def show_web_paragraphs():
splitter = MergeParagraphs(400)
stop = NltkPlusStopWords(True)
ranker = TopTfIdf(stop, 6)
stop_words = stop.words
corpus = TriviaQaWebDataset()
train = corpus.get_train()
points = flatten_iterable([(q, d) for d in q.all_docs] for q in train)
np.random.shuffle(points)
for q, d in points:
q_words = {strip_accents_unicode(w.lower()) for w in q.question}
q_words = {x for x in q_words if x not in stop_words}
doc = corpus.evidence.get_document(d.doc_id)
doc = splitter.split_annotated(doc, d.answer_spans)
ranked = ranker.dists(q.question, doc)
if len(ranked) < 2 or len(ranked[1][0].answer_spans) == 0:
continue
print(" ".join(q.question))
print(q.answer.all_answers)
for i, (para, dist) in enumerate(ranked[0:2]):
text = flatten_iterable(para.text)
print("Start=%d, Rank=%d, Dist=%.4f" % (para.start, i, dist))
if len(para.answer_spans) == 0:
continue
for s, e in para.answer_spans:
text[s] = bcolors.CYAN + text[s]
text[e] = text[e] + bcolors.ENDC
for i, w in enumerate(text):
if strip_accents_unicode(w.lower()) in q_words:
text[i] = bcolors.ERROR + text[i] + bcolors.ENDC
print(" ".join(text))
input()
def main():
data = TriviaQaWebDataset()
stop = NltkPlusStopWords()
splitter = MergeParagraphs(400)
selector = TopTfIdf(stop, 4)
print("Loading data..")
train = data.get_train()
print("Start")
for q in train:
for doc in q.all_docs:
if len(doc.answer_spans) > 3:
text = splitter.split_annotated(
data.evidence.get_document(doc.doc_id), doc.answer_spans)
text = selector.prune(q.question, text)
for para in text:
if len(para.answer_spans) > 3:
print(q.question)
text = flatten_iterable(para.text)
for s, e in para.answer_spans:
text[s] = "{{{" + text[s]
text[e] = text[e] + "}}}"
print(" ".join(text))
input()
def check_preprocess():
data = TriviaQaWebDataset()
merge = MergeParagraphs(400)
questions = data.get_dev()
pre = WithIndicators(False)
remove_cross = WithIndicators(True)
rng = np.random.RandomState(0)
rng.shuffle(questions)
for q in tqdm(questions[:1000]):
doc = rng.choice(q.all_docs, 1)[0]
text = data.evidence.get_document(doc.doc_id, n_tokens=800)
paras = merge.split_annotated(text, doc.answer_spans)
para = paras[np.random.randint(0, len(paras))]
built = pre.encode_extracted_paragraph(q.question, para)
expected_text = flatten_iterable(para.text)
if expected_text != [
x for x in built.text if x not in pre.special_tokens()
]:
raise ValueError()
expected = [expected_text[s:e + 1] for s, e in para.answer_spans]
expected = Counter([tuple(x) for x in expected])
actual = [tuple(built.text[s:e + 1]) for s, e in built.answer_spans]
actual_cleaned = Counter(
tuple(z for z in x if z not in pre.special_tokens())
for x in actual)
if actual_cleaned != expected:
raise ValueError()
r_built = remove_cross.encode_extracted_paragraph(q.question, para)
rc = Counter(
tuple(r_built.text[s:e + 1]) for s, e in r_built.answer_spans)
removed = Counter()
for w in actual:
if all(x not in pre.special_tokens() for x in w):
removed[w] += 1
if rc != removed:
raise ValueError()
def show_stats():
splitter = MergeParagraphs(400)
stop = NltkPlusStopWords(True)
ranker = TopTfIdf(stop, 6)
corpus = TriviaQaWebDataset()
train = corpus.get_train()
points = flatten_iterable([(q, d) for d in q.all_docs] for q in train)
np.random.shuffle(points)
counts = np.zeros(6)
answers = np.zeros(6)
n_answers = []
points = points[:1000]
for q, d in tqdm(points):
doc = corpus.evidence.get_document(d.doc_id)
doc = splitter.split_annotated(doc, d.answer_spans)
ranked = ranker.prune(q.question, doc)
counts[:len(ranked)] += 1
for i, para in enumerate(ranked):
if len(para.answer_spans) > 0:
