def get_softmax_margin_partition(factorexprs, goldfactors, valid_fes, sentlen):
logalpha = [None for _ in range(sentlen)]
for j in range(sentlen):
# full length spans
spanscores = []
if not USE_SPAN_CLIP or j <= ALLOWED_SPANLEN:
spanscores = [
factorexprs[Factor(0, j, y)] +
cost(Factor(0, j, y), goldfactors) for y in valid_fes
]
# recursive case
istart = 0
if USE_SPAN_CLIP and j > ALLOWED_SPANLEN:
istart = max(0, j - ALLOWED_SPANLEN - 1)
for i in range(istart, j):
facscores = [
logalpha[i] + factorexprs[Factor(i + 1, j, y)] +
cost(Factor(i + 1, j, y), goldfactors) for y in valid_fes
]
spanscores.extend(facscores)
if not USE_SPAN_CLIP and len(spanscores) != len(valid_fes) * (j + 1):
raise Exception("counting errors")
logalpha[j] = dy.logsumexp(spanscores)
return logalpha[sentlen - 1]
def get_hinge_partition(factorexprs, goldfacs, valid_fes, sentlen):
alpha = [None for _ in range(sentlen)]
backpointers = [None for _ in range(sentlen)]
for j in range(sentlen):
# full length spans
bestscore = float("-inf")
if not USE_SPAN_CLIP or j <= ALLOWED_SPANLEN:
for y in valid_fes:
factor = Factor(0, j, y)
facscore = factorexprs[factor] + cost(factor, goldfacs)
if facscore.scalar_value() > bestscore:
bestscore = facscore.scalar_value()
alpha[j] = facscore
backpointers[j] = (0, y)
# recursive case
istart = 0
if USE_SPAN_CLIP and j > ALLOWED_SPANLEN:
istart = max(0, j - ALLOWED_SPANLEN - 1)
for i in range(istart, j):
for y in valid_fes:
factor = Factor(i + 1, j, y)
facscore = alpha[i] + factorexprs[factor] + cost(
factor, goldfacs)
if facscore.scalar_value() > bestscore:
bestscore = facscore.scalar_value()
alpha[j] = facscore
backpointers[j] = (i + 1, y)
predfactors = []
j = sentlen - 1
i = backpointers[j][0]
while i >= 0:
fe = backpointers[j][1]
predfactors.append(Factor(i, j, fe))
if i == 0:
break
j = i - 1
i = backpointers[j][0]
return alpha[sentlen - 1], predfactors
def get_factor_expressions(fws,
bws,
tfemb,
tfdict,
valid_fes,
sentence,
spaths_x=None,
cpaths_x=None):
factexprs = {}
sentlen = len(fws)
sortedtfd = sorted(list(tfdict.keys()))
targetspan = (sortedtfd[0], sortedtfd[-1])
for j in range(sentlen):
istart = 0
if USE_SPAN_CLIP and j > ALLOWED_SPANLEN:
istart = max(0, j - ALLOWED_SPANLEN)
for i in range(istart, j + 1):
spanlen = dy.scalarInput(j - i + 1)
logspanlen = dy.scalarInput(math.log(j - i + 1))
spanwidth = sp_x[SpanWidth.howlongisspan(i, j)]
spanpos = ap_x[ArgPosition.whereisarg((i, j), targetspan)]
fbemb_ij_basic = dy.concatenate([
fws[i][j], bws[i][j], tfemb, spanlen, logspanlen, spanwidth,
spanpos
])
if USE_DEPS:
outs = oh_s[OutHeads.getnumouts(i, j, sentence.outheads)]
shp = spaths_x[sentence.shortest_paths[(i, j, targetspan[0])]]
fbemb_ij = dy.concatenate([fbemb_ij_basic, outs, shp])
elif USE_CONSTITS:
isconstit = dy.scalarInput((i, j) in sentence.constitspans)
lca = ct_x[sentence.lca[(i, j)][1]]
phrp = cpaths_x[sentence.cpaths[(i, j, targetspan[0])]]
fbemb_ij = dy.concatenate(
[fbemb_ij_basic, isconstit, lca, phrp])
else:
fbemb_ij = fbemb_ij_basic
for y in valid_fes:
fctr = Factor(i, j, y)
if USE_HIER and y in feparents:
fefixed = dy.esum([fe_x[y]] +
[fe_x[par] for par in feparents[y]])
else:
fefixed = fe_x[y]
fbemb_ijy = dy.concatenate([fefixed, fbemb_ij])
factexprs[fctr] = w_f * dy.rectify(w_z * fbemb_ijy + b_z) + b_f
