def get_batch_generator(word2id, id2idf, qn_uuid_data, context_token_data, qn_token_data, batch_size, context_len, question_len):
"""
This is similar to get_batch_generator in data_batcher.py, but with some
differences (see explanation in refill_batches).
Inputs:
word2id: dictionary mapping word (string) to word id (int)
qn_uuid_data: list of strings that are unique ids
context_token_data, qn_token_data: list of lists of strings (no UNKs, no padding)
batch_size: int. size of batches to make
context_len, question_len: ints. max sizes of context and question. Anything longer is truncated.
Yields:
Batch objects, but they only contain context and question information (no answer information)
"""
batches = []
while True:
if len(batches) == 0:
refill_batches(batches, word2id, qn_uuid_data, context_token_data, qn_token_data, batch_size, context_len, question_len)
if len(batches) == 0:
break
# Get next batch. These are all lists length batch_size
(uuids, context_tokens, context_ids, qn_ids) = batches.pop(0)
# Pad context_ids and qn_ids
qn_ids = padded(qn_ids, question_len) # pad questions to length question_len
context_ids = padded(context_ids, context_len) # pad contexts to length context_len
# Make qn_ids into a np array and create qn_mask
qn_ids = np.array(qn_ids)
qn_mask = (qn_ids != PAD_ID).astype(np.int32)
# Make context_ids into a np array and create context_mask
context_ids = np.array(context_ids)
context_mask = (context_ids != PAD_ID).astype(np.int32)
qn_features = get_question_features(word2id, id2idf, context_ids, qn_ids, qn_mask)
cx_features = get_context_features(word2id, id2idf, context_ids, qn_ids, context_mask)
# Make into a Batch object
batch = Batch(context_ids, context_mask, context_tokens, qn_ids, qn_mask, qn_tokens=None, ans_span=None, ans_tokens=None, qn_features=qn_features, cx_features=cx_features, uuids=uuids)
yield batch
return
def get_batch_generator(word2id, qn_uuid_data, context_token_data,
qn_token_data, batch_size, context_len, question_len):
"""
This is similar to get_batch_generator in data_batcher.py, but with some
differences.
"""
batches = []
while True:
if len(batches) == 0:
refill_batches(batches, word2id, qn_uuid_data, context_token_data,
qn_token_data, batch_size, context_len,
question_len)
if len(batches) == 0:
break
(uuids, context_tokens, context_ids, qn_ids) = batches.pop(0)
qn_ids = padded(qn_ids, question_len)
context_ids = padded(context_ids, context_len)
qn_ids = np.array(qn_ids)
qn_mask = (qn_ids != PAD_ID).astype(np.int32)
context_ids = np.array(context_ids)
context_mask = (context_ids != PAD_ID).astype(np.int32)
batch = Batch(context_ids,
context_mask,
context_tokens,
qn_ids,
qn_mask,
qn_tokens=None,
ans_span=None,
ans_tokens=None,
uuids=uuids)
yield batch
return
def test_append_batch_both_different():
"""Test appending batch with both timestamps, different times."""
archiver = CsvArchiver(SAMPLE_PATH)
batch = Batch([SAMPLE_TIMESTAMP], [SAMPLE_TIMESTAMP2])
archiver.append(batch)
data_expected = {
SAMPLE_TIMESTAMP: {
LABEL_TIMESTAMP: SAMPLE_TIMESTAMP,
EventType.monitoring_started.name: 0,
EventType.monitoring_ended.name: 0,
EventType.person_entered.name: 1,
EventType.person_left.name: 0,
},
SAMPLE_TIMESTAMP2: {
LABEL_TIMESTAMP: SAMPLE_TIMESTAMP2,
EventType.monitoring_started.name: 0,
EventType.monitoring_ended.name: 0,
EventType.person_entered.name: 0,
EventType.person_left.name: 1,
},
}
assert archiver._entries == data_expected
def get_batch_generator(word2id, char2id, qn_uuid_data, context_token_data, qn_token_data, batch_size, context_len, question_len, word_len):
"""
This is similar to get_batch_generator in data_batcher.py, but with some
differences (see explanation in refill_batches).
Inputs:
word2id: dictionary mapping word (string) to word id (int)
qn_uuid_data: list of strings that are unique ids
context_token_data, qn_token_data: list of lists of strings (no UNKs, no padding)
batch_size: int. size of batches to make
context_len, question_len: ints. max sizes of context and question. Anything longer is truncated.