answers[i] += 1
n_answers.append(
tuple(i for i, x in enumerate(ranked) if len(x.answer_spans) > 0))
print(answers / counts)
c = Counter()
other = 0
for tup in n_answers:
if len(tup) <= 2:
c[tup] += 1
else:
other += 1
for p in sorted(c.keys()):
print(p, c.get(p) / len(points))
print(other / len(points))
def contains_question_word():
data = TriviaQaWebDataset()
stop = NltkPlusStopWords(punctuation=True).words
doc_filter = ContainsQuestionWord(NltkPlusStopWords(punctuation=True))
splits = MergeParagraphs(400)
# splits = Truncate(400)
questions = data.get_dev()
pairs = flatten_iterable([(q, doc) for doc in q.all_docs]
for q in questions)
pairs.sort(key=lambda x: (x[0].question_id, x[1].doc_id))
np.random.RandomState(0).shuffle(questions)
has_token = 0
total = 0
used = Counter()
for q, doc in tqdm(pairs[:1000]):
text = data.evidence.get_document(doc.doc_id, splits.reads_first_n)
q_tokens = set(x.lower() for x in q.question)
q_tokens -= stop
for para in splits.split_annotated(text, doc.answer_spans):
# if para.start == 0:
# continue
if len(para.answer_spans) == 0:
continue
if any(x.lower() in q_tokens for x in flatten_iterable(para.text)):
has_token += 1
for x in flatten_iterable(para.text):
if x in q_tokens:
used[x] += 1
# else:
# print_questions(q.question, q.answer.all_answers, para.text, para.answer_spans)
# input()
total += 1
for k, v in used.most_common(200):
print("%s: %d" % (k, v))
print(has_token / total)
def main():
parser = argparse.ArgumentParser(description='Train a model on TriviaQA web')
parser.add_argument('mode', choices=["paragraph-level", "confidence", "merge",
"shared-norm", "sigmoid", "shared-norm-600"])
parser.add_argument("name", help="Where to store the model")
parser.add_argument('-n', '--n_processes', type=int, default=2,
help="Number of processes (i.e., select which paragraphs to train on) "
"the data with")
args = parser.parse_args()
mode = args.mode
out = args.name + "-" + datetime.now().strftime("%m%d-%H%M%S")
model = get_model(100, 140, mode, WithIndicators())
stop = NltkPlusStopWords(True)
if mode == "paragraph-level":
extract = ExtractSingleParagraph(MergeParagraphs(400), TopTfIdf(stop, 1), model.preprocessor, intern=True)
elif mode == "shared-norm-600":
extract = ExtractMultiParagraphs(MergeParagraphs(600), TopTfIdf(stop, 4), model.preprocessor, intern=True)
else:
extract = ExtractMultiParagraphs(MergeParagraphs(400), TopTfIdf(stop, 4), model.preprocessor, intern=True)
if mode == "paragraph-level":
n_epochs = 16
train = ParagraphAndQuestionsBuilder(ClusteredBatcher(60, ContextLenBucketedKey(3), True))
test = ParagraphAndQuestionsBuilder(ClusteredBatcher(60, ContextLenKey(), False))
n_dev, n_train = 21000, 12000
eval = [LossEvaluator(), SpanEvaluator([4, 8], "triviaqa")]
else:
eval = [LossEvaluator(), MultiParagraphSpanEvaluator(8, "triviaqa", mode != "merge")]
# we sample two paragraphs per a (question, doc) pair, so evaluate on fewer questions
n_dev, n_train = 15000, 8000
if mode == "confidence" or mode == "sigmoid":
if mode == "sigmoid":
# Trains very slowly, do this at your own risk
n_epochs = 71
else:
n_epochs = 28
test = RandomParagraphSetDatasetBuilder(120, "flatten", True, 1)
train = StratifyParagraphsBuilder(ClusteredBatcher(60, ContextLenBucketedKey(3), True), 0, 1)
else:
n_epochs = 14
test = RandomParagraphSetDatasetBuilder(120, "merge" if mode == "merge" else "group", True, 1)
train = StratifyParagraphSetsBuilder(35, mode == "merge", True, 1)
data = TriviaQaWebDataset()
params = get_triviaqa_train_params(n_epochs, n_dev, n_train)
data = PreprocessedData(data, extract, train, test, eval_on_verified=False)
