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 read_input_data(model):
data = []
vocab = set()
tokenizer = NltkAndPunctTokenizer()
splitter = Truncate(400) # NOTE: we truncate past 400 tokens
selector = TopTfIdf(NltkPlusStopWords(True), n_to_select=5)
with open(OPTS.input_file) as f:
for i, line in enumerate(f):
try:
document_raw, question_raw = line.strip().split('\t')
except ValueError as e:
print(line.strip())
print('Error at line %d' % i)
raise e
document = re.split("\s*\n\s*", document_raw)
question = tokenizer.tokenize_paragraph_flat(question_raw)
doc_toks = [tokenizer.tokenize_paragraph(p) for p in document]
split_doc = splitter.split(doc_toks)
context = selector.prune(question, split_doc)
if model.preprocessor is not None:
context = [model.preprocessor.encode_text(question, x) for x in context]
else:
context = [flatten_iterable(x.text) for x in context]
vocab.update(question)
for txt in context:
vocab.update(txt)
ex = [ParagraphAndQuestion(x, question, None, "user-question%d"%i)
for i, x in enumerate(context)]
data.append((document_raw, question_raw, context, ex))
return data, vocab
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():
corpus = SquadCorpus()
if OPTS.normalize_before_ranking:
normalizer = WordNormalizer()
else:
normalizer = None
if OPTS.use_vec_dist:
word_vecs = corpus.get_pruned_word_vecs('glove.840B.300d')
prepro = SquadVectorTfIdfRanker(NltkPlusStopWords(True), OPTS.num_per_orig, True, word_vecs, word_normalizer=normalizer)
else:
prepro = SquadTfIdfRanker(NltkPlusStopWords(True), OPTS.num_per_orig, True, word_normalizer=normalizer)
orig_data = corpus.get_train() if OPTS.split == 'train' else corpus.get_dev()
orig_lens = [len(p.text[0]) for doc in orig_data for p in doc.paragraphs
for q in p.questions]
new_data = preprocess_par(orig_data, corpus.evidence, prepro, n_processes=1)
new_lens = [len(p.text) for q in new_data for p in q.paragraphs]
print('%d original, mean %.2f words' % (len(orig_lens), np.mean(orig_lens)))
print('%d new, mean %.2f words'% (len(new_lens), np.mean(new_lens)))
if OPTS.out_file:
write_output(OPTS.split, new_data, OPTS.out_file)
def get_para_filter(filter_name, per_document, n_paragraphs):
filter_name = ('tfidf' if per_document else
'linear') if filter_name is None else filter_name
if filter_name == "tfidf":
para_filter = TopTfIdf(NltkPlusStopWords(punctuation=True),
n_paragraphs)
elif filter_name == "truncate":
para_filter = FirstN(n_paragraphs)
elif filter_name == "linear":
para_filter = ShallowOpenWebRanker(n_paragraphs)
else:
raise ValueError()
return para_filter
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 find_answer(documents, raw_question):
raw_question = raw_question.lower()
documents = [d.lower() for d in documents]
global best_spans, conf
documents = [re.split("\s*\n\s*", doc) for doc in documents]
tokenizer = NltkAndPunctTokenizer()
question = tokenizer.tokenize_paragraph_flat(raw_question)
documents = [[tokenizer.tokenize_paragraph(p) for p in doc]
for doc in documents]
splitter = MergeParagraphs(400)
documents = [splitter.split(doc) for doc in documents]
if len(documents) == 1:
selector = TopTfIdf(NltkPlusStopWords(True), n_to_select=5)
context = selector.prune(question, documents[0])
else:
selector = ShallowOpenWebRanker(n_to_select=10)
context = selector.prune(question, flatten_iterable(documents))
context = [flatten_iterable(x.text) for x in context]
data = [
ParagraphAndQuestion(x, question, None, "user-question%d" % i)
for i, x in enumerate(context)
]
encoded = model.encode(data, is_train=False)
with sess.as_default():
spans, confid = sess.run([best_spans, conf], feed_dict=encoded)
best_para = np.argmax(confid)
ans = " ".join(context[best_para][spans[best_para][0]:spans[best_para][1] +
1])
confidence = confid[best_para]
return ans, confidence
def main():
corpus = SquadCorpus()
prepro = SquadTfIdfRanker(NltkPlusStopWords(True), OPTS.num_per_orig, True)
orig_data = corpus.get_train(
) if OPTS.split == 'train' else corpus.get_dev()
orig_lens = [
len(p.text[0]) for doc in orig_data for p in doc.paragraphs
for q in p.questions
]
new_data = preprocess_par(orig_data,
corpus.evidence,
prepro,
n_processes=1)
new_lens = [len(p.text) for q in new_data for p in q.paragraphs]
print('%d original, mean %.2f words' %
(len(orig_lens), np.mean(orig_lens)))
print('%d new, mean %.2f words' % (len(new_lens), np.mean(new_lens)))
if OPTS.out_file:
write_output(OPTS.split, new_data, OPTS.out_file)
def show_open_paragraphs(start: int, end: int):
splitter = MergeParagraphs(400)
stop = NltkPlusStopWords(True)
ranker = ShallowOpenWebRanker(6)
stop_words = stop.words
print("Loading train")
corpus = TriviaQaOpenDataset()
train = corpus.get_dev()
np.random.shuffle(train)
for q in train:
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}
para = []
for d in q.all_docs:
doc = corpus.evidence.get_document(d.doc_id)
para += splitter.split_annotated(doc, d.answer_spans)
ranked = ranker.prune(q.question, para)
if len(ranked) < start:
continue
ranked = ranked[start:end]
print(" ".join(q.question))
print(q.answer.all_answers)
for i in range(start, end):
para = ranked[i]
text = flatten_iterable(para.text)
print("Start=%d, Rank=%d" % (para.start, i))
if len(para.answer_spans) == 0:
# print("No Answer!")
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 = TriviaQaOpenDataset()
# data = TriviaQaWebDataset()
print("Loading...")
