def get_batch(index, contexts, start_token):
"make ICT batch data"
sentence = [contexts[i]["sentence"]
for i in index] # get sentences of paragraphs
target_sentence = [
random.randint(0,
len(sen) - 1) for sen in sentence
] # set target sentence for ICT training
remove_target = [
random.random() < (1 - config.remove_percent)
for _ in range(len(target_sentence))
] # determine removal of original sentence as mention in paper
target_context = [
sen[:i] + sen[i + remove:]
for i, sen, remove in zip(target_sentence, sentence, remove_target)
] # set sentences of target context
target_context = [[y for x in context for y in x]
for context in target_context
] # concat sentences of context
target_context = [[start_token] + context
for context in target_context]
target_sentence = [sen[i] for i, sen in zip(target_sentence, sentence)]
target_sentence = [[start_token] + sen for sen in target_sentence]
s, s_mask = util.pad_sequence(target_sentence,
max_seq=config.max_seq,
device=config.device) # pad sequence
c, c_mask = util.pad_sequence(target_context,
max_seq=config.max_seq,
device=config.device)
return s, s_mask, c, c_mask
def get_semantic_sim(model):
"make semantic embedding of context, question. and get similarity"
context_embedding = []
question_embedding = []
model.eval()
with torch.no_grad():
for i in tqdm(range(0, len(contexts), config.test_batch_size)):
c = [[y for x in context["sentence"] for y in x]
for context in contexts[i:i + config.test_batch_size]]
c, c_mask = util.pad_sequence(c,
max_seq=config.max_seq,
device=config.device)
c_encode = model(x=c, x_mask=c_mask)
context_embedding.append(c_encode.detach().cpu().numpy())
for i in tqdm(range(0, len(q_wordpiece), config.test_batch_size)):
q = [
tokens
for tokens in q_wordpiece[i:i + config.test_batch_size]
]
q, q_mask = util.pad_sequence(q,
max_seq=config.max_seq,
device=config.device)
q_encode = model(x=q, x_mask=q_mask)
question_embedding.append(q_encode.detach().cpu().numpy())
context_embedding = np.concatenate(context_embedding, axis=0)
question_embedding = np.concatenate(question_embedding, axis=0)
return util.get_sim(question_embedding, context_embedding)
def preprocess_batch(a_batch):
global FEATURE_LENGTH, VECTOR_EMBEDDINGS, NB_ARGUMENT_LABELS
_input_seqs = []
_output_seqs = []
_lengths = []
for _ds in a_batch:
_lengths.append(len(_ds.seq_features))
_max_length = max(_lengths)
for _ds in a_batch:
_seq = []
for feature in _ds.seq_features:
element = []
for l in feature:
if isinstance(l, str):
element.extend(
list(
VECTOR_EMBEDDINGS.get(l,
VECTOR_EMBEDDINGS["_unk_"])))
else:
element.append(l)
_seq.append(element)
_input_seqs.append(util.pad_sequence(_seq, _max_length,
FEATURE_LENGTH))
_output_seqs.append(
util.pad_sequence(list(_ds.seq_labels), _max_length, 1))
_input_seqs = np.array(_input_seqs)
_output_seqs = np.array(_output_seqs)
_lengths = np.array(_lengths)
return _input_seqs, _output_seqs, _lengths, _max_length
def generate_answers(sess, model, dataset, rev_vocab):
"""
Loop over the dev or test dataset and generate answer.
Note: output format must be answers[uuid] = "real answer"
You must provide a string of words instead of just a list, or start and end index
In main() function we are dumping onto a JSON file
evaluate.py will take the output JSON along with the original JSON file
and output a F1 and EM
You must implement this function in order to submit to Leaderboard.
:param sess: active TF session
:param model: a built QASystem model
:param rev_vocab: this is a list of vocabulary that maps index to actual words
:return:
"""
answers = {}
(context, question, question_uuid_data) = dataset
context_data = convert_data_to_list(context)
question_data = convert_data_to_list(question)
context_padded, context_mask = pad_sequence(context_data,
FLAGS.max_context_len)
question_padded, question_mask = pad_sequence(question_data,
FLAGS.max_question_len)
input_data = vectorize(context_padded, context_mask, question_padded,
question_mask)
minibatch_size = 20
for start in tqdm(range(0, len(question_uuid_data), minibatch_size),
desc="predicting on test"):
h_s, h_e = model.decode(sess, input_data[start:start + minibatch_size])
iter_num = min(minibatch_size,
len(question_uuid_data) - start - minibatch_size)
for i in range(iter_num):
a_s = np.argmax(h_s[i])
a_e = np.argmax(h_e[i])
if a_s > a_e:
k = a_e
a_e = a_s
a_s = k
uuid = question_uuid_data[start + i]
sample_dataset = [input_data[start + i]]
context = sample_dataset[0][0]
predicted_answer = model.formulate_answer(context, rev_vocab, a_s,
a_e)
answers[uuid] = predicted_answer
return answers
def preprocess_batch(a_batch):
"""
Preprocess a batch and return input output sequences, actual lengths of each sequence in the batch
and max length in the batch.