data.preprocess(args.n_processes, 1000)
with open(__file__, "r") as f:
notes = f.read()
notes = "*" * 10 + "\nMode: " + args.mode + "\n" + "*"*10 + "\n" + notes
trainer.start_training(data, model, params, eval, model_dir.ModelDir(out), notes)
def main():
parser = argparse.ArgumentParser(description='Evaluate a model on TriviaQA data')
parser.add_argument('model', help='model directory')
parser.add_argument('-p', '--paragraph_output', type=str,
help="Save fine grained results for each paragraph in csv format")
parser.add_argument('-o', '--official_output', type=str, help="Build an offical output file with the model's"
" most confident span for each (question, doc) pair")
parser.add_argument('--no_ema', action="store_true", help="Don't use EMA weights even if they exist")
parser.add_argument('--n_processes', type=int, default=None,
help="Number of processes to do the preprocessing (selecting paragraphs+loading context) with")
parser.add_argument('-i', '--step', type=int, default=None, help="checkpoint to load, default to latest")
parser.add_argument('-n', '--n_sample', type=int, default=None, help="Number of questions to evaluate on")
parser.add_argument('-a', '--async', type=int, default=10)
parser.add_argument('-t', '--tokens', type=int, default=400,
help="Max tokens per a paragraph")
parser.add_argument('-g', '--n_paragraphs', type=int, default=15,
help="Number of paragraphs to run the model on")
parser.add_argument('-f', '--filter', type=str, default=None, choices=["tfidf", "truncate", "linear"],
help="How to select paragraphs")
parser.add_argument('-b', '--batch_size', type=int, default=200,
help="Batch size, larger sizes might be faster but wll take more memory")
parser.add_argument('--max_answer_len', type=int, default=8,
help="Max answer span to select")
parser.add_argument('-c', '--corpus',
choices=["web-dev", "web-test", "web-verified-dev", "web-train",
"open-dev", "open-train"],
default="web-verified-dev")
args = parser.parse_args()
model_dir = ModelDir(args.model)
model = model_dir.get_model()
if args.corpus.startswith('web'):
dataset = TriviaQaWebDataset()
corpus = dataset.evidence
if args.corpus == "web-dev":
test_questions = dataset.get_dev()
elif args.corpus == "web-test":
test_questions = dataset.get_test()
elif args.corpus == "web-verified-dev":
test_questions = dataset.get_verified()
elif args.corpus == "web-train":
test_questions = dataset.get_train()
else:
raise RuntimeError()
else:
dataset = TriviaQaOpenDataset()
corpus = dataset.evidence
if args.corpus == "open-dev":
test_questions = dataset.get_dev()
elif args.corpus == "open-train":
test_questions = dataset.get_train()
else:
raise RuntimeError()
splitter = MergeParagraphs(args.tokens)
per_document = not args.corpus.startswith("open")
filter_name = args.filter
if filter_name is None:
if args.corpus.startswith("open"):
filter_name = "linear"
else:
filter_name = "tfidf"
print("Selecting %d paragraphs using %s method per %s" % (args.n_paragraphs, filter_name,
("question-document pair" if per_document else "question")))
if filter_name == "tfidf":
para_filter = TopTfIdf(NltkPlusStopWords(punctuation=True), args.n_paragraphs)
elif filter_name == "truncate":
para_filter = FirstN(args.n_paragraphs)
elif filter_name == "linear":
para_filter = ShallowOpenWebRanker(args.n_paragraphs)
else:
raise ValueError()
n_questions = args.n_sample
if n_questions is not None:
test_questions.sort(key=lambda x:x.question_id)
np.random.RandomState(0).shuffle(test_questions)
test_questions = test_questions[:n_questions]
print("Building question/paragraph pairs...")