all_questions = data.get_dev()
questions = [
q for q in all_questions if any(
len(x.answer_spans) > 0 for x in q.all_docs)
]
print(
"%d/%d (%.4f) have an answer" % (len(questions), len(all_questions),
len(questions) / len(all_questions)))
np.random.shuffle(questions)
pre = ExtractMultiParagraphsPerQuestion(MergeParagraphs(400),
TopTfIdf(NltkPlusStopWords(), 20),
require_an_answer=False)
print("Done")
out = preprocess_par(questions[:2000], data.evidence, pre, 2, 1000)
n_counts = np.zeros(20)
n_any = np.zeros(20)
n_any_all = np.zeros(20)
for q in out.data:
for i, p in enumerate(q.paragraphs):
n_counts[i] += 1
n_any[i] += len(p.answer_spans) > 0
for i, p in enumerate(q.paragraphs):
if len(p.answer_spans) > 0:
n_any_all[i:] += 1
break
print(n_any_all / out.true_len)
print(n_any / n_counts)
print(n_counts)
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 __init__(self, n_to_select):
self.n_to_select = n_to_select
self._stop = NltkPlusStopWords(True).words
self._tfidf = TfidfVectorizer(strip_accents="unicode", stop_words=self._stop)
def main():
parser = argparse.ArgumentParser(
description='Train a model on document-level SQuAD')
parser.add_argument(
'mode',
choices=["paragraph", "confidence", "shared-norm", "merge", "sigmoid"])
parser.add_argument("name", help="Output directory")
args = parser.parse_args()
mode = args.mode
out = args.name + "-" + datetime.now().strftime("%m%d-%H%M%S")
corpus = SquadCorpus()
if mode == "merge":
# Adds paragraph start tokens, since we will be concatenating paragraphs together
pre = WithIndicators(True, para_tokens=False, doc_start_token=False)
else:
pre = None
model = get_model(50, 100, args.mode, pre)
if mode == "paragraph":
# Run in the "standard" known-paragraph setting
if model.preprocessor is not None:
raise NotImplementedError()
n_epochs = 26
train_batching = ClusteredBatcher(45, ContextLenBucketedKey(3), True,
False)
eval_batching = ClusteredBatcher(45, ContextLenKey(), False, False)
data = DocumentQaTrainingData(corpus, None, train_batching,
eval_batching)
eval = [LossEvaluator(), SpanEvaluator(bound=[17], text_eval="squad")]
else:
eval_set_mode = {
"confidence": "flatten",
"sigmoid": "flatten",
"shared-norm": "group",
"merge": "merge"
}[mode]
eval_dataset = RandomParagraphSetDatasetBuilder(
100, eval_set_mode, True, 0)
if mode == "confidence" or mode == "sigmoid":
if mode == "sigmoid":
# needs to be trained for a really long time for reasons unknown, even this might be too small
n_epochs = 100
else:
n_epochs = 50 # more epochs since we only "see" the label very other epoch-osh
train_batching = ClusteredBatcher(45, ContextLenBucketedKey(3),
True, False)
data = PreprocessedData(
SquadCorpus(),
SquadTfIdfRanker(NltkPlusStopWords(True), 4, True,
model.preprocessor),
StratifyParagraphsBuilder(train_batching, 1),
eval_dataset,
eval_on_verified=False,
)
else:
n_epochs = 26
data = PreprocessedData(
SquadCorpus(),
SquadTfIdfRanker(NltkPlusStopWords(True), 4, True,
model.preprocessor),
StratifyParagraphSetsBuilder(25, args.mode == "merge", True,
1),
eval_dataset,
eval_on_verified=False,
)
eval = [LossEvaluator(), MultiParagraphSpanEvaluator(17, "squad")]
data.preprocess(1)
with open(__file__, "r") as f:
notes = f.read()
notes = args.mode + "\n" + notes
trainer.start_training(data, model, train_params(n_epochs), eval,
model_dir.ModelDir(out), notes)
def main():
parser = argparse.ArgumentParser(description="Run an ELMo model on user input")
# parser.add_argument("model", type=int, help="Model directory")
parser.add_argument("question", help="Question to answer")
parser.add_argument("documents", help="List of text documents to answer the question with", nargs='+')
args = parser.parse_args()
# Models path
SQUAD_MODEL_DIRECTORY_PATH = 'docqa/models-cpu/squad'
SQUAD_SHARED_NORM_MODEL_DIRECTORY_PATH = 'docqa/models-cpu/squad-shared-norm'
TRIVIAQA_MODEL_DIRECTORY_PATH = 'docqa/models-cpu/triviaqa-unfiltered-shared-norm'
TRIVIAQA_SHARED_NORM_MODEL_DIRECTORY_PATH = 'docqa/models-cpu/triviaqa-web-shared-norm'
models_directory = [
SQUAD_MODEL_DIRECTORY_PATH,
SQUAD_SHARED_NORM_MODEL_DIRECTORY_PATH,
TRIVIAQA_MODEL_DIRECTORY_PATH,
TRIVIAQA_SHARED_NORM_MODEL_DIRECTORY_PATH
]
print("Preprocessing...")
