def set_sim(answer, prediction):
ground_truths = []
for a in answer:
ground_truths.append(" ".join([w for w in a]))
res = utils.metric_max_over_ground_truths(utils.f1_score, prediction,
ground_truths)
return res
def validate_with_doc(args, data_loader, model, global_stats,
exs_with_doc, docs_by_question, mode):
'''Run one full unofficial validation with docs.
Unofficial = doesn't use SQuAD script.
'''
eval_time = utils.Timer()
f1 = utils.AverageMeter()
exact_match = utils.AverageMeter()
logger.info('validate_with_doc')
# Intialize counters
examples = 0
aa = [0.0 for i in range(num_docs)] # increment only if example has answer
bb = [0.0 for i in range(num_docs)] # increment regardless
for idx, ex_with_doc in enumerate(data_loader):
ex = ex_with_doc[0]
batch_size, ex_id = ex[0].size(0), ex[-1]
# ---------------------------------------------------------------------
# Document Selector
# ---------------------------------------------------------------------
'''
ex_with_doc =
[tensor] x1 = document word indices [batch * len_d]
[tensor] x1_f = document word features indices [batch * len_d * nfeat]
[tensor] x1_mask = document padding mask [batch * len_d]
[tensor] x2 = question word indices [batch * len_q]
[tensor] x2_mask = question padding mask [batch * len_q]
[list] indices [batch]
'''
scores_doc_num = model.predict_with_doc(ex_with_doc)
scores = [{} for i in range(batch_size)]
# ---------------------------------------------------------------------
# Document Reader
# ---------------------------------------------------------------------
for idx_doc in range(0, num_docs):
ex = ex_with_doc[idx_doc]
pred_s, pred_e, pred_score = model.predict(
ex, top_n=display_num)
for i in range(batch_size):
idx_doc_i = idx_doc %len(docs_by_question[ex_id[i]])
doc_text = docs_by_question[ex_id[i]][idx_doc_i]['document']
# try to read the 10 best predicted answers (this may trigger
# an 'index out of range' exception)
for k in range(display_num):
try:
prediction = [doc_text[j] for j in range(pred_s[i][k],
pred_e[i][k]+1)]
prediction = ' '.join(prediction).lower()
# update prediction scores
if (prediction not in scores[i]):
scores[i][prediction] = 0
scores[i][prediction] += (pred_score[i][k] *
scores_doc_num[i][idx_doc])
except:
pass
# Get the 10 most likely answers from the batch and see if the answer
# is actually in there
for i in range(batch_size):
_, indices = scores_doc_num[i].sort(0, descending = True)
for j in range(0, display_num):
idx_doc = indices[j]
idx_doc_i = idx_doc %len(docs_by_question[ex_id[i]])
doc_text = docs_by_question[ex_id[i]][idx_doc_i]['document']
ex_answer = exs_with_doc[ex_id[i]]['answer']
# Looking for the answer in the document...
if (has_answer(args,
ex_answer,
doc_text)[0]):
aa[j]= aa[j] + 1
bb[j]= bb[j]+1
# Update performance metrics
for i in range(batch_size):
best_score = 0
prediction = ''
for key in scores[i]:
if (scores[i][key] > best_score):
best_score = scores[i][key]
prediction = key
ground_truths = []
ex_answer = exs_with_doc[ex_id[i]]['answer']
# Ground truth answers
if (args.dataset == 'CuratedTrec'): # not applicable
ground_truths = ex_answer
else:
for a in ex_answer:
ground_truths.append(' '.join([w for w in a]))
exact_match.update(
utils.metric_max_over_ground_truths(
utils.exact_match_score, prediction, ground_truths))
f1.update(
utils.metric_max_over_ground_truths(
utils.f1_score, prediction, ground_truths))
examples += batch_size
if (mode=='train' and examples>=1000):
break
try:
for j in range(display_num):
if (j>0):
aa[j]= aa[j]+aa[j-1]
bb[j]= bb[j]+bb[j-1]
except:
pass
txt = '{} valid official with doc: Epoch = {} | EM = {:.2f} | '
txt += 'F1 = {:.2f} | examples = {} | valid time = {:.2f} (s)'
logger.info(txt.format(
mode, global_stats['epoch'], exact_match.avg * 100,
f1.avg * 100, examples, eval_time.time()))
return {'exact_match': exact_match.avg * 100, 'f1': f1.avg * 100}
def validate_unofficial_with_doc(args, data_loader, model, global_stats,
exs_with_doc, docs_by_question, mode):
"""Run one full unofficial validation with docs.