return factexprs
def get_all_span_reduced_bert_embeddings(examples):
my_dict = {}
pca_list = []
for trex in examples:
reg_sentence = get_sentence_from_conll_example(trex)
senlen = len(reg_sentence.split(" "))
for j in range(senlen):
istart = 0
if USE_SPAN_CLIP and j > ALLOWED_SPANLEN: istart = max(0, j - ALLOWED_SPANLEN)
for i in range(istart, j + 1):
# Get span expression through BERT
fbemb_ij = get_span_bert_expression(reg_sentence, i, j)
pca_list.append(fbemb_ij)
if USE_PCA:
pca = PCA(435)
new_matrix = pca.fit_transform(pca_list)
else:
new_matrix = transform_auto_encoder(pca_list)
index_keeper = 0
for trex in examples:
# get valid fes per example
tfdict = trex.targetframedict
tfkeys = sorted(tfdict)
tg_start = tfkeys[0]
lu, frame = tfdict[tg_start]
valid_fes = frmfemap[frame.id] + [NOTANFEID]
factexprs = {}
reg_sentence = get_sentence_from_conll_example(trex)
senlen = len(reg_sentence.split(" "))
for j in range(senlen):
istart = 0
if USE_SPAN_CLIP and j > ALLOWED_SPANLEN: istart = max(0, j - ALLOWED_SPANLEN)
for i in range(istart, j + 1):
for y in valid_fes:
fctr = Factor(i, j, y)
if USE_HIER and y in feparents:
fefixed = dy.esum([fe_x[y]] + [fe_x[par] for par in feparents[y]])
else:
fefixed = fe_x[y]
fbemb_ij = dy.inputTensor(new_matrix[index_keeper])
fbemb_ijy = dy.concatenate([fefixed, fbemb_ij])
factexprs[fctr] = w_f * dy.rectify(w_z * fbemb_ijy + b_z) + b_f
index_keeper = index_keeper + 1
my_dict[trex] = factexprs
return my_dict
def get_loss(factorexprs, gold_fes, valid_fes, sentlen):
if options.loss == "hinge":
return get_hinge_loss(factorexprs, gold_fes, valid_fes, sentlen)
goldfactors = [
Factor(span[0], span[1], feid) for feid in gold_fes
for span in gold_fes[feid]
]
numeratorexprs = [factorexprs[gf] for gf in goldfactors]
numerator = dy.esum(numeratorexprs)
if options.loss == "log":
partition = get_logloss_partition(factorexprs, valid_fes, sentlen)
elif options.loss == "softmaxm":
partition = get_softmax_margin_partition(factorexprs, goldfactors,
valid_fes, sentlen)
else:
raise Exception("undefined loss function", options.loss)
lossexp = partition - numerator
if partition.scalar_value() < numerator.scalar_value():
sys.stderr.write(
"WARNING: partition ~~ numerator! possibly overfitting difference = %f\n"
% lossexp.scalar_value())
return None
if lossexp.scalar_value() < 0.0:
sys.stderr.write(str(gold_fes) + "\ngolds\n")
gsum = 0
for fac in goldfactors:
gsum += factorexprs[fac].scalar_value()
sys.stderr.write(
fac.to_str(FEDICT) + " " +
str(factorexprs[fac].scalar_value()) + "\n")
sys.stderr.write("my calculation = " + str(gsum) + " vs " +
str(numerator.scalar_value()) + "\n")
for j in range(sentlen):
sys.stderr.write(":" + str(j) + "\t")
if not USE_SPAN_CLIP or j <= ALLOWED_SPANLEN:
sys.stderr.write("0 ")
istart = 0
if USE_SPAN_CLIP and j > ALLOWED_SPANLEN:
istart = max(0, j - ALLOWED_SPANLEN - 1)
for i in range(istart, j):
sys.stderr.write(str(i + 1) + " ")
sys.stderr.write("\n")
raise Exception("negative probability! probably overcounting spans?",
numerator.scalar_value(), partition.scalar_value(),
lossexp.scalar_value())
return lossexp
def get_hinge_loss(factorexprs, gold_fes, valid_fes, sentlen):
goldfactors = [Factor(span[0], span[1], feid) for feid in gold_fes for span in gold_fes[feid]]