# Load the model
# model_dir = ModelDir(args.model)
model_dir = ModelDir(models_directory[0])
model = model_dir.get_model()
if not isinstance(model, ParagraphQuestionModel):
raise ValueError("This script is built to work for ParagraphQuestionModel models only")
# Read the documents
documents = []
for doc in args.documents:
if not isfile(doc):
raise ValueError(doc + " does not exist")
with open(doc, "r") as f:
documents.append(f.read())
print("Loaded %d documents" % len(documents))
# Split documents into lists of paragraphs
documents = [re.split("\s*\n\s*", doc) for doc in documents]
# Tokenize the input, the models expects data to be tokenized using `NltkAndPunctTokenizer`
# Note the model expects case-sensitive input
tokenizer = NltkAndPunctTokenizer()
question = tokenizer.tokenize_paragraph_flat(args.question) # List of words
# Now list of document->paragraph->sentence->word
documents = [[tokenizer.tokenize_paragraph(p) for p in doc] for doc in documents]
# Now group the document into paragraphs, this returns `ExtractedParagraph` objects
# that additionally remember the start/end token of the paragraph within the source document
splitter = MergeParagraphs(400)
# splitter = PreserveParagraphs() # Uncomment to use the natural paragraph grouping
documents = [splitter.split(doc) for doc in documents]
# Now select the top paragraphs using a `ParagraphFilter`
if len(documents) == 1:
# Use TF-IDF to select top paragraphs from the document
selector = TopTfIdf(NltkPlusStopWords(True), n_to_select=5)
context = selector.prune(question, documents[0])
else:
# Use a linear classifier to select top paragraphs among all the documents
selector = ShallowOpenWebRanker(n_to_select=10)
context = selector.prune(question, flatten_iterable(documents))
print("Select %d paragraph" % len(context))
if model.preprocessor is not None:
# Models are allowed to define an additional pre-processing step
# This will turn the `ExtractedParagraph` objects back into simple lists of tokens
context = [model.preprocessor.encode_text(question, x) for x in context]
else:
# Otherwise just use flattened text
context = [flatten_iterable(x.text) for x in context]
print("Setting up model")
# Tell the model the batch size (can be None) and vocab to expect, This will load the
# needed word vectors and fix the batch size to use when building the graph / encoding the input
voc = set(question)
for txt in context:
voc.update(txt)
model.set_input_spec(ParagraphAndQuestionSpec(batch_size=len(context)), voc)
# Now we build the actual tensorflow graph, `best_span` and `conf` are
# tensors holding the predicted span (inclusive) and confidence scores for each
# element in the input batch, confidence scores being the pre-softmax logit for the span
print("Build tf graph")
sess = tf.Session(config=tf.ConfigProto(allow_soft_placement=True))
# We need to use sess.as_default when working with the cuNND stuff, since we need an active
# session to figure out the # of parameters needed for each layer. The cpu-compatible models don't need this.
with sess.as_default():
# 8 means to limit the span to size 8 or less
best_spans, conf = model.get_prediction().get_best_span(10)
# Loads the saved weights
model_dir.restore_checkpoint(sess)
# Now the model is ready to run
# The model takes input in the form of `ContextAndQuestion` objects, for example:
data = [ParagraphAndQuestion(x, question, None, "user-question%d"%i)
for i, x in enumerate(context)]
print("Starting run")
# The model is run in two steps, first it "encodes" a batch of paragraph/context pairs
# into numpy arrays, then we use `sess` to run the actual model get the predictions
encoded = model.encode(data, is_train=False) # batch of `ContextAndQuestion` -> feed_dict
best_spans, conf = sess.run([best_spans, conf], feed_dict=encoded) # feed_dict -> predictions
best_para = np.argmax(conf) # We get output for each paragraph, select the most-confident one to print
print("Best Paragraph: " + str(best_para))
para_id = int(str(best_para))
# print("Best Paragraph: \n" + (" ".join((paras[para_id].text)[0])))
print("Best Paragraph: \n" + " ".join(context[para_id]))
print("Best span: " + str(best_spans[best_para]))
print("Answer text: " + " ".join(context[best_para][best_spans[best_para][0]:best_spans[best_para][1]+1]))
print("Confidence: " + str(conf[best_para]))
def main():
parser = argparse.ArgumentParser()
parser.add_argument(
'--n_processes',
type=int,
default=1,
help=
"Number of processes to do the preprocessing (selecting paragraphs+loading context) with"
)
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('-n',
'--n_sample',
type=int,
default=None,
help="Number of questions to evaluate on")
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(
'-c',
'--corpus',
choices=[
"en_dev", "en_test", "fr_dev", "fr_test", "de_dev", "de_test",
"ru_dev", "ru_test", "pt_dev", "pt_test", "zh_dev", "zh_test",
"pl_dev", "pl_test", "uk_dev", "uk_test", "ta_dev", "ta_test",
"fr_trans_en_dev", "fr_trans_en_test", "de_trans_en_dev",
"de_trans_en_test", "ru_trans_en_dev", "ru_trans_en_test",
"pt_trans_en_dev", "pt_trans_en_test", "zh_trans_en_dev",
"zh_trans_en_test", "pl_trans_en_dev", "pl_trans_en_test",
"uk_trans_en_dev", "uk_trans_en_test", "ta_trans_en_dev",
"ta_trans_en_test"
],
required=True)
args = parser.parse_args()
corpus_name = args.corpus[:args.corpus.rfind("_")]
eval_set = args.corpus[args.corpus.rfind("_") + 1:]
dataset = XQADataset(corpus_name)
if eval_set == "dev":
test_questions = dataset.get_dev()
elif eval_set == "test":
test_questions = dataset.get_test()
else:
raise AssertionError()
corpus = dataset.evidence
splitter = MergeParagraphs(args.tokens)
per_document = args.corpus.startswith(
"web") # wiki and web are both multi-document
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
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]
preprocessor = WithIndicators()
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,
preprocessor,
require_an_answer=False)
else:
prep = ExtractMultiParagraphsPerQuestion(splitter,
para_filter,
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)
# dump eval data for bert
import pickle
pickle.dump(questions,
open("%s_%d.pkl" % (args.corpus, args.n_paragraphs), "wb"))
def main():
parser = argparse.ArgumentParser(
description='Evaluate a model on document-level SQuAD')
parser.add_argument('model', help='model to use')
parser.add_argument(
'output',
type=str,
help="Store the per-paragraph results in csv format in this file")
parser.add_argument('-n',
'--n_sample',
type=int,
default=None,
help="(for testing) sample documents")
parser.add_argument(
'-s',
'--async',
type=int,
default=10,
help="Encoding batch asynchronously, queueing up to this many")
parser.add_argument('-a',
'--answer_bound',
type=int,
default=17,
help="Max answer span length")
parser.add_argument('-p',
'--n_paragraphs',
type=int,
default=None,
help="Max number of paragraphs to use")
parser.add_argument(
'-b',
'--batch_size',
type=int,
default=200,
help="Batch size, larger sizes can be faster but uses more memory")
parser.add_argument('-c',
'--corpus',
choices=["dev", "train", "doc-rd-dev"],
default="dev")
parser.add_argument('--no_ema',
action="store_true",
help="Don't use EMA weights even if they exist")
args = parser.parse_args()
model_dir = ModelDir(args.model)
print("Loading data")
questions = []
ranker = SquadTfIdfRanker(NltkPlusStopWords(True),
args.n_paragraphs,
force_answer=False)
if args.corpus == "doc-rd-dev":
docs = SquadCorpus().get_dev()
if args.n_sample is not None:
docs.sort(key=lambda x: x.doc_id)
np.random.RandomState(0).shuffle(docs)
docs = docs[:args.n_sample]
print("Fetching document reader docs...")