# Loads the relevant questions/documents, selects the right paragraphs, and runs the model's preprocessor
if per_document:
prep = ExtractMultiParagraphs(splitter, para_filter, model.preprocessor, require_an_answer=False)
else:
prep = ExtractMultiParagraphsPerQuestion(splitter, para_filter, model.preprocessor, require_an_answer=False)
prepped_data = preprocess_par(test_questions, corpus, prep, args.n_processes, 1000)
data = []
for q in prepped_data.data:
for i, p in enumerate(q.paragraphs):
if q.answer_text is None:
ans = None
else:
ans = TokenSpans(q.answer_text, p.answer_spans)
data.append(DocumentParagraphQuestion(q.question_id, p.doc_id,
(p.start, p.end), q.question, p.text,
ans, i))
# Reverse so our first batch will be the largest (so OOMs happen early)
questions = sorted(data, key=lambda x: (x.n_context_words, len(x.question)), reverse=True)
print("Done, starting eval")
if args.step is not None:
if args.step == "latest":
checkpoint = model_dir.get_latest_checkpoint()
else:
checkpoint = model_dir.get_checkpoint(int(args.step))
else:
checkpoint = model_dir.get_best_weights()
if checkpoint is not None:
print("Using best weights")
else:
print("Using latest checkpoint")
checkpoint = model_dir.get_latest_checkpoint()
test_questions = ParagraphAndQuestionDataset(questions, FixedOrderBatcher(args.batch_size, True))
evaluation = trainer.test(model,
[RecordParagraphSpanPrediction(args.max_answer_len, True)],
{args.corpus:test_questions}, ResourceLoader(), checkpoint, not args.no_ema, args.async)[args.corpus]
if not all(len(x) == len(data) for x in evaluation.per_sample.values()):
raise RuntimeError()
df = pd.DataFrame(evaluation.per_sample)
if args.official_output is not None:
print("Saving question result")
# I didn't store the unormalized filenames exactly, so unfortunately we have to reload
# the source data to get exact filename to output an official test script
fns = {}
print("Loading proper filenames")
if args.corpus == 'web-test':
source = join(TRIVIA_QA, "qa", "web-test-without-answers.json")
elif args.corpus == "web-dev":
source = join(TRIVIA_QA, "qa", "web-dev.json")
else:
raise NotImplementedError()
with open(join(source)) as f:
data = json.load(f)["Data"]
for point in data:
for doc in point["EntityPages"]:
filename = doc["Filename"]
fn = join("wikipedia", filename[:filename.rfind(".")])
fn = normalize_wiki_filename(fn)
fns[(point["QuestionId"], fn)] = filename
answers = {}
scores = {}
for q_id, doc_id, start, end, txt, score in df[["question_id", "doc_id", "para_start", "para_end",
"text_answer", "predicted_score"]].itertuples(index=False):
filename = dataset.evidence.file_id_map[doc_id]
if filename.startswith("web"):
true_name = filename[4:] + ".txt"
else:
true_name = fns[(q_id, filename)]
key = q_id + "--" + true_name
prev_score = scores.get(key)
if prev_score is None or prev_score < score:
scores[key] = score
answers[key] = txt
with open(args.official_output, "w") as f:
json.dump(answers, f)
if per_document:
group_by = ["question_id", "doc_id"]
else:
group_by = ["question_id"]
# Print a table of scores as more paragraphs are used
df.sort_values(group_by + ["rank"], inplace=True)
f1 = compute_model_scores(df, "predicted_score", "text_f1", group_by)
em = compute_model_scores(df, "predicted_score", "text_em", group_by)
table = [["N Paragraphs", "EM", "F1"]]
table += list([str(i+1), "%.4f" % e, "%.4f" % f] for i, (e, f) in enumerate(zip(em, f1)))
print_table(table)
output_file = args.paragraph_output
if output_file is not None:
print("Saving paragraph result")
if output_file.endswith("json"):
with open(output_file, "w") as f:
json.dump(evaluation.per_sample, f)
elif output_file.endswith("pkl"):
with open(output_file, "wb") as f:
pickle.dump(evaluation.per_sample, f)
elif output_file.endswith("csv"):
df.to_csv(output_file, index=False)
else:
raise ValueError("Unrecognized file format")
np.random.RandomState(0).shuffle(questions)
has_token = 0
total = 0
used = Counter()
for q, doc in tqdm(pairs[:1000]):
text = data.evidence.get_document(doc.doc_id, splits.reads_first_n)
q_tokens = set(x.lower() for x in q.question)
q_tokens -= stop
for para in splits.split_annotated(text, doc.answer_spans):
# if para.start == 0:
# continue
if len(para.answer_spans) == 0:
continue
if any(x.lower() in q_tokens for x in flatten_iterable(para.text)):
has_token += 1
for x in flatten_iterable(para.text):
if x in q_tokens:
used[x] += 1
# else:
# print_questions(q.question, q.answer.all_answers, para.text, para.answer_spans)
# input()
total += 1