Unofficial = doesn't use SQuAD script.
"""
eval_time = utils.Timer()
f1 = utils.AverageMeter()
exact_match = utils.AverageMeter()
out_set = set({33, 42, 45, 70, 39})
logger.info("validate_unofficial_with_doc")
# Run through examples
examples = 0
aa = [0.0 for i in range(vector.num_docs)]
bb = [0.0 for i in range(vector.num_docs)]
aa_sum = 0.0
display_num = 10
for idx, ex_with_doc in enumerate(data_loader):
ex = ex_with_doc[0]
batch_size, question, ex_id = ex[0].size(0), ex[3], ex[-1]
scores_doc_num = model.predict_with_doc(ex_with_doc)
scores = [{} for i in range(batch_size)]
tot_sum = [0.0 for i in range(batch_size)]
tot_sum1 = [0.0 for i in range(batch_size)]
neg_sum = [0.0 for i in range(batch_size)]
min_sum = [[] for i in range(batch_size)]
min_sum1 = [[] for i in range(batch_size)]
for idx_doc in range(0, vector.num_docs):
ex = ex_with_doc[idx_doc]
pred_s, pred_e, pred_score = model.predict(ex, top_n=10)
for i in range(batch_size):
doc_text = docs_by_question[ex_id[i]][idx_doc % len(
docs_by_question[ex_id[i]])]["document"]
has_answer_t = has_answer(args,
exs_with_doc[ex_id[i]]['answer'],
doc_text)
for k in range(10):
try:
prediction = []
for j in range(pred_s[i][k], pred_e[i][k] + 1):
prediction.append(doc_text[j])
prediction = " ".join(prediction).lower()
if (prediction not in scores[i]):
scores[i][prediction] = 0
scores[i][prediction] += pred_score[i][
k] * scores_doc_num[i][idx_doc]
except:
pass
for i in range(batch_size):
_, indices = scores_doc_num[i].sort(0, descending=True)
for j in range(0, display_num):
idx_doc = indices[j]
doc_text = docs_by_question[ex_id[i]][idx_doc % len(
docs_by_question[ex_id[i]])]["document"]
if (has_answer(args, exs_with_doc[ex_id[i]]['answer'],
doc_text)[0]):
aa[j] = aa[j] + 1
bb[j] = bb[j] + 1
for i in range(batch_size):
best_score = 0
prediction = ""
for key in scores[i]:
if (scores[i][key] > best_score):
best_score = scores[i][key]
prediction = key
# Compute metrics
ground_truths = []
answer = exs_with_doc[ex_id[i]]['answer']
if (args.dataset == "CuratedTrec"):
ground_truths = answer
else:
for a in answer:
ground_truths.append(" ".join([w for w in a]))
#logger.info(prediction)
#logger.info(ground_truths)
exact_match.update(
utils.metric_max_over_ground_truths(utils.exact_match_score,
prediction, ground_truths))
f1.update(
utils.metric_max_over_ground_truths(utils.f1_score, prediction,
ground_truths))
a = sorted(scores[i].items(), key=lambda d: d[1], reverse=True)
examples += batch_size
if (mode == "train" and examples >= 1000):
break
try:
for j in range(0, display_num):
if (j > 0):
aa[j] = aa[j] + aa[j - 1]
bb[j] = bb[j] + bb[j - 1]
logger.info(aa[j] / bb[j])
except:
pass
logger.info(aa_sum)
if (mode == 'dev' or mode == 'train'):
g.write("*" * 50 + "\n")
g.close()
logger.info('%s valid official with doc: Epoch = %d | EM = %.2f | ' %
(mode, global_stats['epoch'], exact_match.avg * 100) +
'F1 = %.2f | examples = %d | valid time = %.2f (s)' %
(f1.avg * 100, examples, eval_time.time()))
return {'exact_match': exact_match.avg * 100, 'f1': f1.avg * 100}