doc_rd_versions = get_doc_rd_doc(docs)
print("Ranking and matching with questions...")
for doc in tqdm(docs):
doc_questions = flatten_iterable(x.questions
for x in doc.paragraphs)
paragraphs = doc_rd_versions[doc.title]
ranks = ranker.rank([x.words for x in doc_questions],
[x.text for x in paragraphs])
for i, question in enumerate(doc_questions):
para_ranks = np.argsort(ranks[i])
for para_rank, para_num in enumerate(
para_ranks[:args.n_paragraphs]):
# Just use dummy answers spans for these pairs
questions.append(
RankedParagraphQuestion(
question.words,
TokenSpans(question.answer.answer_text,
np.zeros((0, 2), dtype=np.int32)),
question.question_id, paragraphs[para_num],
para_rank, para_num))
rl = ResourceLoader()
else:
if args.corpus == "dev":
docs = SquadCorpus().get_dev()
else:
docs = SquadCorpus().get_train()
rl = SquadCorpus().get_resource_loader()
if args.n_sample is not None:
docs.sort(key=lambda x: x.doc_id)
np.random.RandomState(0).shuffle(docs)
docs = docs[:args.n_sample]
for q in ranker.ranked_questions(docs):
for i, p in enumerate(q.paragraphs):
questions.append(
RankedParagraphQuestion(
q.question, TokenSpans(q.answer_text, p.answer_spans),
q.question_id,
ParagraphWithInverse([p.text], p.original_text,
p.spans), i, p.paragraph_num))
print("Split %d docs into %d paragraphs" % (len(docs), len(questions)))
questions = sorted(questions,
key=lambda x: (x.n_context_words, len(x.question)),
reverse=True)
for q in questions:
if len(q.answer.answer_spans.shape) != 2:
raise ValueError()
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()
if checkpoint is None:
raise ValueError("No checkpoints found")
data = ParagraphAndQuestionDataset(
questions, FixedOrderBatcher(args.batch_size, True))
model = model_dir.get_model()
evaluation = trainer.test(
model, [RecordParagraphSpanPrediction(args.answer_bound, True)],
{args.corpus: data}, rl, checkpoint, not args.no_ema,
args. async)[args.corpus]
print("Saving result")
output_file = args.output
df = pd.DataFrame(evaluation.per_sample)
df.sort_values(["question_id", "rank"], inplace=True, ascending=True)
group_by = ["question_id"]
f1 = compute_ranked_scores(df, "predicted_score", "text_f1", group_by)
em = compute_ranked_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)
df.to_csv(output_file, index=False)
def paragraph_stats(corpus, splitter: DocumentSplitter, sample):
stop = NltkPlusStopWords(punctuation=True).words
data = corpus.get_dev()
pairs = flatten_iterable([(q, doc) for doc in q.all_docs] for q in data)
data = [
pairs[i] for i in np.random.choice(
np.arange(0, len(pairs)), sample, replace=False)
]
word_matches = Counter()
n_para = []
n_answers = []
n_question_words = []
for q, doc in data:
if len(doc.answer_spans) == 0:
continue
q_words = set(x.lower() for x in q.question)
q_words -= stop
# q_words = set(norm.normalize(w) for w in q_words)
text = corpus.evidence.get_document(doc.doc_id)
para = splitter.split_annotated(text, doc.answer_spans)
n_para.append(len(para))
n_answers += [len(x.answer_spans) for x in para]
for x in para:
match_set = set()
n_matches = 0
text = flatten_iterable(x.text)
for word in text:
word = word.lower()
if word in q_words:
n_matches += 1
match_set.add(word)
if len(match_set) == 0 and len(x.answer_spans) > 0:
print_paragraph(q, x)
input()
word_matches.update(match_set)
n_question_words.append(n_matches)
n_answers = np.array(n_answers)
n_question_words = np.array(n_question_words)
any_answers = n_answers > 0
any_question_word = n_question_words > 0
total_para = len(any_answers)
total_q = len(n_para)
no_quesiton_and_answer = any_answers[np.logical_not(any_question_word)]
print("%d/%d (%.4f) pairs have an answer" %
(total_q, len(data), total_q / len(data)))
print("%d para in %d questions (av %.4f)" %
(sum(n_para), total_q, sum(n_para) / total_q))
print("%d/%d (%.4f) paragraphs have answers" %
(any_answers.sum(), total_para, any_answers.mean()))
print("%d/%d (%.4f) paragraphs have question word" %
(any_question_word.sum(), total_para, any_question_word.mean()))
print("%d/%d (%.4f) no question words have answers" %
(no_quesiton_and_answer.sum(), len(no_quesiton_and_answer),
no_quesiton_and_answer.mean()))
def main(Data: pd.DataFrame, nlp, model_dir, model):
#parser = argparse.ArgumentParser(description="Run an ELMo model on user input")
#parser.add_argument("model", help="Model directory")
#parser.add_argument("question", help="Question to answer")
#parser.add_argument("documents", help="List of text documents to answer the question with", nargs='+')
#args = parser.parse_args()
#print("Preprocessing...")
