def dict_score_prons(dict_, candidate_prons):
# adds "score" information to candidate pronunciations. If a candidate
# pronunciation matches a dictionary entry, it's "right." If it matches
# except for stress, it's "partial." Otherwise, it's "wrong."
#
# wsj/s5/local/dict/score_prons.pl
word_and_pron = set()
word_and_pron_nostress = set()
num_pattern = re.compile(r"\d")
for entry in cat(dict_):
word, pron = entry.split(maxsplit=1)
pron_nostress = num_pattern.sub("", pron)
word_and_pron.add(word + ";" + pron)
word_and_pron_nostress.add(word + ";" + pron_nostress)
if isinstance(candidate_prons, str):
candidate_prons = cat(candidate_prons)
for line in candidate_prons:
word, pron = line.split(";", 2)[:2]
pron_nostress = num_pattern.sub("", pron)
if (word + ";" + pron) in word_and_pron:
score = ";right"
elif (word + ";" + pron_nostress) in word_and_pron_nostress:
score = ";partial"
else:
score = ";wrong"
yield line + score
def __init__(self, key: str, img: List[str]):
self.key = key
self.img = img
self.stored_ops: List[int] = []
# Store edge as ids to speed up later matches. Convert to binary, read clockwise
def edge_to_int(s: str) -> int:
return int(str_replace(s, ['.', '#'], ['1', '0']), 2)
top, bottom = img[0], img[-1][::-1]
right, left = cat([l[-1] for l in img]), cat([l[0] for l in img])[::-1]
self.edges = [edge_to_int(side) for side in [left, top, right, bottom]]
def solve_part_two(day_input: List[str]) -> int:
def mask_addresses(mask: str) -> Iterator[int]:
"""Returns all the possible addresses by substituting X by 0 and 1"""
if mask.find('X') == -1:
yield int(mask, 2)
else:
yield from mask_addresses(mask.replace('X', '0', 1))
yield from mask_addresses(mask.replace('X', '1', 1))
mem = dict()
for line in day_input:
op, arg = line.split(' = ')
if op == 'mask':
# Safeguard, don't accept more than 2^10 addresses to write
if (arg.count('X') > 10): raise ValueError(f"An input with more than 2^10 possibilities was encountered: \n{line}")
mask = arg
elif op.startswith('mem'):
# Strategy is to:
# 1. create the address mask doing a manual *or*
# 2. get all possible addresses and set them
addr = f"{int(op[4:-1]):036b}"
new_mask = [a if m == '0' else m for m, a in zip(mask, addr)]
for m in mask_addresses(cat(new_mask)):
mem[m] = int(arg)
return sum(mem.values())
def dict_limit_candidate_prons(rule_hierarchy, candidate_prons):
# This takes the candidate pronunciations from dict_get_candidate_prons
# and the rule hierarchy from dict_get_rule_hierarchy and limits the
# candidate pronunciations to those that use the most specific rules
#
# see wsj/s5/local/dict/limit_candidate_prons.pl etc etc...
hierarchy = set(cat(rule_hierarchy))
def process_word(cur_lines):
pair2rule_list = dict()
for line in cur_lines:
baseword, basepron = line.split(";", 4)[2:4]
key = baseword + ";" + basepron
pair2rule_list.setdefault(key, []).append(line)
for lines in pair2rule_list.values():
stress, rules = [], []
for line in lines:
rulename, destress = line.split(";")[4:6]
stress.append(destress)
rules.append(rulename)
for m in range(len(lines)):
ok = True
for n in range(len(lines)):
if m == n or stress[m] != stress[n]:
continue
if (rules[n] + ";" + rules[m]) in hierarchy:
ok = False
break
if ok:
yield lines[m]
if isinstance(candidate_prons, str):
candidate_prons = cat(candidate_prons)
cur_word = None
cur_lines = []
for line in candidate_prons:
word = line.split(";", 1)[0]
if cur_word is not None and cur_word != word:
for x in process_word(cur_lines):
yield x
cur_lines.clear()
cur_lines.append(line)
cur_word = word
for x in process_word(cur_lines):
yield x
def valid(name, checksum):
"Determine if name is valid according to checksum."
counts = Counter(name.replace('-', ''))
print(counts)
# Note: counts.most_common(5) doesn't work because it breaks ties arbitrarily.
letters = sorted(counts, key=lambda L: (-counts[L], L))
print(letters)
return checksum == cat(letters[:5])
def img_after_ops(img: List[str], ops: List[int]) -> List[str]:
"""Apply rotation and flip *ops* to image *img* returning the result"""
new_img = img[:]
for op in ops:
if op == Tile.ROTATE:
new_img = [cat(l)[::-1] for l in zip(*new_img)]
elif op == Tile.FLIP:
new_img = [l[::-1] for l in new_img]
return new_img
def wsj_prepare_char_dict(data_root, src_dict_suffix, dst_dict_suffix):
phone_dir = os.path.join(data_root, "local", "dict" + src_dict_suffix)
dir_ = os.path.join(data_root, "local", "dict" + dst_dict_suffix)
mkdir(dir_)
lexicon1_raw_nosil_txt = os.path.join(phone_dir, "lexicon1_raw_nosil.txt")
phn_lexicon2_raw_nosil_txt = os.path.join(phone_dir,
"lexicon2_raw_nosil.txt")
unique = OrderedDict()
for entry in cat(lexicon1_raw_nosil_txt):
unique.setdefault(entry.split(" ")[0], entry)
pipe_to(unique.values(), phn_lexicon2_raw_nosil_txt)
char_lexicon2_raw_nosil_txt = os.path.join(dir_, "lexicon2_raw_nosil.txt")
bad_chars = set('!~@#$%^&*()+=/",;:?_{}-')
pipe_to(
(" ".join([x] + [y for y in x if y not in bad_chars])
for x in unique.keys()),
char_lexicon2_raw_nosil_txt,
)
del unique
pipe_to(["SIL", "SPN", "NSN"], os.path.join(dir_, "silence_phones.txt"))
pipe_to(["SIL"], os.path.join(dir_, "optional_silence.txt"))
pipe_to(
sort(
set(
cat(
["!SIL SIL", "<SPOKEN_NOISE> SPN", "<NOISE> NSN"],
char_lexicon2_raw_nosil_txt,
))),
os.path.join(dir_, "lexicon.txt"),
)
pipe_to(
sort(
set(
cat(*(x.split(" ")[1:]
for x in cat(char_lexicon2_raw_nosil_txt))))),
os.path.join(dir_, "nonsilence_phones.txt"),
)
def display(board: List[Position]):
"""Displays a board"""
boundaries = get_boundaries(board)
for w in range(*boundaries['w']):
for z in range(*boundaries['z']):
print(f'z={z}, w={w}')
for y in range(*boundaries['y']):
line = ['.'] * (boundaries['x'][1] - boundaries['x'][0])
for p in board:
if p.z == z and p.y == y:
line[p.x - boundaries['x'][0]] = '#'
print(cat(line))
def wsj_train_lms(data_root, src_dict_suffix, out_dir="local_lm", max_order=4):