for k, v in used.most_common(200):
print("%s: %d" % (k, v))
print(has_token / total)
if __name__ == "__main__":
paragraph_stats(TriviaQaWebDataset(), MergeParagraphs(400), 1000)
def main():
parser = argparse.ArgumentParser(
description='Evaluate a model on TriviaQA data')
parser.add_argument('model', help='model directory')
parser.add_argument(
'-p',
'--paragraph_output',
type=str,
help="Save fine grained results for each paragraph in csv format")
parser.add_argument('-o',
'--official_output',
type=str,
help="Build an offical output file with the model's"
" most confident span for each (question, doc) pair")
parser.add_argument('--no_ema',
action="store_true",
help="Don't use EMA weights even if they exist")
parser.add_argument(
'--n_processes',
type=int,
default=None,
help=
"Number of processes to do the preprocessing (selecting paragraphs+loading context) with"
)
parser.add_argument('-i',
'--step',
type=int,
default=None,
help="checkpoint to load, default to latest")
parser.add_argument('-n',
'--n_sample',
type=int,
default=None,
help="Number of questions to evaluate on")
parser.add_argument('-a', '--async', type=int, default=10)
parser.add_argument('-t',
'--tokens',
type=int,
default=400,
help="Max tokens per a paragraph")
parser.add_argument('-g',
'--n_paragraphs',
type=int,
default=15,
help="Number of paragraphs to run the model on")
parser.add_argument('-f',
'--filter',
type=str,
default=None,
choices=["tfidf", "truncate", "linear"],
help="How to select paragraphs")
parser.add_argument(
'-b',
'--batch_size',
type=int,
default=200,
help="Batch size, larger sizes might be faster but wll take more memory"
)
parser.add_argument('--max_answer_len',
type=int,
default=8,
help="Max answer span to select")
parser.add_argument('-c',
'--corpus',
choices=[
"web-dev", "web-test", "web-verified-dev",
"web-train", "open-dev", "open-train", "wiki-dev",
"wiki-test"
],
default="web-verified-dev")
parser.add_argument("-s",
"--source_dir",
type=str,
default=None,
help="where to take input files")
parser.add_argument("--n_span_per_q",
type=int,
default=1,
help="where to take input files")
args = parser.parse_args()
dataset_name = args.source_dir.split('/')[-1]
model_name = args.model.split('/')[-1]
ElasticLogger().write_log('INFO',
'Start Evaluation',
context_dict={
'model': model_name,
'dataset': dataset_name
})
model_dir = ModelDir(args.model)
model = model_dir.get_model()
if args.corpus.startswith('web'):
dataset = TriviaQaWebDataset()
if args.corpus == "web-dev":
test_questions = dataset.get_dev()
elif args.corpus == "web-test":
test_questions = dataset.get_test()
elif args.corpus == "web-verified-dev":
test_questions = dataset.get_verified()
elif args.corpus == "web-train":
test_questions = dataset.get_train()
else:
raise AssertionError()
elif args.corpus.startswith("wiki"):
dataset = TriviaQaWikiDataset()
if args.corpus == "wiki-dev":
test_questions = dataset.get_dev()
elif args.corpus == "wiki-test":
test_questions = dataset.get_test()
else:
raise AssertionError()
else:
dataset = TriviaQaOpenDataset(args.source_dir)
if args.corpus == "open-dev":
# just loading the pkl that was saved in build_span_corpus
test_questions = dataset.get_dev()
elif args.corpus == "open-train":
test_questions = dataset.get_train()
else:
raise AssertionError()
### ALON debuging
#test_questions = test_questions[0:5]
corpus = dataset.evidence
splitter = MergeParagraphs(args.tokens)
per_document = args.corpus.startswith(
"web") # wiki and web are both multi-document
#per_document = True
filter_name = args.filter
if filter_name is None:
# Pick default depending on the kind of data we are using
if per_document:
filter_name = "tfidf"
else:
filter_name = "linear"
print("Selecting %d paragraphs using method \"%s\" per %s" %
(args.n_paragraphs, filter_name,
("question-document pair" if per_document else "question")))
if filter_name == "tfidf":
para_filter = TopTfIdf(NltkPlusStopWords(punctuation=True),
args.n_paragraphs)
elif filter_name == "truncate":
para_filter = FirstN(args.n_paragraphs)
elif filter_name == "linear":
para_filter = ShallowOpenWebRanker(args.n_paragraphs)
else:
raise ValueError()
n_questions = args.n_sample
docqa.config.SPANS_PER_QUESTION = args.n_span_per_q
#n_questions = 1
if n_questions is not None:
test_questions.sort(key=lambda x: x.question_id)
np.random.RandomState(0).shuffle(test_questions)
test_questions = test_questions[:n_questions]
print("Building question/paragraph pairs...")