# Load the model
#model_dir = ModelDir(MODEL_DIR)
#model = model_dir.get_model()
print(model)
if not isinstance(model, ParagraphQuestionModel):
raise ValueError(
"This script is built to work for ParagraphQuestionModel models only"
)
#print(model)
# Read the documents
documents = []
documents.append(Data.at[0, 'Filetext'])
"""import pyodbc
conn = pyodbc.connect("Driver={ODBC Driver 13 for SQL Server};"
"Server=192.168.100.15;"
"Database=PharmaAce;"
"UID=sa;"
"PWD=admin@123;"
"Trusted_Connection=no;")
cursor=conn.cursor()
#(23211,28690,33214,25638,25837,26454,28693,26137,31428,32087)
for doc in cursor.execute("select cast(filetext as varchar(max)) as filetext from kpl_tmp"):
documents.append(doc[0])
#doc="D:\Document QnA\document-qa-master\Data\Drug_Delivery_Surveying_Global_Competitive_Landscape_BMI.txt"
if not isfile(doc):
raise ValueError(doc + " does not exist")
with open(doc, "r") as f:
documents.append(f.read())
"""
#print("Loaded %d documents" % len(documents))
#temp=documents[0].split()
# Split documents into lists of paragraphs
#documents=[" ".join(temp[i:(i+400)]) for i in range(1,len(temp),400)]
documents = [re.split("\s*\n\s*", doc) for doc in documents]
# Tokenize the input, the models expects data to be tokenized using `NltkAndPunctTokenizer`
# Note the model expects case-sensitive input
tokenizer = NltkAndPunctTokenizer()
question = tokenizer.tokenize_paragraph_flat(
Data.at[0, 'Question']) # List of words
# Now list of document->paragraph->sentence->word
documents = [[tokenizer.tokenize_paragraph(p) for p in doc]
for doc in documents]
# Now group the document into paragraphs, this returns `ExtractedParagraph` objects
# that additionally remember the start/end token of the paragraph within the source document
splitter = MergeParagraphs(400)
#splitter = PreserveParagraphs() # Uncomment to use the natural paragraph grouping
documents = [splitter.split(doc) for doc in documents]
#print(str(len(documents))+" kpl") #kpl
# Now select the top paragraphs using a `ParagraphFilter`
print(len(documents)) #kpl
if len(documents) == 1:
# Use TF-IDF to select top paragraphs from the document
selector = TopTfIdf(NltkPlusStopWords(True), n_to_select=5)
context = selector.prune(question, documents[0])
else:
# Use a linear classifier to select top paragraphs among all the documents
selector = ShallowOpenWebRanker(n_to_select=10)
context = selector.prune(question, flatten_iterable(documents))
#print("Select %d paragraph" % len(context))
if model.preprocessor is not None:
# Models are allowed to define an additional pre-processing step
# This will turn the `ExtractedParagraph` objects back into simple lists of tokens
context = [
model.preprocessor.encode_text(question, x) for x in context
]
else:
# Otherwise just use flattened text
context = [flatten_iterable(x.text) for x in context]
print("Setting up model")
# Tell the model the batch size (can be None) and vocab to expect, This will load the
# needed word vectors and fix the batch size to use when building the graph / encoding the input
voc = set(question)
for txt in context:
voc.update(txt)
model.set_input_spec(nlp,
ParagraphAndQuestionSpec(batch_size=len(context)),
voc)
# Now we build the actual tensorflow graph, `best_span` and `conf` are
# tensors holding the predicted span (inclusive) and confidence scores for each
# element in the input batch, confidence scores being the pre-softmax logit for the span
#print("Build tf graph") #kpl
print("after set input spec")
sess = tf.Session(config=tf.ConfigProto(allow_soft_placement=True))
# We need to use sess.as_default when working with the cuNND stuff, since we need an active
# session to figure out the # of parameters needed for each layer. The cpu-compatible models don't need this.
with sess.as_default():
# 8 means to limit the span to size 8 or less
best_spans, conf = model.get_prediction().get_best_span(8)
# Loads the saved weights
model_dir.restore_checkpoint(sess)
print("after loading weights")
# Now the model is ready to run
# The model takes input in the form of `ContextAndQuestion` objects, for example:
data = [
ParagraphAndQuestion(x, question, None, "user-question%d" % i)
for i, x in enumerate(context)
]
#print("Starting run")
# The model is run in two steps, first it "encodes" a batch of paragraph/context pairs
# into numpy arrays, then we use `sess` to run the actual model get the predictions
encoded = model.encode(
data, is_train=True) # batch of `ContextAndQuestion` -> feed_dict
best_spans, conf = sess.run([best_spans, conf],
feed_dict=encoded) # feed_dict -> predictions
best_para = np.argmax(
conf
) # We get output for each paragraph, select the most-confident one to print
#print("Best Paragraph: " + str(best_para))
#print("Best span: " + str(best_spans[best_para]))
#print("Answer text: " + " ".join(context[best_para][best_spans[best_para][0]:best_spans[best_para][1]+1]))
#print("Confidence: " + str(conf[best_para]))
return " ".join(
context[best_para][best_spans[best_para][0]:best_spans[best_para][1] +
1])
#if __name__ == "__main__":
# main()
def main():
print('Starting...')
model_dir = ModelDir(OPTS.model)
model = model_dir.get_model()
tokenizer = NltkAndPunctTokenizer()
if not isinstance(model, ParagraphQuestionModel):
raise ValueError(
"This script is built to work for ParagraphQuestionModel models only"
)
if OPTS.reload_vocab:
loader = ResourceLoader()
else:
loader = CachingResourceLoader()
print('Loading word vectors...')
model.set_input_spec(ParagraphAndQuestionSpec(batch_size=None),
set([',']),
word_vec_loader=loader,
allow_update=True)
print('Starting Tensorflow session...')