# Train a language model on WSJ lm training corpus
# Here we don't do things the Kaldi way. Kaldi uses its own
# derivative-based language modeling. We'll do modified Kneser-Ney
# smoothing, which is a little more widespread.
src_dict_dir = os.path.join(data_root, "local",
"dict" + src_dict_suffix + "_larger")
dst_lm_dir = os.path.join(data_root, "local", out_dir)
mkdir(dst_lm_dir)
vocab = set(x.split()[0]
for x in cat(os.path.join(src_dict_dir, "lexicon.txt")))
vocab.remove("!SIL")
pipe_to(sorted(vocab), os.path.join(dst_lm_dir, "wordlist.txt"))
with gzip.open(os.path.join(src_dict_dir, "cleaned.gz"), "rt") as f:
text = f.read()
sents = ngram_lm.text_to_sents(text,
sent_end_expr=r"\n",
word_delim_expr=r" +")
del text
ngram_counts = ngram_lm.sents_to_ngram_counts(sents, max_order)
ngram_counts[0]["<UNK>"] = 0 # add to vocab
# find any ngrams that contain words that aren't part of the vocabulary.
# we'll prune them. By pruning them we mean completely removing them.
# Modified Kneser-Ney can use the frequency statistics before removing
# them
to_prune = set(ngram_counts[0]) - vocab
to_prune.remove("<S>")
for i, ngram_count in enumerate(ngram_counts[1:]):
if i:
to_prune.update(x for x in ngram_count
if x[:-1] in to_prune or x[-1] in to_prune)
else:
to_prune.update(x for x in ngram_count
if x[0] in to_prune or x[-1] in to_prune)
prob_list = ngram_lm.ngram_counts_to_prob_list_kneser_ney(
ngram_counts, sos="<S>", to_prune=to_prune)
del ngram_counts
lm = ngram_lm.BackoffNGramLM(prob_list, sos="<S>", eos="</S>", unk="<UNK>")
# "pruning" here means removing the probability mass of the sos token and
# redistributing to the other unigrams. "<S>" will remain in the
# vocabulary
lm.prune_by_name({"<S>"})
del prob_list
print("Corpus PPL:", lm.corpus_perplexity(sents))
def solve_part_one(day_input: List[str]) -> str:
lst = [int(i) for i in day_input[0]]
M, m = max(lst), min(lst)
for _ in range(100):
val = lst[0] - 1 if lst[0] > m else M
while val in lst[1:4]:
val = val - 1 if val > m else M
idx = lst.index(val)
lst = lst[4:idx+1] + lst[1:4] + lst[idx+1:] + lst[0:1]
idx = lst.index(1)
return cat(str(i) for i in lst[idx + 1:] + lst[0:idx])
def dict_score_rules(counted_rules,
partial_score=0.8,
ballast=1,
destress_penalty=1e-5):
# weigh the counted rules to derive a score for each rule
#
# wsj/s5/local/dict/score_rules.pl
if isinstance(counted_rules, str):
counted_rules = cat(counted_rules)
for counted_rule in counted_rules:
rule, destress, right, partial, wrong = counted_rule.split(";")
rule_score = int(right) + int(partial) * partial_score
rule_score /= int(right) + int(partial) + int(wrong) + ballast
if destress == "yes":
rule_score -= destress_penalty
yield "{};{};{:.5f}".format(rule, destress, rule_score)
def dict_select_candidate_prons(candidates, max_prons=4, min_rule_score=0.35):