# Loads the relevant questions/documents, selects the right paragraphs, and runs the model's preprocessor
if per_document:
prep = ExtractMultiParagraphs(splitter,
para_filter,
model.preprocessor,
require_an_answer=False)
else:
prep = ExtractMultiParagraphsPerQuestion(splitter,
para_filter,
model.preprocessor,
require_an_answer=False)
prepped_data = preprocess_par(test_questions, corpus, prep,
args.n_processes, 1000)
data = []
for q in prepped_data.data:
for i, p in enumerate(q.paragraphs):
if q.answer_text is None:
ans = None
else:
ans = TokenSpans(q.answer_text, p.answer_spans)
data.append(
DocumentParagraphQuestion(q.question_id, p.doc_id,
(p.start, p.end), q.question, p.text,
ans, i))
# Reverse so our first batch will be the largest (so OOMs happen early)
questions = sorted(data,
key=lambda x: (x.n_context_words, len(x.question)),
reverse=True)
print("Done, starting eval")
if args.step is not None:
if args.step == "latest":
checkpoint = model_dir.get_latest_checkpoint()
else:
checkpoint = model_dir.get_checkpoint(int(args.step))
else:
checkpoint = model_dir.get_best_weights()
if checkpoint is not None:
print("Using best weights")
else:
print("Using latest checkpoint")
checkpoint = model_dir.get_latest_checkpoint()
test_questions = ParagraphAndQuestionDataset(
questions, FixedOrderBatcher(args.batch_size, True))
evaluation = trainer.test(
model, [RecordParagraphSpanPrediction(args.max_answer_len, True)],
{args.corpus: test_questions}, ResourceLoader(), checkpoint,
not args.no_ema, args. async)[args.corpus]
if not all(len(x) == len(data) for x in evaluation.per_sample.values()):
raise RuntimeError()
df = pd.DataFrame(evaluation.per_sample)
if args.official_output is not None:
print("Saving question result")
fns = {}
if per_document:
# I didn't store the unormalized filenames exactly, so unfortunately we have to reload
# the source data to get exact filename to output an official test script
print("Loading proper filenames")
if args.corpus == 'web-test':
source = join(TRIVIA_QA, "qa", "web-test-without-answers.json")
elif args.corpus == "web-dev":
source = join(TRIVIA_QA, "qa", "web-dev.json")
else:
raise AssertionError()
with open(join(source)) as f:
data = json.load(f)["Data"]
for point in data:
for doc in point["EntityPages"]:
filename = doc["Filename"]
fn = join("wikipedia", filename[:filename.rfind(".")])