sess = tf.Session(config=tf.ConfigProto(allow_soft_placement=True))
with sess.as_default():
prediction = model.get_prediction()
# Take 0-th here because we know we only truncate to one paragraph
start_logits_tf = prediction.start_logits[0]
end_logits_tf = prediction.end_logits[0]
none_logit_tf = prediction.none_logit[0]
#best_spans_tf, conf_tf = prediction.get_best_span(MAX_SPAN_LENGTH)
model_dir.restore_checkpoint(sess)
splitter = Truncate(400) # NOTE: we truncate past 400 tokens
selector = TopTfIdf(NltkPlusStopWords(True), n_to_select=5)
app = bottle.Bottle()
@app.route('/')
def index():
return bottle.template('index')
@app.route('/post_query', method='post')
def post_query():
document_raw = bottle.request.forms.getunicode('document').strip()
question_raw = bottle.request.forms.getunicode('question').strip()
document = re.split("\s*\n\s*", document_raw)
question = tokenizer.tokenize_paragraph_flat(question_raw)
doc_toks = [tokenizer.tokenize_paragraph(p) for p in document]
split_doc = splitter.split(doc_toks)
context = selector.prune(question, split_doc)
if model.preprocessor is not None:
context = [
model.preprocessor.encode_text(question, x) for x in context
]
else:
context = [flatten_iterable(x.text) for x in context]
vocab = set(question)
for txt in context:
vocab.update(txt)
data = [
ParagraphAndQuestion(x, question, None, "user-question%d" % i)
for i, x in enumerate(context)
]
model.word_embed.update(loader, vocab)
encoded = model.encode(data, is_train=False)
start_logits, end_logits, none_logit = sess.run(
[start_logits_tf, end_logits_tf, none_logit_tf], feed_dict=encoded)
beam, p_na = logits_to_probs(document_raw,
context[0],
start_logits,
end_logits,
none_logit,
beam_size=BEAM_SIZE)
return bottle.template('results',
document=document_raw,
question=question_raw,
beam=beam,
p_na=p_na)
cur_dir = os.path.abspath(os.path.dirname(__file__))
bottle.TEMPLATE_PATH.insert(0, os.path.join(cur_dir, 'views'))
bottle.run(app, host=OPTS.hostname, port=OPTS.port, debug=OPTS.debug)
def getAnswer(self):
#parser = argparse.ArgumentParser(description="Run an ELMo model on user input")
#parser.add_argument("model", help="Model directory")
#parser.add_argument("question", help="Question to answer")
#parser.add_argument("documents", help="List of text documents to answer the question with", nargs='+')
#args = parser.parse_args()
#print("Preprocessing...")
# Load the model
model_dir = ModelDir(MODEL_DIR)
model = model_dir.get_model()
if not isinstance(model, ParagraphQuestionModel):
raise ValueError(
"This script is built to work for ParagraphQuestionModel models only"
)
conn = pyodbc.connect(DB_CONN)
cursor = conn.cursor()
#(23211,28690,33214,25638,25837,26454,28693,26137,31428,32087)
query="select cast(filetext as varchar(max)) as filetext, name, type from dbo.UserworkspaceData where objectmasterid= "+\
str(self.ObjectMasterId)+\
" order by id asc"
#query="select cast(filetext as varchar(max)) as filetext from kpl_tmp"
documents = []
document = ""
name = ""
filetype = 0
for doc in cursor.execute(query):
document = document + doc[0]
name = doc[1]
filetype = doc[2]
#open("E:/kpl.txt","w+").write(document)
documents.append(document)
#documents.replace("\n\n","\n")
#r.sub("",documents)
#documents=" ".join(documents.split())
#open("E:\kpl_test.txt","w+").write(document)
#doc="D:\Document QnA\document-qa-master\Data\Drug_Delivery_Surveying_Global_Competitive_Landscape_BMI.txt"
# =============================================================================
# if not isfile(doc):
# raise ValueError(doc + " does not exist")
# with open(doc, "r") as f:
# documents.append(f.read())
# =============================================================================
#print("Loaded %d documents" % len(documents))
#temp=documents[0].split()
# Split documents into lists of paragraphs
#documents=[" ".join(temp[i:(i+400)]) for i in range(1,len(temp),400)]
documents = [re.split("\s*\n\s*", doc) for doc in documents]
# Tokenize the input, the models expects data to be tokenized using `NltkAndPunctTokenizer`
# Note the model expects case-sensitive input
tokenizer = NltkAndPunctTokenizer()
question = tokenizer.tokenize_paragraph_flat(
self.Question) # List of words
# Now list of document->paragraph->sentence->word
documents = [[tokenizer.tokenize_paragraph(p) for p in doc]
for doc in documents]
# Now group the document into paragraphs, this returns `ExtractedParagraph` objects
# that additionally remember the start/end token of the paragraph within the source document
splitter = MergeParagraphs(400)
#splitter = PreserveParagraphs() # Uncomment to use the natural paragraph grouping
documents = [splitter.split(doc) for doc in documents]
#print(str(len(documents))+" kpl") #kpl
# Now select the top paragraphs using a `ParagraphFilter`
if len(documents) == 1:
# Use TF-IDF to select top paragraphs from the document
selector = TopTfIdf(NltkPlusStopWords(True), n_to_select=5)
context = selector.prune(question, documents[0])
else:
# Use a linear classifier to select top paragraphs among all the documents
selector = ShallowOpenWebRanker(n_to_select=10)
context = selector.prune(question, flatten_iterable(documents))
#print("Select %d paragraph" % len(context))
if model.preprocessor is not None:
# Models are allowed to define an additional pre-processing step
# This will turn the `ExtractedParagraph` objects back into simple lists of tokens
context = [
model.preprocessor.encode_text(question, x) for x in context
]
else:
# Otherwise just use flattened text
context = [flatten_iterable(x.text) for x in context]
#x=open("E:\context.txt","a+")
#[x.write(" ".join(cont)) for cont in context]
#x.write("\n.......................................................\n")
#print("Setting up model")
# Tell the model the batch size (can be None) and vocab to expect, This will load the
# needed word vectors and fix the batch size to use when building the graph / encoding the input
voc = set(question)
for txt in context:
voc.update(txt)
model.set_input_spec(self.nlp,
ParagraphAndQuestionSpec(batch_size=len(context)),
voc)
# Now we build the actual tensorflow graph, `best_span` and `conf` are
# tensors holding the predicted span (inclusive) and confidence scores for each
# element in the input batch, confidence scores being the pre-softmax logit for the span
#print("Build tf graph") #kpl
sess = tf.Session(config=tf.ConfigProto(allow_soft_placement=True))