Yields:
Batch objects, but they only contain context and question information (no answer information)
"""
batches = []
while True:
if len(batches) == 0:
# refill_batches(batches, word2id, qn_uuid_data, context_token_data, qn_token_data, batch_size, context_len, question_len)
refill_batches(batches, word2id, char2id, qn_uuid_data, context_token_data, qn_token_data, batch_size, context_len, question_len, word_len)
if len(batches) == 0:
break
# Get next batch. These are all lists length batch_size
(uuids, context_tokens, context_ids, qn_ids,\
context_char_tokens, context_char_ids, context_char_mask,\
qn_char_tokens, qn_char_ids, qn_char_mask) = batches.pop(0)
# Pad context_ids and qn_ids
qn_ids = padded(qn_ids, question_len) # pad questions to length question_len
context_ids = padded(context_ids, context_len) # pad contexts to length context_len
# Make qn_ids into a np array and create qn_mask
qn_ids = np.array(qn_ids)
qn_mask = (qn_ids != PAD_ID).astype(np.int32)
# Make context_ids into a np array and create context_mask
context_ids = np.array(context_ids)
context_mask = (context_ids != PAD_ID).astype(np.int32)
curr_batch_size = len(qn_ids)
# Fill in blanks
context_char_ids_np = np.ones(shape=(curr_batch_size, context_len, word_len))
context_char_ids = np.array(context_char_ids)
(a,b,c) = context_char_ids.shape
context_char_ids_np[:a,:b,:c] = context_char_ids
context_char_mask_np = np.ones(shape=(curr_batch_size, context_len, word_len))
context_char_mask = np.array(context_char_mask)
context_char_mask_np[:a,:b,:c] = context_char_mask
qn_char_ids_np = np.ones(shape=(curr_batch_size, question_len, word_len))
qn_char_ids = np.array(qn_char_ids)
(a,b,c) = qn_char_ids.shape
qn_char_ids_np[:a,:b,:c] = qn_char_ids
qn_char_mask_np = np.ones(shape=(curr_batch_size, question_len, word_len))
qn_char_mask = np.array(qn_char_mask)
qn_char_mask_np[:a,:b,:c] = qn_char_mask
# word_len_chars = word_len*[1]
# context_len_chars = curr_batch_size *[context_len * [word_len_chars]]
# question_len_chars = curr_batch_size *[question_len * [word_len_chars]]
# Make into a Batch object
batch = Batch(context_ids, context_mask, context_tokens, qn_ids, qn_mask, qn_tokens=None, ans_span=None, ans_tokens=None,
char_context_ids=context_char_ids_np, char_context_mask=context_char_mask_np, char_context_tokens=context_char_tokens,
char_qn_ids=qn_char_ids_np, char_qn_mask=qn_char_mask_np, char_qn_tokens=qn_char_tokens, uuids=uuids)
yield batch
return
def get_batch_generator(word2id, qn_uuid_data, context_token_data,
qn_token_data, batch_size, context_len, question_len,
num_feats, word_len, mcids_dict):
"""
This is similar to get_batch_generator in data_batcher.py, but with some
differences (see explanation in refill_batches).
Inputs:
word2id: dictionary mapping word (string) to word id (int)
qn_uuid_data: list of strings that are unique ids
context_token_data, qn_token_data: list of lists of strings (no UNKs, no padding)
batch_size: int. size of batches to make
context_len, question_len: ints. max sizes of context and question. Anything longer is truncated.
Yields:
Batch objects, but they only contain context and question information (no answer information)
"""
batches = []
while True:
if len(batches) == 0:
refill_batches(batches, word2id, qn_uuid_data, context_token_data,
qn_token_data, batch_size, context_len,
question_len, word_len, mcids_dict)
if len(batches) == 0:
break
# Get next batch. These are all lists length batch_size
(uuids, context_tokens, context_ids, qn_ids, feats, char_ids,
commonQ_mask, commonQ_emb_indices, charQ_ids, commonC_mask,
commonC_emb_indices) = batches.pop(0)
# Pad context_ids and qn_ids
qn_ids = padded(qn_ids,
question_len) # pad questions to length question_len
context_ids = padded(context_ids,
context_len) # pad contexts to length context_len
# Make qn_ids into a np array and create qn_mask
qn_ids = np.array(qn_ids)
qn_mask = (qn_ids != PAD_ID).astype(np.int32)
# Make context_ids into a np array and create context_mask
context_ids = np.array(context_ids)
context_mask = (context_ids != PAD_ID).astype(np.int32)
# Make feats into an np array
feats = np.array(padded2(feats, num_feats, context_len))
# Pad character ids (first for word length, then for context length), then make into array
char_ids = padded2(char_ids, word_len, context_len, islist=True)
char_ids = np.array(char_ids)
char_mask = (char_ids != PAD_ID).astype(np.int32)
charQ_ids = padded2(charQ_ids, word_len, question_len, islist=True)
charQ_ids = np.array(charQ_ids)
charQ_mask = (charQ_ids != PAD_ID).astype(np.int32)
# Pad commonQ_mask and commonQ_emb_indices / convert to np.array
commonQ_mask = np.array(paddedBool(commonQ_mask, question_len))
commonQ_emb_indices = np.array(
padded(commonQ_emb_indices, question_len))
commonC_mask = np.array(paddedBool(commonC_mask, context_len))
commonC_emb_indices = np.array(padded(commonC_emb_indices,
context_len))
# Make into a Batch object
batch = Batch(context_ids, context_mask, context_tokens, qn_ids, qn_mask, qn_tokens=None, ans_span=None, ans_tokens=None, \
feats=feats, char_ids=char_ids, char_mask=char_mask, commonQ_mask=commonQ_mask, commonQ_emb_indices=commonQ_emb_indices, \
charQ_ids=charQ_ids, charQ_mask=charQ_mask, commonC_mask=commonC_mask, commonC_emb_indices=commonC_emb_indices, uuids=uuids)
yield batch
return
def test_append_empty_batch():
"""Test appending empty batch."""