# for a given word, sort its pronunciations by score and return up to
# max_prons of the top-scored candidates, subject to the constraint
# that all returned candidates have a score >= min_rule_score
#
# wsj/s5/local/dict/select_candidates.pl
#
# AFAICT Perl sort-keys-by-value introduces non-determinism in the order
# of pronunciations of the same score. Unfortunately, this determines the
# ultimate candidate pronunciations of some OOV words.
# We pre-empt a bit of that non-determinism, but don't expect perfect
# matching values with Kaldi.
def process_word(cur_lines):
pron2rule_score = dict()
pron2line = dict()
for line in cur_lines:
word, pron, _, _, _, _, score = line.split(";")
score = float(score)
if score >= min_rule_score and score > pron2rule_score.get(
pron, -1):
pron2rule_score[pron] = score
pron2line[pron] = line
prons = sorted(pron2rule_score,
key=lambda x: (-pron2rule_score[x], x),
reverse=False)
for pron, _ in zip(prons, range(max_prons)):
yield pron2line[pron]
cur_lines = []
cur_word = None
if isinstance(candidates, str):
candidates = cat(candidates)
for candidate in candidates:
word, _ = candidate.split(";", 1)
if word != cur_word:
for line in process_word(cur_lines):
yield line
cur_word, cur_lines = word, []
cur_lines.append(candidate)
for line in process_word(cur_lines):
yield line
def dict_reverse_candidates(candidates):
# reverse prefix/suffixes in candidate pronunciation list
#
# wsj/s5/local/dict/reverse_candidates.pl
def reverse_str(x):
return x[::-1]
def reverse_pron(x):
return " ".join(x.split(" ")[::-1])
if isinstance(candidates, str):
candidates = cat(candidates)
for candidate in candidates:
word, pron, baseword, basepron, rule, rest = candidate.split(";", 5)
word, pron = reverse_str(word), reverse_pron(pron)
baseword, basepron = reverse_str(baseword), reverse_pron(basepron)
r_suff, r_bsuff, r_pron, r_bpron = rule.split(",")
r_suff, r_bsuff = reverse_str(r_suff), reverse_str(r_bsuff)
r_pron, r_bpron = reverse_pron(r_pron), reverse_pron(r_bpron)
rule = ",".join((r_suff, r_bsuff, r_pron, r_bpron))
yield ";".join((word, pron, baseword, basepron, rule, rest))
def dict_count_rules(scored_prons):
# count the number of times a rule, stress pair was scored right, partial
# or wrong in dict_score_prons
#
# wsj/s5/local/dict/count_rules.pl
counts = dict()
if isinstance(scored_prons, str):
scored_prons = cat(scored_prons)
for scored_pron in scored_prons:
rulename, destress, score = scored_pron.split(";")[4:]
ref = counts.setdefault(rulename + ";" + destress, [0, 0, 0])
if score == "right":
ref[0] += 1
elif score == "partial":
ref[1] += 1
elif score == "wrong":
ref[2] += 1
else:
raise ValueError("Bad score")
for key, value in counts.items():
yield ";".join([key] + [str(x) for x in value])
def dict_get_rule_hierarchy(rules_path):
# this function looks at the pairs of suffixes from dict_get_rules.
# Whenever those spelling/pronunciation suffixes share a non-empty prefix,
# it implies that that rule could be generalized to one that excludes that
# shared prefix. For (a real) example: ("TICS", ["S"]) and ("TIC", [])
# are a paired rule. Recall that this says you can generate a new word by
# removing "TICS" from the word and adding "TIC", and removing ["S"] from
# its pronunciation and adding [] (or vice versa). Because the suffixes
# share a prefix "TIC", a more general rule might exclude the "TIC" part:
# ("S", ["S"]), ("", []). If this more general rule is also included in the
# set of rules, then the output stores the specific -> general relationship
#
# See wsj/s5/local/dict/get_rule_hierarchy.pl for a more formal description
rules = set(cat(rules_path))
for rule in rules:
A = rule.split(",")
suffix_a, suffix_b = tuple(A[0]), tuple(A[1])
psuffix_a, psuffix_b = A[2].split(), A[3].split()
common_suffix_len = 0
while common_suffix_len < min(len(suffix_a), len(suffix_b)):
if suffix_a[common_suffix_len] != suffix_b[common_suffix_len]:
break
common_suffix_len += 1
common_psuffix_len = 0
while common_psuffix_len < min(len(psuffix_a), len(psuffix_b)):
if psuffix_a[common_psuffix_len] != psuffix_b[common_psuffix_len]:
break
common_psuffix_len += 1
for m in range(common_suffix_len + 1):
sa, sb = "".join(suffix_a[m:]), "".join(suffix_b[m:])
for n in range(common_psuffix_len + 1):
if not n and not m:
continue
psa, psb = " ".join(psuffix_a[n:]), " ".join(psuffix_b[n:])
more_general_rule = ",".join((sa, sb, psa, psb))
if more_general_rule in rules:
yield ";".join((rule, more_general_rule))
elif cmd.startswith('rotate row'):
screen[A, :] = rotate(screen[A,:], B)
elif cmd.startswith('rotate col'):
screen[:,A] = rotate(screen[:, A], B)
def rotate(items, n): return np.append(items[-n:], items[:-n])
def Screen(): return np.zeros((6, 50), dtype = np.int)
def run(commands, screen):
for cmd in commands:
interpret(cmd,screen)
return screen
screen = run(Input(8), Screen())
print(np.sum(screen))
for row in screen:
# print(cat(' @'[pixel] for pixel in row))
# print([' @'[pixel] for pixel in row])
# print([[pixel] for pixel in row])
print(cat( ' @' if pixel == 1 else ' ' for pixel in row))
# a = np.zeros((6,6), dtype=int)
# b = [1,2, 3]
# c = [' @'[m] for m in b]
# print(c)
from common import transpose, first, groupby, cat
import unittest
if __name__ == '__main__':
assert tuple(transpose(((1, 2, 3), (4, 5, 6)))) == ((1, 4), (2, 5), (3, 6))
assert first('abc') == first(['a', 'b', 'c']) == 'a'
assert cat(['a', 'b', 'c']) == 'abc'
assert (groupby(['test', 'one', 'two', 'three', 'four'], key=len) == {
3: ['one', 'two'],
4: ['test', 'four'],
5: ['three']
})
unittest.main()
def parallel_exonerate(ex, subfiles, dbFile, grp_descr,
minScore=77, n=1, x=22, l=30, q=True, via="local"):
futures = [fastqToFasta.nonblocking(ex,sf,n=n,x=x,via=via) for sf in subfiles]
futures2 = []
res = []
resExonerate = []
faSubFiles = []
all_ambiguous = []
all_ambiguous_fastq = []
all_unaligned = []
all_discarded = []
gid, grp_name = grp_descr
my_minscore = _get_minscore(dbFile)
for sf in futures:
subResFile = unique_filename_in()
faSubFiles.append(sf.wait())
