def start_state(root):
''' None -> <s>^{g-1} . TOP </s>^{g-1} '''
## LMState.cache = {}
lmstr = LMState.lm.raw_startsyms()
lhsstr = lmstr + [root] + LMState.lm.raw_stopsyms()
edge = Hyperedge(None, [root], Vector(), lhsstr)
edge.lmlhsstr = LMState.lm.startsyms() + [root] + LMState.lm.stopsyms()
edge.rule = Rule.parse("ROOT(TOP) -> x0 ### ")
sc = root.bestres[0] if FLAGS.futurecost else 0
return LMState(None, [DottedRule(edge, dot=len(lmstr))], LMState.lm.startsyms(),
step=0, score=sc) # future cost
def load(filename, is_tforest=False, lower=False, sentid=0, first=None, lm=None):
'''now returns a generator! use load().next() for singleton.
and read the last line as the gold tree -- TODO: optional!
and there is an empty line at the end
'''
if first is None: # N.B.: must be here, not in the param line (after program initializes)
first = FLAGS.first
file = getfile(filename)
line = None
total_time = 0
num_sents = 0
while True:
start_time = time.time()
##'\tThe complicated language in ...\n"
## tag is often missing
line = file.readline() # emulate seek
if len(line) == 0:
break
try:
## strict format, no consecutive breaks
# if line is None or line == "\n":
# line = "\n"
# while line == "\n":
# line = file.readline() # emulate seek
tag, sent = line.split("\t") # foreign sentence
except:
## no more forests
yield None
continue
num_sents += 1
# caching the original, word-based, true-case sentence
sent = sent.split() ## no splitting with " "
cased_sent = sent [:]
if lower:
sent = [w.lower() for w in sent] # mark johnson: lowercase all words
#sent = words_to_chars(sent, encode_back=True) # split to chars
## read in references
refnum = int(file.readline().strip())
refs = []
for i in xrange(refnum):
refs.append(file.readline().strip())
## sizes: number of nodes, number of edges (optional)
num, nedges = map(int, file.readline().split("\t"))
forest = Forest(sent, cased_sent, tag, is_tforest)
forest.tag = tag
forest.refs = refs
forest.bleu = Bleu(refs=refs) ## initial (empty test) bleu; used repeatedly later
forest.labelspans = {}
forest.short_edges = {}
forest.rules = {}
for i in xrange(1, num+1):
## '2\tDT* [0-1]\t1 ||| 1232=2 ...\n'
## node-based features here: wordedges, greedyheavy, word(1), [word(2)], ...
line = file.readline()
try:
keys, fields = line.split(" ||| ")
except:
keys = line
fields = ""
iden, labelspan, size = keys.split("\t") ## iden can be non-ints
size = int(size)
fvector = Vector(fields) #
## remove_blacklist(fvector)
node = Node(iden, labelspan, size, fvector, sent)
forest.add_node(node)
if cache_same:
if labelspan in forest.labelspans:
node.same = forest.labelspans[labelspan]
node.fvector = node.same.fvector
else:
forest.labelspans[labelspan] = node
for j in xrange(size):
is_oracle = False
## '\t1 ||| 0=8.86276 1=2 3\n'
## N.B.: can't just strip! "\t... ||| ... ||| \n" => 2 fields instead of 3
tails, rule, fields = file.readline().strip("\t\n").split(" ||| ")
if tails != "" and tails[0] == "*": #oracle edge
is_oracle = True
tails = tails[1:]
tails = tails.split() ## N.B.: don't split by " "!
tailnodes = []
lhsstr = [] # 123 "thank" 456
lmstr = []
lmscore = 0
lmlhsstr = []
for x in tails:
if x[0]=='"': # word
word = desymbol(x[1:-1])
lhsstr.append(word) ## desymbol here and only here; ump will call quoteattr
if lm is not None:
this = lm.word2index(word)
lmscore += lm.ngram.wordprob(this, lmstr)
lmlhsstr.append(this)
lmstr += [this,]
else: # variable
assert x in forest.nodes, "BAD TOPOL ORDER: node #%s is referred to " % x + \
"(in a hyperedge of node #%s) before being defined" % iden
tail = forest.nodes[x]
tailnodes.append(tail)
lhsstr.append(tail)
if lm is not None:
lmstr = [] # "..." "..." x0 "..."
lmlhsstr.append(tail) # sync with lhsstr
fvector = Vector(fields)
if lm is not None:
fvector["lm1"] = lmscore # hack
edge = Hyperedge(node, tailnodes, fvector, lhsstr)
edge.lmlhsstr = lmlhsstr
## new
x = rule.split()
edge.ruleid = int(x[0])
if len(x) > 1:
edge.rule = Rule.parse(" ".join(x[1:]) + " ### " + fields)
forest.rules[edge.ruleid] = edge.rule #" ".join(x[1:]) #, None)
else:
edge.rule = forest.rules[edge.ruleid] # cahced rule
node.add_edge(edge)
if is_oracle:
node.oracle_edge = edge
if node.sp_terminal():
node.word = node.edges[0].subs[0].word
## splitted nodes 12-3-4 => (12, 3, 4)
tmp = sorted([(map(int, x.iden.split("-")), x) for x in forest.nodeorder])
forest.nodeorder = [x for (_, x) in tmp]
forest.rehash()
sentid += 1
## print >> logs, "sent #%d %s, %d words, %d nodes, %d edges, loaded in %.2lf secs" \
## % (sentid, forest.tag, forest.len, num, forest.num_edges, time.time() - basetime)
forest.root = node
node.set_root(True)
line = file.readline()
if line is not None and line.strip() != "":
if line[0] == "(":
forest.goldtree = Tree.parse(line.strip(), trunc=True, lower=False)
line = file.readline()
else:
line = None
forest.number_nodes()
#print forest.root.position_id
total_time += time.time() - start_time
if num_sents % 100 == 0:
print >> logs, "... %d sents loaded (%.2lf secs per sent) ..." \
% (num_sents, total_time/num_sents)
forest.subtree() #compute the subtree string for each node
yield forest
if first is not None and num_sents >= first:
break
# better check here instead of zero-division exception
if num_sents == 0:
print >> logs, "NO FORESTS FOUND!!! (empty input file?)"
sys.exit(1)
# yield None # new: don't halt -- WHY?
Forest.load_time = total_time
print >> logs, "%d forests loaded in %.2lf secs (avg %.2lf per sent)" \
% (num_sents, total_time, total_time/(num_sents+0.001))