# We need to use sess.as_default when working with the cuNND stuff, since we need an active
# session to figure out the # of parameters needed for each layer. The cpu-compatible models don't need this.
with sess.as_default():
# 8 means to limit the span to size 8 or less
best_spans, conf = model.get_prediction().get_best_span(8)
# Loads the saved weights
model_dir.restore_checkpoint(sess)
# Now the model is ready to run
# The model takes input in the form of `ContextAndQuestion` objects, for example:
data = [
ParagraphAndQuestion(x, question, None, "user-question%d" % i)
for i, x in enumerate(context)
]
#print("Starting run")
# The model is run in two steps, first it "encodes" a batch of paragraph/context pairs
# into numpy arrays, then we use `sess` to run the actual model get the predictions
encoded = model.encode(
data, is_train=True) # batch of `ContextAndQuestion` -> feed_dict
best_spans, conf = sess.run(
[best_spans, conf], feed_dict=encoded) # feed_dict -> predictions
best_para = np.argmax(
conf
) # We get output for each paragraph, select the most-confident one to print
#print("Best Paragraph: " + str(best_para))
#print("Best span: " + str(best_spans[best_para]))
#print("Answer text: " + " ".join(context[best_para][best_spans[best_para][0]:best_spans[best_para][1]+1]))
#print("Confidence: " + str(conf[best_para]))
Answer = " ".join(context[best_para]
[best_spans[best_para][0]:best_spans[best_para][1] +
1])
print("Confidence: " + str(conf[best_para]))
print("Best Paragraph: " + str(best_para))
print("Best span: " + str(best_spans[best_para]))
print("Answer text: " + Answer)
print(" ".join(context[best_para]))
context[best_para][best_spans[best_para][
0]] = r"<em>" + context[best_para][best_spans[best_para][0]]
context[best_para][best_spans[best_para][1]] = context[best_para][
best_spans[best_para][1]] + r"</em>"
start = 0
end = len(context[best_para])
positions = [
x for x, n in enumerate(context[best_para]
[0:best_spans[best_para][0]]) if n == "."
]
if len(positions) >= 2: start = positions[len(positions) - 2] + 1
positions = [
x
for x, n in enumerate(context[best_para][best_spans[best_para][1] +
1:]) if n == "."
]
if len(positions) > 1:
end = best_spans[best_para][1] + 1 + positions[1]
d = dict()
if conf[best_para] > 10:
d["answer"] = Answer
else:
d["answer"] = ""
d["name"] = name
d["filetype"] = filetype
d["paragraph"] = re.sub(r' (?=\W)', '',
" ".join(context[best_para][start:end]))
d["ObjectMasterId"] = self.ObjectMasterId
return d
#if __name__ == "__main__":
# main()
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")
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 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 predict():
json_data = {"success": False, "predictions": []}
print("Preprocessing...")
# Load the model
model_dir = ModelDir(
"/home/antriv/conversation_ai/Transfer_Learning/ALLENAI_DocumentQA/document-qa/pretrained_models/models/triviaqa-unfiltered-shared-norm"
)
model = model_dir.get_model()
if not isinstance(model, ParagraphQuestionModel):
raise ValueError(
"This script is built to work for ParagraphQuestionModel models only"
)
# Load the question
question = (flask.request.data).decode("utf-8")
# Read the documents
documents = []
doclist = ["/home/antriv/data/The-Future-Computed.txt"]
for doc in doclist:
if not isfile(doc):
raise ValueError(doc + " does not exist")
with open(doc, "r") as f:
documents.append(f.read())
print("Loaded %d documents" % len(documents))
# Split documents into lists of paragraphs
documents = [re.split("\s*\n\s*", doc) for doc in documents]
# Tokenize the input, the models expects data to be tokenized using `NltkAndPunctTokenizer`
# Note the model expects case-sensitive input
tokenizer = NltkAndPunctTokenizer()
question = tokenizer.tokenize_paragraph_flat(question) # List of words
# Now list of document->paragraph->sentence->word
documents = [[tokenizer.tokenize_paragraph(p) for p in doc]
for doc in documents]
# Now group the document into paragraphs, this returns `ExtractedParagraph` objects
# that additionally remember the start/end token of the paragraph within the source document
splitter = MergeParagraphs(400)
#splitter = PreserveParagraphs() # Uncomment to use the natural paragraph grouping
documents = [splitter.split(doc) for doc in documents]
# Now select the top paragraphs using a `ParagraphFilter`
if len(documents) == 1:
# Use TF-IDF to select top paragraphs from the document
selector = TopTfIdf(NltkPlusStopWords(True), n_to_select=1000)
context = selector.prune(question, documents[0])
else:
# Use a linear classifier to select top paragraphs among all the documents
selector = ShallowOpenWebRanker(n_to_select=1000)
context = selector.prune(question, flatten_iterable(documents))
print("Select %d paragraph" % len(context))
if model.preprocessor is not None:
# Models are allowed to define an additional pre-processing step
# This will turn the `ExtractedParagraph` objects back into simple lists of tokens
context = [
model.preprocessor.encode_text(question, x) for x in context
]
else:
# Otherwise just use flattened text
context = [flatten_iterable(x.text) for x in context]
print("Setting up model")
# Tell the model the batch size (can be None) and vocab to expect, This will load the
# needed word vectors and fix the batch size to use when building the graph / encoding the input
voc = set(question)
for txt in context:
voc.update(txt)
model.set_input_spec(ParagraphAndQuestionSpec(batch_size=len(context)),
voc)
# Now we build the actual tensorflow graph, `best_span` and `conf` are
# tensors holding the predicted span (inclusive) and confidence scores for each
# element in the input batch, confidence scores being the pre-softmax logit for the span
print("Build tf graph")
sess = tf.Session(config=tf.ConfigProto(allow_soft_placement=True))