futures2.append(exonerate.nonblocking(ex, faSubFiles[-1], dbFile, minScore=my_minscore,
via=via, stdout=subResFile, memory=6))
resExonerate.append(subResFile)
for nf,f in enumerate(resExonerate):
futures2[nf].wait()
(resSplitExonerate,alignments) = split_exonerate(f,minScore,l=l,n=n)
all_unaligned.append(alignments["unaligned"])
all_ambiguous.append(alignments["ambiguous"])
all_ambiguous_fastq.append(alignments["ambiguous_fastq"])
all_discarded.append(alignments["discarded"])
res.append(resSplitExonerate)
gzipfile(ex,cat(all_unaligned[1:],out=all_unaligned[0]))
ex.add(all_unaligned[0]+".gz",
description=set_file_descr(grp_name+"_unaligned.txt.gz",
groupId=gid,step="exonerate",type="txt",
view="admin", comment="scores between %i and %i"%(my_minscore,minScore)) )
# add ambiguous file only if it is not empty
n = count_lines(ex,all_ambiguous[0])
if n > 1:
gzipfile(ex,cat(all_ambiguous[1:],out=all_ambiguous[0]))
ex.add(all_ambiguous[0]+".gz",
description=set_file_descr(grp_name+"_ambiguous.txt.gz",
groupId=gid,step="exonerate",type="txt",
view="admin", comment="multiple equally good classifications") )
# add ambiguous fastq file only if it is not empty
tot_ambiguous = count_lines(ex,all_ambiguous_fastq[0])/4
if n > 1:
gzipfile(ex,cat(all_ambiguous_fastq[1:],out=all_ambiguous_fastq[0]))
ex.add(all_ambiguous_fastq[0]+".gz",
description=set_file_descr(grp_name+"_ambiguous.fastq.gz",
groupId=gid,step="exonerate",type="fastq", comment="multiple equally good classifications") )
# add discarded file only if it is not empty
tot_discarded = count_lines(ex,all_discarded[0])/4
if n > 1:
gzipfile(ex,cat(all_discarded[1:],out=all_discarded[0]))
ex.add(all_discarded[0]+".gz",
description=set_file_descr(grp_name+"_discarded.fastq.gz",
groupId=gid, step="exonerate", type="fastq",
view="admin", comment="< %i bps" %l ) )
gzipfile(ex,faSubFiles[0])
ex.add(faSubFiles[0]+".gz",
description=set_file_descr(grp_name+"_input_part.fa.gz",
groupId=gid, step="init", type="fa",
view="admin", comment="part") )
gzipfile(ex,resExonerate[0])
ex.add(resExonerate[0]+".gz",
description=set_file_descr(grp_name+"_exonerate_part.txt.gz",
groupId=gid, step="exonerate", type="txt",
view="admin", comment="part") )
resFiles = dict((k,'') for d in res for k in d.keys())
for k in resFiles.keys():
v = [d[k] for d in res if k in d]
resFiles[k] = cat(v[1:],out=v[0])
return (resFiles, tot_ambiguous, tot_discarded)
def demultiplex_workflow(ex,
job,
gl,
file_path="../",
via='lsf',
logfile=sys.stdout,
debugfile=sys.stderr):
script_path = gl['script_path']
file_names = {}
job_groups = job.groups
resFiles = {}
for gid, group in job_groups.iteritems():
file_names[gid] = {}
primersFilename = 'group_' + group['name'] + "_barcode_file.fa"
primersFile = group.get("primersfile",
os.path.join(file_path, primersFilename))
ex.add(primersFile,
description=set_file_descr(primersFilename,
groupId=gid,
step="init",
type="fa"))
paramsFilename = 'group_' + group['name'] + "_param_file.txt"
paramsFile = group.get("paramsfile",
os.path.join(file_path, paramsFilename))
ex.add(paramsFile,
description=set_file_descr(paramsFilename,
groupId=gid,
step="init",
type="txt"))
params = load_paramsFile(paramsFile)
infiles = []
tot_counts = 0
allSubFiles = []
for rid, run in group['runs'].iteritems():
infiles.append(run)
n = count_lines(ex, run)
tot_counts += n / 4
if n > 10000000:
allSubFiles.extend(split_file(ex, run, n_lines=8000000))
else:
allSubFiles.append(run)
(resExonerate, tot_ambiguous,
tot_discarded) = parallel_exonerate(ex,
allSubFiles,
primersFile, (gid, group['name']),
via=via,
**params)
gzipfile(ex, cat(infiles))
ex.add(run + ".gz",
description=set_file_descr(group['name'] + "_full_fastq.gz",
groupId=gid,
step='exonerate',
view='debug',
type="fastq"))
logfile.write("Will get sequences to filter\n")
logfile.flush()
seqToFilter = getSeqToFilter(ex, primersFile)
logfile.write("Will filter the sequences\n")
filteredFastq = filterSeq(ex,
resExonerate,
seqToFilter,
gid,
group['name'],
via=via)
logfile.write("After filterSeq, filteredFastq=%s\n" % filteredFastq)
logfile.flush()
counts_primers = {}
counts_primers_filtered = {}
global bcDelimiter
if len(filteredFastq):
archive = unique_filename_in()
tgz = tarfile.open(archive, "w:gz")
for k, f in resExonerate.iteritems():
counts_primers[k] = count_lines(ex, f) / 4
if k in filteredFastq:
k2 = k.replace(bcDelimiter, "_")
file_names[gid][k2] = group['name'] + "_" + k2 + "_filtered"
ex.add(filteredFastq[k],
description=set_file_descr(file_names[gid][k2] +
".fastq",
groupId=gid,
step="final",
type="fastq"))
counts_primers_filtered[k] = count_lines(ex,
filteredFastq[k]) / 4
tgz.add(f,
arcname=group['name'] + "_" +
k.replace(bcDelimiter, "_") + ".fastq")
else:
k2 = k.replace(bcDelimiter, "_")
file_names[gid][k2] = group['name'] + "_" + k2
ex.add(f,
description=set_file_descr(file_names[gid][k2] +
".fastq",
groupId=gid,
step="final",
type="fastq"))
counts_primers_filtered[k] = 0
if len(filteredFastq):
tgz.close()
ex.add(archive,
description=set_file_descr(group['name'] +
"_unfiltered_fastq.tgz",
groupId=gid,
step="exonerate",
type="tar"))
# Prepare report per group of runs
report_ok, reportFile = prepareReport(ex, group['name'], tot_counts,
counts_primers,
counts_primers_filtered,
tot_ambiguous, tot_discarded)
ex.add(reportFile,
description=set_file_descr(group['name'] +
"_report_demultiplexing.txt",
groupId=gid,
step="final",
type="txt",
view="admin"))
if report_ok:
reportFile_pdf = unique_filename_in()
createReport(ex, reportFile, reportFile_pdf, script_path)
ex.add(reportFile_pdf,
description=set_file_descr(group['name'] +
"_report_demultiplexing.pdf",
groupId=gid,
step="final",
type="pdf"))
else:
logfile.write(
"*** Probable ambiguous classification: total_reads < sum(reads_by_primers) ***\n"
)
logfile.flush()
add_pickle(
ex, file_names,
set_file_descr('file_names', step="final", type='py', view='admin'))
return resFiles
def wsj_prepare_dict(data_root, dict_suffix=""):
dir_ = os.path.join(data_root, "local", "dict" + dict_suffix)
cmudict = os.path.join(dir_, "cmudict")
mkdir(cmudict)