fn = normalize_wiki_filename(fn)
fns[(point["QuestionId"], fn)] = filename
answers = {}
scores = {}
for q_id, doc_id, start, end, txt, score in df[[
"question_id", "doc_id", "para_start", "para_end",
"text_answer", "predicted_score"
]].itertuples(index=False):
filename = dataset.evidence.file_id_map[doc_id]
if per_document:
if filename.startswith("web"):
true_name = filename[4:] + ".txt"
else:
true_name = fns[(q_id, filename)]
# Alon Patch for triviaqa test results
true_name = true_name.replace('TriviaQA_Org/', '')
key = q_id + "--" + true_name
else:
key = q_id
prev_score = scores.get(key)
if prev_score is None or prev_score < score:
scores[key] = score
answers[key] = txt
with open(args.official_output, "w") as f:
json.dump(answers, f)
output_file = args.paragraph_output
if output_file is not None:
print("Saving paragraph result")
df.to_csv(output_file, index=False)
print("Computing scores")
if per_document:
group_by = ["question_id", "doc_id"]
else:
group_by = ["question_id"]
# Print a table of scores as more paragraphs are used
df.sort_values(group_by + ["rank"], inplace=True)
df_scores = df.copy(deep=True)
df_scores['predicted_score'] = df_scores['predicted_score'].apply(
lambda x: pd.Series(x).max())
em = compute_ranked_scores(df_scores, "predicted_score", "text_em",
group_by)
f1 = compute_ranked_scores(df_scores, "predicted_score", "text_f1",
group_by)
table = [["N Paragraphs", "EM", "F1"]]
table += list([str(i + 1), "%.4f" % e, "%.4f" % f]
for i, (e, f) in enumerate(zip(em, f1)))
table_df = pd.DataFrame(table[1:], columns=table[0]).drop(['N Paragraphs'],
axis=1)
ElasticLogger().write_log('INFO', 'Results', context_dict={'model': model_name, 'dataset': dataset_name, \
'max_EM':table_df.max().ix['EM'], \
'max_F1':table_df.max().ix['F1'], \
'result_table': str(table_df)})
df_flat = []
for id, question in df.iterrows():
for text_answer, predicted_span, predicted_score in zip(
question['text_answer'], question['predicted_span'],
question['predicted_score']):
new_question = dict(question.copy())
new_question.update({
'text_answer': text_answer,
'predicted_span': predicted_span,
'predicted_score': predicted_score
})
df_flat.append(new_question)
results_df = pd.DataFrame(df_flat)
#Alon: outputing the estimates for all the
#results_df = results_df.groupby(['question_id', 'text_answer']).apply(lambda df: df.ix[df['predicted_score'].argmax()]).reset_index(drop=True)
results_df.sort_values(by=['question_id', 'predicted_score'],
ascending=False).set_index([
'question_id', 'text_answer'
])[['question', 'predicted_score',
'text_em']].to_csv('results.csv')
print_table(table)
def main():
parser = argparse.ArgumentParser(
description='Train a model on TriviaQA unfiltered')
parser.add_argument(
'mode',
choices=["confidence", "merge", "shared-norm", "sigmoid", "paragraph"])
parser.add_argument("name", help="Where to store the model")
parser.add_argument("-t",
"--n_tokens",
default=400,
type=int,
help="Paragraph size")
parser.add_argument(
'-n',
'--n_processes',
type=int,
default=2,
help="Number of processes (i.e., select which paragraphs to train on) "
"the data with")
args = parser.parse_args()
mode = args.mode
out = args.name + "-" + datetime.now().strftime("%m%d-%H%M%S")
model = get_model(100, 140, mode, WithIndicators())
extract = ExtractMultiParagraphsPerQuestion(MergeParagraphs(args.n_tokens),
ShallowOpenWebRanker(16),
model.preprocessor,
intern=True)
eval = [
LossEvaluator(),
MultiParagraphSpanEvaluator(8, "triviaqa", mode != "merge")
]
oversample = [1] * 4
if mode == "paragraph":
n_epochs = 120
test = RandomParagraphSetDatasetBuilder(120, "flatten", True,
oversample)
train = StratifyParagraphsBuilder(ClusteredBatcher(
60, ContextLenBucketedKey(3), True),
oversample,
only_answers=True)
elif mode == "confidence" or mode == "sigmoid":
if mode == "sigmoid":
n_epochs = 640
else:
n_epochs = 160
test = RandomParagraphSetDatasetBuilder(120, "flatten", True,
oversample)
train = StratifyParagraphsBuilder(
ClusteredBatcher(60, ContextLenBucketedKey(3), True), oversample)
else:
n_epochs = 80
test = RandomParagraphSetDatasetBuilder(
120, "merge" if mode == "merge" else "group", True, oversample)
train = StratifyParagraphSetsBuilder(30, mode == "merge", True,
oversample)
data = TriviaQaWebDataset()
params = TrainParams(SerializableOptimizer("Adadelta",
dict(learning_rate=1)),
num_epochs=n_epochs,
ema=0.999,
max_checkpoints_to_keep=2,
async_encoding=10,
log_period=30,
eval_period=1800,
save_period=1800,
eval_samples=dict(dev=None, train=6000))
data = PreprocessedData(data, extract, train, test, eval_on_verified=False)
data.preprocess(args.n_processes, 1000)
with open(__file__, "r") as f:
notes = f.read()
notes = "Mode: " + args.mode + "\n" + notes
trainer.start_training(data, model, params, eval, model_dir.ModelDir(out),
notes)
def main():
parser = argparse.ArgumentParser(description='')
parser.add_argument('answers', help='answer file')
parser.add_argument('question_source')
args = parser.parse_args()
print("Loading answers..")