numeratorexprs = [factorexprs[gf] for gf in goldfactors]
numerator = dy.esum(numeratorexprs)
denominator, predfactors = get_hinge_partition(factorexprs, goldfactors, valid_fes, sentlen)
if set(predfactors) == set(goldfactors):
return None
hingeloss = denominator - numerator
if denominator.scalar_value() < numerator.scalar_value():
raise Exception("ERROR: predicted cost less than gold!",
denominator.scalar_value(),
numerator.scalar_value(),
hingeloss.scalar_value())
return hingeloss
def decode(factexprscalars, sentlen, valid_fes):
alpha = [None for _ in range(sentlen)]
backpointers = [None for _ in range(sentlen)]
if USE_DROPOUT:
raise Exception("incorrect usage of dropout, turn off!\n")
for j in range(sentlen):
if USE_SPAN_CLIP and j > ALLOWED_SPANLEN: continue
bestscore = float("-inf")
bestlabel = None
for y in valid_fes:
fac = Factor(0, j, y)
facscore = math.exp(factexprscalars[fac])
if facscore > bestscore:
bestscore = facscore
bestlabel = y
alpha[j] = bestscore
backpointers[j] = (0, bestlabel)
for j in range(sentlen):
bestscore = float("-inf")
bestbeg = bestlabel = None
if alpha[j] is not None:
bestscore = alpha[j]
bestbeg, bestlabel = backpointers[j]
istart = 0
if USE_SPAN_CLIP and j > ALLOWED_SPANLEN: istart = max(0, j - ALLOWED_SPANLEN - 1)
for i in range(istart, j):
for y in valid_fes:
fac = Factor(i + 1, j, y)
facscore = math.exp(factexprscalars[fac])
if facscore * alpha[i] > bestscore:
bestscore = facscore * alpha[i]
bestlabel = y
bestbeg = i + 1
alpha[j] = bestscore
backpointers[j] = (bestbeg, bestlabel)
j = sentlen - 1
i = backpointers[j][0]
argmax = {}
while i >= 0:
fe = backpointers[j][1]
if fe in argmax:
argmax[fe].append((i, j))
else:
argmax[fe] = [(i, j)]
if i == 0:
break
j = i - 1
i = backpointers[j][0]
# merging neighboring spans in prediction (to combat spurious ambiguity)
mergedargmax = {}
for fe in argmax:
mergedargmax[fe] = []
if fe == NOTANFEID:
mergedargmax[fe].extend(argmax[fe])
continue
argmax[fe].sort()
mergedspans = [argmax[fe][0]]
for span in argmax[fe][1:]:
prevsp = mergedspans[-1]
if span[0] == prevsp[1] + 1:
prevsp = mergedspans.pop()
mergedspans.append((prevsp[0], span[1]))
else:
mergedspans.append(span)
mergedargmax[fe] = mergedspans
return mergedargmax
sys.stderr.write("Ran out of patience, ending training.\n")
sys.stderr.write("Best model evaluation:\n{}\n".format(best_dev_eval_str))
sys.stderr.write("Best model saved to {}\n".format(model_file_name))
sys.stderr.write(" [took %.3fs]\n" % (time.time() - starttime))
break
loss = 0.0
elif options.mode == "ensemble":
exfs = {x: {} for x in range(len(devexamples))}
USE_DROPOUT = False
sys.stderr.write("reading ensemble factors...")
enf = open(in_ens_file, "rb")
for l in enf:
fields = l.split("\t")
fac = Factor(int(fields[1]), int(fields[2]), FEDICT.getid(fields[3]))
exfs[int(fields[0])][fac] = float(fields[4])
enf.close()
sys.stderr.write("done!\n")
teststarttime = time.time()
sys.stderr.write("testing " + str(len(devexamples)) + " examples ...\n")
testpredictions = []
for tidx, testex in enumerate(devexamples, 1):
if tidx % 100 == 0:
sys.stderr.write(str(tidx) + "...")
valid_fes_for_frame = frmfemap[testex.frame.id] + [NOTANFEID]
testargmax = decode(exfs[tidx - 1], len(testex.tokens), valid_fes_for_frame)
testpredictions.append(testargmax)