# we use the github URL mentioned here
# http://www.speech.cs.cmu.edu/cgi-bin/cmudict
# to avoid using subversion.
url = ("https://raw.githubusercontent.com/Alexir/CMUdict/"
"7a37de79f7e650fd6334333b1b5d2bcf0dee8ad3/")
for x in {"cmudict.0.7a", "cmudict-0.7b.symbols"}:
# 0.7b.symbols the same as 0.7a.symbols
path = os.path.join(cmudict, x)
if not os.path.exists(path):
request.urlretrieve(url + x, path)
silence_phones_txt = os.path.join(dir_, "silence_phones.txt")
optional_silence_txt = os.path.join(dir_, "optional_silence.txt")
pipe_to(["SIL", "SPN", "NSN"], silence_phones_txt)
pipe_to(["SIL"], optional_silence_txt)
nonsilence_phones_txt = os.path.join(dir_, "nonsilence_phones.txt")
phone_pattern = re.compile(r"^(\D+)\d*$")
phones_of = dict()
for phone in cat(os.path.join(cmudict, "cmudict-0.7b.symbols")):
match = phone_pattern.match(phone)
if not match:
raise ValueError("Bad phone {}".format(phone))
base = match.groups(1) # no stress
phones_of.setdefault(base, []).append(phone)
pipe_to((" ".join(x) for x in phones_of.values()), nonsilence_phones_txt)
del phones_of
# skip extra_questions.txt
# there were a few updates to 0.7.a that make the resulting lexicon
# slightly different from Kaldi's
lexicon_raw_nosil_txt = os.path.join(dir_, "lexicon1_raw_nosil.txt")
entry_pattern = re.compile(r"^(\S+)\(\d+\) (.*)$")
lexicon_raw_nosil_lines = []
for line in cat(os.path.join(cmudict, "cmudict.0.7a")):
if line.startswith(";;;"):
continue
match = entry_pattern.match(line)
if match is None:
lexicon_raw_nosil_lines.append(line)
else:
lexicon_raw_nosil_lines.append(" ".join(
[match.group(1), match.group(2)]))
pipe_to(lexicon_raw_nosil_lines, lexicon_raw_nosil_txt)
del lexicon_raw_nosil_lines
lexicon_txt = os.path.join(dir_, "lexicon.txt")
pipe_to(
sort(
set(
cat(
[
"!SIL SIL", "<SPOKEN_NOISE> SPN", "<UNK> SPN",
"<NOISE> NSN"
],
lexicon_raw_nosil_txt,
))),
lexicon_txt,
)
def wsj_extend_dict(dir_13_32_1, data_root, src_dict_suffix, mincount=2):
src_dict_dir = os.path.join(data_root, "local", "dict" + src_dict_suffix)
dst_dict_dir = os.path.join(data_root, "local",
"dict" + src_dict_suffix + "_larger")
if os.path.isdir(dst_dict_dir):
rmtree(dst_dict_dir)
copytree(src_dict_dir, dst_dict_dir)
# lexicon1_raw_nosil.txt is an unsorted (?) version of dict.cmu
dict_cmu = os.path.join(dst_dict_dir, "dict.cmu")
pipe_to(
sort(set(cat(os.path.join(src_dict_dir, "lexicon1_raw_nosil.txt")))),
dict_cmu)
pipe_to(
sort(set(x.split()[0] for x in cat(dict_cmu))),
os.path.join(dst_dict_dir, "wordlist.cmu"),
)
cleaned_gz = os.path.join(dst_dict_dir, "cleaned.gz")
train_data_root = os.path.join(dir_13_32_1, "wsj1", "doc", "lng_modl",
"lm_train", "np_data")
assert os.path.isdir(train_data_root)
train_data_files = []
for subdir in ("87", "88", "89"):
train_data_files.extend(
glob(os.path.join(train_data_root, subdir), r"*.z"))
isword = set(x.strip()
for x in cat(os.path.join(dst_dict_dir, "wordlist.cmu")))
with gzip.open(cleaned_gz, "wt") as out:
for train_data_file in train_data_files:
with open(train_data_file, "rb") as in_:
compressed = in_.read()
decompressed = unlzw(compressed)
in_ = io.TextIOWrapper(io.BytesIO(decompressed))
for line in in_:
if line.startswith("<"):
continue
A = line.strip().upper().split(" ")
for n, a in enumerate(A):
if a not in isword and len(a) > 1 and a.endswith("."):
out.write(a[:-1])
if n < len(A) - 1:
out.write("\n")
else:
out.write(a + " ")
out.write("\n")
del in_, compressed, decompressed
counts = Counter()
with gzip.open(cleaned_gz, "rt") as cleaned:
for line in cleaned:
for token in line.strip().split():
counts[token] += 1
counts = sorted(((v, k) for (k, v) in counts.items()), reverse=True)
digits = set(str(x) for x in range(10))
oov_counts_path = os.path.join(dst_dict_dir, "oov.counts")
oovlist_path = os.path.join(dst_dict_dir, "oovlist")
with open(os.path.join(dst_dict_dir, "unigrams"),
"w") as unigrams, open(oov_counts_path,
"w") as oov_cnts, open(oovlist_path,
"w") as oov_lst:
for count, word in counts:
line = "{} {}\n".format(count, word)
unigrams.write(line)
if word not in isword:
oov_cnts.write(line)
if not (set(word) & digits) and count >= mincount:
oov_lst.write(word + "\n")
del counts
dict_acronyms_path = os.path.join(dst_dict_dir, "dict.acronyms")
pipe_to(
dict_get_acronym_prons(oovlist_path, dict_cmu),
dict_acronyms_path,
)
f_dir = os.path.join(dst_dict_dir, "f")
b_dir = os.path.join(dst_dict_dir, "b")
mkdir(f_dir, b_dir)
banned = set(",;")
pipe_to(
(x for x in cat(dict_cmu) if not (set(x.split()[0]) & banned)),
os.path.join(f_dir, "dict"),
)
pipe_to(
(x for x in cat(os.path.join(dst_dict_dir, "oovlist"))
if not (set(x.split()[0]) & banned)),
os.path.join(f_dir, "oovs"),
)
pipe_to(
(" ".join([w[::-1]] + p.split()[::-1])
for (w, p) in (x.split(" ", 1)
for x in cat(os.path.join(f_dir, "dict")))),
os.path.join(b_dir, "dict"),
)
pipe_to((x[::-1] for x in cat(os.path.join(f_dir, "oovs"))),
