped stuff, so dump it all in a single line
dif = 0.05 * (t0 - last) # so we're not touching boundaries
sf(full[:n].strip(), last + dif, t0 - dif)
#print '%d: %s' % ((full.index(s) if s in full else -1), s)
full = full[n + len(s):].strip()
# see if we have any skipped things todo
sf(s, t0, t1)
last = t1