archiver = CsvArchiver(SAMPLE_PATH)
batch = Batch([], [])
archiver.append(batch)
assert not archiver._entries
def convert_text_as_batch(self, context, question):
# 转为 unicode
context = context.decode("utf-8")
question = question.decode("utf-8")
# 1.分词，pos，ner，词干，tf
logger.debug("#" * 20 + " 开始转换为模型的输入 " + "#" * 20)
t1 = time.time()
context_tokens, context_poses, context_ners, context_lemmas, context_tf = tokenize_pos_ner(context)
question_tokens, question_poses, question_ners, question_lemmas, _ = tokenize_pos_ner(question)
t2 = time.time()
logger.debug("分词、词性、实体处理: {}s".format(t2 - t1))
# 2.获取文章与问题的3种匹配特征
question_tokens_lower = [w.lower() for w in question_tokens]
context_tokens_lower = [w.lower() for w in context_tokens]
exact_match = c2q_match(context_tokens, set(question_tokens)) # 精确匹配
lower_match = c2q_match(context_tokens_lower, set(question_tokens_lower)) # 小写匹配
lemma_match = c2q_match(context_lemmas, set(question_lemmas)) # 提取文章 token 的词干是否出现在问题中
context_tf_match = [[f1, f2, f3, f4] for f1, f2, f3, f4 in
zip(context_tf, exact_match, lower_match, lemma_match)]
t3 = time.time()
logger.debug("文章与问题匹配处理: {}s".format(t3 - t2))
# 3. 转为 id
context_ids = [self.word2id.get(token.lower(), UNK_ID) for token in context_tokens]
ques_ids = [self.word2id.get(token.lower(), UNK_ID) for token in question_tokens]
context_pos_ids = [self.tag2id.get(p, 0) for p in context_poses]
context_ner_ids = [self.ner2id.get(n, 0) for n in context_ners]
t4 = time.time()
logger.debug("token->id 处理: {}s".format(t4 - t3))
def batchry(item):
''' 将一条数据封装为 batch 的形式，其实就是加一层[]'''
return [item]
batch_uuids = batchry(1)
batch_context_ids = batchry(context_ids)
batch_context_tokens = batchry(context_tokens)
batch_context_pos_ids = batchry(context_pos_ids)
batch_context_ner_ids = batchry(context_ner_ids)
batch_context_features = batchry(context_tf_match)
batch_ques_ids = batchry(ques_ids)
batch_ques_tokens = batchry(question_tokens)
batch_ans_span = batchry([0, 1])
batch_ans_tokens = batchry([])
# 进行pad
context_len = self.FLAGS.context_len
ques_len = self.FLAGS.ques_len
batch_context_ids = pad(batch_context_ids, context_len)
batch_context_pos_ids = pad(batch_context_pos_ids, context_len)
batch_context_ner_ids = pad(batch_context_ner_ids, context_len)
batch_ques_ids = pad(batch_ques_ids, ques_len)
batch_context_features = pad(batch_context_features, context_len, np.array([0, 0, 0, 0]))
# np化
batch_context_ids = np.asarray(batch_context_ids)
batch_context_pos_ids = np.asarray(batch_context_pos_ids)
batch_context_ner_ids = np.asarray(batch_context_ner_ids)
batch_context_features = np.asarray(batch_context_features)
batch_ques_ids = np.asarray(batch_ques_ids)
batch_ans_span = np.asarray(batch_ans_span)
# 进行mask，只有进行np化后，才能进行这样的操作
batch_context_mask = (batch_context_ids != PAD_ID).astype(np.int32)
batch_ques_mask = (batch_ques_ids != PAD_ID).astype(np.int32)
batch = Batch(batch_context_ids, batch_context_mask, batch_context_tokens, batch_context_pos_ids,
batch_context_ner_ids,
batch_context_features, batch_ques_ids, batch_ques_mask, batch_ques_tokens, batch_ans_span,
batch_ans_tokens, batch_uuids)
t5 = time.time()
logger.debug("封装为 batch 处理: {}s".format(t5 - t4))
logger.debug("#" * 10 + " 完成转换为模型的输入，共耗时：{} ".format(t5 - t1) + "#" * 10)
return batch