os.path.join(b_dir, "oovs"))
for dir_ in (f_dir, b_dir):
dict_path = os.path.join(dir_, "dict")
rules_path = os.path.join(dir_, "rules")
hierarchy_path = os.path.join(dir_, "hierarchy")
oovs_path = os.path.join(dir_, "oovs")
rule_counts_path = os.path.join(dir_, "rule.counts")
rules_with_scores_path = os.path.join(dir_, "rules.with_scores")
oov_candidates_path = os.path.join(dir_, "oovs.candidates")
pipe_to(dict_get_rules(cat(dict_path)), rules_path)
pipe_to(dict_get_rule_hierarchy(rules_path), hierarchy_path)
pipe_to(
dict_count_rules(
dict_score_prons(
dict_path,
dict_limit_candidate_prons(
hierarchy_path,
dict_get_candidate_prons(rules_path, dict_path,
dict_path),
),
)),
rule_counts_path,
)
pipe_to(
sorted(
dict_score_rules(rule_counts_path),
key=lambda x: (float(x.split(";")[2]), x),
reverse=True,
),
rules_with_scores_path,
)
pipe_to(
dict_limit_candidate_prons(
hierarchy_path,
dict_get_candidate_prons(rules_with_scores_path, dict_path,
oovs_path),
),
oov_candidates_path,
)
oov_candidates_path = os.path.join(dst_dict_dir, "oovs.candidates")
pipe_to(
sorted(
cat(
dict_reverse_candidates(os.path.join(b_dir,
"oovs.candidates")),
os.path.join(f_dir, "oovs.candidates"),
)),
oov_candidates_path,
)
dict_oovs_path = os.path.join(dst_dict_dir, "dict.oovs")
pipe_to(
("{0} {1}".format(*x.split(";"))
for x in dict_select_candidate_prons(oov_candidates_path)),
dict_oovs_path,
)
dict_oovs_merged_path = os.path.join(dst_dict_dir, "dict.oovs_merged")
pipe_to(
sorted(set(cat(dict_acronyms_path, dict_oovs_path))),
dict_oovs_merged_path,
)
pipe_to(
sorted(
set(
cat(
[
"!SIL SIL", "<SPOKEN_NOISE> SPN", "<UNK> SPN",
"<NOISE> NSN"
],
dict_cmu,
dict_oovs_merged_path,
))),
os.path.join(dst_dict_dir, "lexicon.txt"),
)
def dict_get_acronym_prons(oovlist, dict_):
# this function first extracts single-letter acronyms from the CMU dict.
# It then looks for words in the oovlist file that look like acronyms, and
# builds up some pronunciations based on them
#
# consult Kaldi's wsj/s5/local/dict/get_acronym_prons.pl for more details
def get_letter_prons(letters, letter_prons):
acronym = list(letters)
prons = [""]
while acronym:
letter = acronym.pop(0)
n = 1
while acronym and acronym[0] == letter:
acronym.pop(0)
n += 1
letter_pron = letter_prons[letter]
prons_of_block = []
if n == 2:
for lpron in letter_pron:
prons_of_block.append("D AH1 B AH0 L " + lpron)
prons_of_block.append(lpron + " " + lpron)
elif n == 3:
for lpron in letter_pron:
prons_of_block.append("T R IH1 P AH0 L " + lpron)
prons_of_block.append(" ".join([lpron] * 3))
else:
for lpron in letter_pron:
prons_of_block.append(" ".join([lpron] * n))
new_prons = []
for pron in prons:
for pron_of_block in prons_of_block:
if pron:
new_prons.append(pron + " " + pron_of_block)
else:
new_prons.append(pron_of_block)
prons = new_prons
assert prons[0] != ""
for pron in prons:
yield pron
if isinstance(dict_, str):
dict_ = cat(dict_)
letter_pattern = re.compile(r"^[A-Z]\.$")
letter_prons = dict()
for line in dict_:
word, pron = line.strip().split(" ", maxsplit=1)
if letter_pattern.match(word):
letter = word[0]
letter_prons.setdefault(letter, []).append(pron.strip())
if isinstance(oovlist, str):
oovlist = cat(oovlist)
acro_wo_points_pattern = re.compile(r"^[A-Z]{1,5}$")
acro_w_points_pattern = re.compile(r"^([A-Z]\.){1,4}[A-Z]\.?$")
for word in oovlist:
word = word.strip()
if acro_wo_points_pattern.match(word):
for pron in get_letter_prons(word, letter_prons):
yield word + " " + pron
elif acro_w_points_pattern.match(word):
for pron in get_letter_prons(word.replace(".", ""), letter_prons):
yield word + " " + pron
def parallel_exonerate(ex,
subfiles,
dbFile,
grp_descr,
minScore=77,
n=1,
x=22,
l=30,
trim=True,
via="local"):
futures = [
fastqToFasta.nonblocking(ex, sf, n=n, x=x, via=via) for sf in subfiles
]
futures2 = []
res = []
resExonerate = []
faSubFiles = []
all_ambiguous = []
all_ambiguous_fastq = []
all_unaligned = []
all_discarded = []
gid, grp_name = grp_descr
my_minscore = _get_minscore(dbFile)
primersDict = get_primersList(dbFile)
for sf in futures:
subResFile = unique_filename_in()
faSubFiles.append(sf.wait())
futures2.append(
exonerate.nonblocking(ex,
faSubFiles[-1],
dbFile,
minScore=my_minscore,
via=via,
stdout=subResFile,
memory=6))
resExonerate.append(subResFile)
for nf, f in enumerate(resExonerate):
futures2[nf].wait()
(resSplitExonerate, alignments) = split_exonerate(f,
primersDict,
minScore,
l=l,
n=n,
trim=trim)
all_unaligned.append(alignments["unaligned"])
all_ambiguous.append(alignments["ambiguous"])
all_ambiguous_fastq.append(alignments["ambiguous_fastq"])
all_discarded.append(alignments["discarded"])
res.append(resSplitExonerate)
# add unaligned file only if it is not empty
n = count_lines(ex, all_unaligned[0])
if n > 1:
catfile = cat(all_unaligned)
gzipfile(ex, catfile)
ex.add(catfile + ".gz",
description=set_file_descr(grp_name + "_unaligned.txt.gz",