answer_df = pd.read_csv(args.answers)
print("Loading questions..")
if args.question_source == "open":
corpus = TriviaQaOpenDataset()
questions = {q.question_id: q for q in corpus.get_dev()}
elif args.question_source == "web":
corpus = TriviaQaWebDataset()
questions = {}
for q in corpus.get_dev():
for d in q.all_docs:
questions[(q.question_id, d.doc_id)] = q
elif args.question_source == "squad":
show_squad_errors(args.answers)
return
else:
raise ValueError()
pre = WithIndicators()
answer_df.sort_values(["question_id", "rank"], inplace=True)
if args.question_source == "open":
iter = answer_df.groupby(["question_id"])
else:
iter = answer_df.groupby(["question_id", "doc_id"])
grouped = list(iter)
np.random.shuffle(grouped)
for key, group in grouped:
print(list(questions.keys())[:10])
q = questions[key]
cur_best_score = group.text_f1.iloc[0]
cur_best_conf = group.predicted_score.iloc[0]
cur_best_ix = group.index[0]
for i in range(1, len(group)):
ix = group.index[i]
conf = group.predicted_score[ix]
if conf > cur_best_conf:
score = group.text_f1[ix]
if score < cur_best_score:
# We hurt our selves!
print("Oh no!")
print(" ".join(q.question))
print(q.answer.all_answers)
print("Best score was %.4f (conf=%.4f), but not is %.4f (conf=%.4f)" % (
cur_best_score, cur_best_conf, score, conf
))
d1 = [d for d in q.all_docs if d.doc_id == group.doc_id[cur_best_ix]][0]
p1 = extract_paragraph(corpus.evidence.get_document(d1.doc_id), group.para_start[cur_best_ix], group.para_end[cur_best_ix])
s, e = group.para_start[cur_best_ix], group.para_end[cur_best_ix]
answers = d1.answer_spans[np.logical_and(d1.answer_spans[:, 0] >= s, d1.answer_spans[:, 1] < s)] - s
p1 = pre.encode_extracted_paragraph(q.question, ExtractedParagraphWithAnswers(
p1, group.para_start[cur_best_ix], group.para_end[cur_best_ix], answers))
d2 = [d for d in q.all_docs if d.doc_id == group.doc_id[ix]][0]
p2 = extract_paragraph(corpus.evidence.get_document(d2.doc_id), group.para_start[ix], group.para_end[ix])
s, e = group.para_start[ix], group.para_end[ix]
answers = d2.answer_spans[np.logical_and(d2.answer_spans[:, 0] >= s, d2.answer_spans[:, 1] < s)] - s
p2 = pre.encode_extracted_paragraph(q.question, ExtractedParagraphWithAnswers(
p2, group.para_start[ix], group.para_end[ix], answers))
p1_s, p1_e = group.predicted_start[cur_best_ix], group.predicted_end[cur_best_ix]
p2_s, p2_e = group.predicted_start[ix], group.predicted_end[ix]
print(" ".join(display_para(p1.text, p1.answer_spans, q.question, p1_s, p1_e)))
print()
print(" ".join(display_para(p2.text, p2.answer_spans, q.question, p2_s, p2_e)))
input()
else:
cur_best_score = score
cur_best_ix = ix
cur_best_conf = conf