# We need to use sess.as_default when working with the cuNND stuff, since we need an active
# session to figure out the # of parameters needed for each layer. The cpu-compatible models don't need this.
with sess.as_default():
# 8 means to limit the span to size 8 or less
best_spans, conf = model.get_prediction().get_best_span(8)
# Loads the saved weights
model_dir.restore_checkpoint(sess)
# Now the model is ready to run
# The model takes input in the form of `ContextAndQuestion` objects, for example:
data = [
ParagraphAndQuestion(x, question, None, "user-question%d" % i)
for i, x in enumerate(context)
]
print("Starting run")
# The model is run in two steps, first it "encodes" a batch of paragraph/context pairs
# into numpy arrays, then we use `sess` to run the actual model get the predictions
encoded = model.encode(
data, is_train=False) # batch of `ContextAndQuestion` -> feed_dict
best_spans, conf = sess.run([best_spans, conf],
feed_dict=encoded) # feed_dict -> predictions
best_para = np.argmax(
conf
) # We get output for each paragraph, select the most-confident one to print
print("Best Paragraph: " + str(best_para))
print("Best span: " + str(best_spans[best_para]))
print("Answer text: " +
" ".join(context[best_para]
[best_spans[best_para][0]:best_spans[best_para][1] + 1]))
print("Confidence: " + str(conf[best_para]))
y_output = " ".join(
context[best_para][best_spans[best_para][0]:best_spans[best_para][1] +
1])
print(y_output)
json_data["predictions"].append(str(y_output))
#indicate that the request was a success
json_data["success"] = True
#return the data dictionary as a JSON response
return flask.jsonify(json_data)
def main():
parser = argparse.ArgumentParser("Train our ELMo model on SQuAD")
parser.add_argument("loss_mode", choices=['default', 'confidence'])
parser.add_argument("output_dir")
parser.add_argument("--dim", type=int, default=90)
parser.add_argument("--l2", type=float, default=0)
parser.add_argument("--mode",
choices=["input", "output", "both", "none"],
default="both")
parser.add_argument("--top_layer_only", action="store_true")
parser.add_argument("--no-tfidf",
action='store_true',
help="Don't add TF-IDF negative examples")
args = parser.parse_args()
out = args.output_dir + "-" + datetime.now().strftime("%m%d-%H%M%S")
dim = args.dim
recurrent_layer = CudnnGru(dim, w_init=TruncatedNormal(stddev=0.05))
if args.loss_mode == 'default':
n_epochs = 24
answer_encoder = SingleSpanAnswerEncoder()
predictor = BoundsPredictor(
ChainBiMapper(first_layer=recurrent_layer,
second_layer=recurrent_layer))
batcher = ClusteredBatcher(45, ContextLenKey(), False, False)
data = DocumentQaTrainingData(SquadCorpus(), None, batcher, batcher)
elif args.loss_mode == 'confidence':
if args.no_tfidf:
prepro = SquadDefault()
n_epochs = 15
else:
prepro = SquadTfIdfRanker(NltkPlusStopWords(True), 4, True)
n_epochs = 50
answer_encoder = DenseMultiSpanAnswerEncoder()
predictor = ConfidencePredictor(ChainBiMapper(
first_layer=recurrent_layer,
second_layer=recurrent_layer,
),
AttentionEncoder(),
FullyConnected(80, activation="tanh"),
aggregate="sum")
eval_dataset = RandomParagraphSetDatasetBuilder(
100, 'flatten', True, 0)
train_batching = ClusteredBatcher(45, ContextLenBucketedKey(3), True,
False)
data = PreprocessedData(SquadCorpus(),
prepro,
StratifyParagraphsBuilder(train_batching, 1),
eval_dataset,
eval_on_verified=False)
data.preprocess(1)
params = trainer.TrainParams(trainer.SerializableOptimizer(
"Adadelta", dict(learning_rate=1.0)),
ema=0.999,
max_checkpoints_to_keep=2,
async_encoding=10,
num_epochs=n_epochs,
log_period=30,
eval_period=1200,
save_period=1200,
best_weights=("dev", "b17/text-f1"),
eval_samples=dict(dev=None, train=8000))
lm_reduce = MapperSeq(
ElmoLayer(args.l2,
layer_norm=False,
top_layer_only=args.top_layer_only),
DropoutLayer(0.5),
)
model = AttentionWithElmo(
encoder=DocumentAndQuestionEncoder(answer_encoder),
lm_model=SquadContextConcatSkip(),
append_before_atten=(args.mode == "both" or args.mode == "output"),
append_embed=(args.mode == "both" or args.mode == "input"),
max_batch_size=128,
word_embed=FixedWordEmbedder(vec_name="glove.840B.300d",
word_vec_init_scale=0,
learn_unk=False,
cpu=True),
char_embed=CharWordEmbedder(LearnedCharEmbedder(word_size_th=14,
char_th=49,
char_dim=20,
init_scale=0.05,
force_cpu=True),
MaxPool(Conv1d(100, 5, 0.8)),
shared_parameters=True),
embed_mapper=SequenceMapperSeq(
VariationalDropoutLayer(0.8),
recurrent_layer,
VariationalDropoutLayer(0.8),
),
lm_reduce=None,
lm_reduce_shared=lm_reduce,
per_sentence=False,
memory_builder=NullBiMapper(),
attention=BiAttention(TriLinear(bias=True), True),
match_encoder=SequenceMapperSeq(
FullyConnected(dim * 2, activation="relu"),
ResidualLayer(
SequenceMapperSeq(
VariationalDropoutLayer(0.8),
recurrent_layer,
VariationalDropoutLayer(0.8),
StaticAttentionSelf(TriLinear(bias=True),
ConcatWithProduct()),
FullyConnected(dim * 2, activation="relu"),
)), VariationalDropoutLayer(0.8)),
predictor=predictor)
with open(__file__, "r") as f:
notes = f.read()
notes = str(sorted(args.__dict__.items(),
key=lambda x: x[0])) + "\n" + notes
trainer.start_training(
data, model, params,
[LossEvaluator(),
SpanEvaluator(bound=[17], text_eval="squad")], ModelDir(out), notes)
return out
if on_token + len(sent) < start:
pass
on_para += sent
on_token += len(sent)
if len(on_para) > 0:
out.append(on_para)
on_para = []
out.append(on_para)
if len(flatten_iterable(out)) != end - start:
raise ValueError(len(flatten_iterable(out)), end-start)
return out
stop = NltkPlusStopWords(True).words
class bcolors:
CORRECT = '\033[94m'
ERROR = '\033[91m'
CYAN = "\033[96m"
ENDC = '\033[0m'
def display_para(text: List[str], answers, question, p_start, p_end):
words = {w.lower() for w in question if w.lower() not in stop}
text = list(text)
if answers is not None:
for s,e in answers:
text[s] = bcolors.CORRECT + text[s]
text[e] = text[e] + bcolors.ENDC