groupId=gid,
step="exonerate",
type="txt",
view="admin",
comment="scores between %i and %i" %
(my_minscore, minScore)))
# add ambiguous file only if it is not empty
n = count_lines(ex, all_ambiguous[0])
if n > 1:
catfile = cat(all_ambiguous)
gzipfile(ex, catfile)
ex.add(catfile + ".gz",
description=set_file_descr(
grp_name + "_ambiguous.txt.gz",
groupId=gid,
step="exonerate",
type="txt",
view="admin",
comment="multiple equally good classifications"))
# add ambiguous fastq file only if it is not empty
tot_ambiguous = count_lines(ex, all_ambiguous_fastq[0]) / 4
if tot_ambiguous > 1:
catfile = cat(all_ambiguous_fastq)
gzipfile(ex, catfile)
ex.add(catfile + ".gz",
description=set_file_descr(
grp_name + "_ambiguous.fastq.gz",
groupId=gid,
step="exonerate",
type="fastq",
comment="multiple equally good classifications"))
# add discarded file only if it is not empty
tot_discarded = count_lines(ex, all_discarded[0]) / 4
if tot_discarded > 1:
catfile = cat(all_discarded)
gzipfile(ex, catfile)
ex.add(catfile + ".gz",
description=set_file_descr(grp_name + "_discarded.fastq.gz",
groupId=gid,
step="exonerate",
type="fastq",
view="admin",
comment="remaining seq too short"))
## add part input fasta file only if it is not empty
n = count_lines(ex, faSubFiles[0])
if n > 1:
gzipfile(ex, faSubFiles[0])
ex.add(faSubFiles[0] + ".gz",
description=set_file_descr(grp_name + "_input_part.fa.gz",
groupId=gid,
step="init",
type="fa",
view="admin",
comment="part"))
## add part res exonerate only if it is not empty
n = count_lines(ex, resExonerate[0])
if n > 1:
gzipfile(ex, resExonerate[0])
ex.add(resExonerate[0] + ".gz",
description=set_file_descr(grp_name + "_exonerate_part.txt.gz",
groupId=gid,
step="exonerate",
type="txt",
view="admin",
comment="part"))
resFiles = dict((k, '') for d in res for k in d.keys())
for k in resFiles.keys():
v = [d[k] for d in res if k in d]
resFiles[k] = cat(v[1:], out=v[0])
return (resFiles, tot_ambiguous, tot_discarded)
if k in fastqFiles:
indexFiles[k] = bowtie_build.nonblocking(ex,f,via=via)
unalignedFiles = {}
futures = []
bwtarg = ["-a","-q","-n","2","-l","20","--un"]
for k,f in indexFiles.iteritems():
unalignedFiles[k] = unique_filename_in()
touch(ex,unalignedFiles[k])
futures.append(bowtie.nonblocking( ex, f.wait(), fastqFiles[k],
bwtarg+[unalignedFiles[k]], via='lsf'))
for f in futures: f.wait()
return unalignedFiles
if __name__ == "__main__":
primersDict=get_primersList(sys.argv[2])
(resSplitExonerate,alignments)=split_exonerate(sys.argv[1],primersDict,minScore=8,l=30,n=1,trim='False')
res = []
res.append(resSplitExonerate)
resFiles = dict((k,'') for d in res for k in d.keys())
for k in resFiles.keys():
v = [d[k] for d in res if k in d]
resFiles[k] = cat(v[1:],out=v[0])
print resFiles
print resSplitExonerate
print alignments
print "Done"
from common import Input, transpose, cat
from collections import Counter
if __name__ == '__main__':
data = Input(6).read().split()
t_data = [col for col in transpose(data)]
couters = [Counter(m) for m in t_data]
frequenc_items = [m.most_common(1)[0][0] for m in couters]
print(cat(frequenc_items))
def demultiplex_workflow(ex, job, gl, file_path="../", via='lsf',
logfile=sys.stdout, debugfile=sys.stderr):
script_path = gl['script_path']
file_names = {}
job_groups=job.groups
resFiles={}
for gid, group in job_groups.iteritems():
file_names[gid] = {}
primersFilename = 'group_' + group['name'] + "_barcode_file.fa"
primersFile = group.get("primersfile",os.path.join(file_path,primersFilename))
ex.add(primersFile,description=set_file_descr(primersFilename,groupId=gid,step="init",type="fa"))
paramsFilename = 'group_' + group['name'] + "_param_file.txt"
paramsFile = group.get("paramsfile",os.path.join(file_path,paramsFilename))
ex.add(paramsFile,description=set_file_descr(paramsFilename,groupId=gid,step="init",type="txt"))
params = load_paramsFile(paramsFile)
infiles = []
tot_counts = 0
allSubFiles = []
for rid,run in group['runs'].iteritems():
infiles.append(run)
n=count_lines(ex,run)
tot_counts += n/4
if n>10000000:
allSubFiles.extend(split_file(ex,run,n_lines=8000000))
else:
allSubFiles.append(run)
(resExonerate, tot_ambiguous, tot_discarded) = parallel_exonerate( ex, allSubFiles, primersFile,
(gid, group['name']), via=via, **params )
gzipfile(ex, cat(infiles))
ex.add(run+".gz",description=set_file_descr(group['name']+"_full_fastq.gz",
groupId=gid,step='exonerate',view='debug',type="fastq"))
logfile.write("Will get sequences to filter\n");logfile.flush()
seqToFilter = getSeqToFilter(ex,primersFile)
logfile.write("Will filter the sequences\n")
filteredFastq = filterSeq(ex,resExonerate,seqToFilter,gid,group['name'],via=via)
logfile.write("After filterSeq, filteredFastq=%s\n" %filteredFastq);logfile.flush()
counts_primers = {}
counts_primers_filtered = {}
global bcDelimiter
if len(filteredFastq):
archive = unique_filename_in()
tgz = tarfile.open(archive, "w:gz")