:param a_batch:
:return:
"""
global FEATURE_LENGTH, VECTOR_EMBEDDINGS, NB_ARGUMENT_LABELS, USE_BI_LSTM
_input_seqs = []
_output_seqs = []
_lengths = []
_input_fw_seqs = []
_input_bw_seqs = []
for _ds in a_batch:
_lengths.append(len(_ds.seq_features))
_max_length = max(_lengths)
for _ds in a_batch:
_seq = []
for feature in _ds.seq_features:
element = []
for l in feature:
if isinstance(l, str):
element.extend(
list(
VECTOR_EMBEDDINGS.get(l,
VECTOR_EMBEDDINGS["_unk_"])))
else:
element.append(l)
_seq.append(element)
if USE_BI_LSTM:
_input_fw_seqs.append(
util.pad_sequence(_seq, _max_length, FEATURE_LENGTH))
_input_bw_seqs.append(
util.pad_sequence(list(reversed(_seq)), _max_length,
FEATURE_LENGTH))
else:
_input_seqs.append(
util.pad_sequence(_seq, _max_length, FEATURE_LENGTH))
_output_seqs.append(
util.pad_sequence(list(_ds.seq_labels), _max_length, 1))
_input_seqs = np.array(_input_seqs)
_input_fw_seqs = np.array(_input_fw_seqs)
_input_bw_seqs = np.array(_input_bw_seqs)
_output_seqs = np.array(_output_seqs)
_lengths = np.array(_lengths)
if USE_BI_LSTM:
return _input_fw_seqs, _input_bw_seqs, _output_seqs, _lengths, _max_length
else:
return _input_seqs, _output_seqs, _lengths, _max_length
def generate_answers(sess, model, dataset, rev_vocab):
"""
Loop over the dev or test dataset and generate answer.
Note: output format must be answers[uuid] = "real answer"
You must provide a string of words instead of just a list, or start and end index
In main() function we are dumping onto a JSON file
evaluate.py will take the output JSON along with the original JSON file
and output a F1 and EM
You must implement this function in order to submit to Leaderboard.
:param sess: active TF session
:param model: a built QASystem model
:param rev_vocab: this is a list of vocabulary that maps index to actual words
:return:
"""
answers = {}
(context, question, question_uuid_data) = dataset
context_data = convert_data_to_list(context)
question_data = convert_data_to_list(question)
context_padded, context_mask = pad_sequence(context_data,
FLAGS.max_context_len)
question_padded, question_mask = pad_sequence(question_data,
FLAGS.max_question_len)
input_data = vectorize(context_padded, context_mask, question_padded,
question_mask, question_uuid_data)
batch_size = 32
num_batches = int(len(input_data) / batch_size) + 1
prog = Progbar(target=num_batches)
for i, batch in enumerate(minibatches(input_data, batch_size)):
a_s_vec, a_e_vec = model.answer(sess, batch)
prog.update(i + 1)
for (a_s, a_e, context, uuid) in zip(a_s_vec, a_e_vec, batch[0],
batch[4]):
if a_s > a_e:
tmp = a_s
a_s = a_e
a_e = tmp
predicted_answer = model.formulate_answer(context, rev_vocab, a_s,
a_e)
answers[uuid] = predicted_answer
return answers
def get_inputs(question, evidences, word2idx):
question_list = []
evidence_list = []
q_list = []
e_list = []
q_mask_list = []
e_mask_list = []
ques, q_len = get_chars(question, word2idx)
question, q_mask = pad_sequence(ques, param.question_size, word2idx)
nb_evid = len(evidences)
for i, e in enumerate(evidences):
e, e_len = get_chars(e, word2idx)
if e_len == 0: continue
other_id = random.randint(0, nb_evid - 1)
if nb_evid != 1:
while other_id == i:
other_id = random.randint(0, nb_evid - 1)
other_evidence = evidences[other_id]
other_evidence, _ = get_chars(other_evidence, word2idx)
q_feat = get_feats(ques, e)
e_feat = get_feats(other_evidence, e)
evidence, e_mask = pad_sequence(e, param.evidence_size, word2idx)
q_tags, _ = pad_sequence(q_feat, param.evidence_size, word2idx)
e_tags, _ = pad_sequence(e_feat, param.evidence_size, word2idx)
question_list.append(question)
evidence_list.append(evidence)
q_list.append(q_tags)
e_list.append(e_tags)
q_mask_list.append(q_mask)
e_mask_list.append(e_mask)
question = Variable(torch.LongTensor(question_list)).cuda()
evidence = Variable(torch.LongTensor(evidence_list)).cuda()
e_feat = Variable(torch.LongTensor(e_list)).cuda()
q_feat = Variable(torch.LongTensor(q_list)).cuda()
q_mask = Variable(torch.ByteTensor(q_mask_list)).cuda()
e_mask = Variable(torch.ByteTensor(e_mask_list)).cuda()
return question, evidence, q_mask, e_mask, q_feat, e_feat