for k,f in resExonerate.iteritems():
counts_primers[k] = count_lines(ex,f)/4
if k in filteredFastq:
k2 = k.replace(bcDelimiter,"_")
file_names[gid][k2] = group['name']+"_"+k2+"_filtered"
ex.add(filteredFastq[k],description=set_file_descr(file_names[gid][k2]+".fastq",
groupId=gid,step="final",
type="fastq"))
counts_primers_filtered[k] = count_lines(ex,filteredFastq[k])/4
tgz.add( f, arcname=group['name']+"_"+k.replace(bcDelimiter,"_")+".fastq" )
else:
k2 = k.replace(bcDelimiter,"_")
file_names[gid][k2] = group['name']+"_"+k2
ex.add(f,description=set_file_descr(file_names[gid][k2]+".fastq",
groupId=gid,step="final",
type="fastq"))
counts_primers_filtered[k] = 0
if len(filteredFastq):
tgz.close()
ex.add(archive,description=set_file_descr(group['name']+"_unfiltered_fastq.tgz",
groupId=gid,step="exonerate",
type="tar"))
# Prepare report per group of runs
report_ok,reportFile = prepareReport(ex,group['name'],
tot_counts, counts_primers,counts_primers_filtered,
tot_ambiguous, tot_discarded)
ex.add(reportFile,description = set_file_descr(
group['name']+"_report_demultiplexing.txt",
groupId=gid,step="final",type="txt",view="admin"))
if report_ok:
reportFile_pdf = unique_filename_in()
createReport(ex,reportFile,reportFile_pdf,script_path)
ex.add(reportFile_pdf,description=set_file_descr(
group['name']+"_report_demultiplexing.pdf",
groupId=gid,step="final",type="pdf"))
else:
logfile.write("*** Probable ambiguous classification: total_reads < sum(reads_by_primers) ***\n");logfile.flush()
add_pickle( ex, file_names,
set_file_descr('file_names',step="final",type='py',view='admin') )
return resFiles
futures.append(
bowtie.nonblocking(ex,
f.wait(),
fastqFiles[k],
bwtarg + [unalignedFiles[k]],
via='lsf'))
for f in futures:
f.wait()
return unalignedFiles
if __name__ == "__main__":
primersDict = get_primersList(sys.argv[2])
(resSplitExonerate, alignments) = split_exonerate(sys.argv[1],
primersDict,
minScore=8,
l=30,
n=1,
trim='False')
res = []
res.append(resSplitExonerate)
resFiles = dict((k, '') for d in res for k in d.keys())
for k in resFiles.keys():
v = [d[k] for d in res if k in d]
resFiles[k] = cat(v[1:], out=v[0])
print resFiles
print resSplitExonerate
print alignments
print "Done"
def dict_get_candidate_prons(rules, dict_, words, min_prefix_len=3):
# the purpose of this script is to apply the rules from dict_get_rules to
# the word list, given the dictionary. It does no pruning based on
# hierarchy. We're basically removing one suffix and attaching another.
#
# see wsj/s5/local/dict/get_candidate_prons.pl for a more formal descript.
isrule = set()
suffix2rule = dict()
rule_and_stress_to_rule_score = dict()
for rule in cat(rules):
rule = rule.split(";", 3)
rule_score = rule.pop() if len(rule) == 3 else -1
destress = rule.pop() if len(rule) == 2 else None
assert len(rule) == 1
rule = rule[0]
R = rule.split(",")
if len(R) != 4:
raise ValueError("Bad rule {}".format(rule))
suffix = R[0]
if rule not in isrule:
isrule.add(rule)
suffix2rule.setdefault(suffix, []).append(R)
if destress is None:
rule_and_stress_to_rule_score[rule + ";yes"] = rule_score
rule_and_stress_to_rule_score[rule + ";no"] = rule_score
else:
rule_and_stress_to_rule_score[rule + ";" + destress] = rule_score
word2prons = dict()
for entry in cat(dict_):
word, pron = entry.split(maxsplit=1)
word2prons.setdefault(word, []).append(pron)
prefixcount = Counter(
chain.from_iterable(
(w[:p] for p in range(len(w) + 1)) for w in word2prons))
if isinstance(words, str):
words = cat(words)
for word in words:
word = word.split(maxsplit=1)[0]
ownword = 1 if word in word2prons else 0
for prefix_len in range(min_prefix_len, len(word) + 1):
prefix, suffix = word[:prefix_len], word[prefix_len:]
if prefixcount.get(prefix, 0) - ownword == 0:
continue
rules_array_ref = suffix2rule.get(suffix, None)
if rules_array_ref is None:
continue
for R in rules_array_ref:
_, base_suffix, psuffix, base_psuffix = R
base_word = prefix + base_suffix
base_prons_ref = word2prons.get(base_word, None)
if base_prons_ref is None:
continue
if base_psuffix:
base_psuffix = " " + base_psuffix
# FIXME(sdrobert): I think this might split up phones. This'll
# be bad when some phones share prefixes with others.
for base_pron in base_prons_ref:
base_pron_prefix_len = len(base_pron) - len(base_psuffix)
if (base_pron_prefix_len < 0 or
base_pron[base_pron_prefix_len:] != base_psuffix):
continue
pron_prefix = base_pron[:base_pron_prefix_len]
rule = ",".join(R)
for destress in range(2):
if destress:
pron_prefix = pron_prefix.replace("2", "1")
destress_mark = "yes"
else:
destress_mark = "no"
pron = pron_prefix
if psuffix:
pron += " " + psuffix
rule_score = rule_and_stress_to_rule_score.get(
rule + ";" + destress_mark, None)
if rule_score is None:
continue
output = [
word, pron, base_word, base_pron, rule,
destress_mark
]
if rule_score != -1:
output.append(str(rule_score))
yield ";".join(output)