def load_dataset(sentences, tags, word_to_id, tag_to_id, lower=False):
"""
Prepare the dataset. Return a list of lists of dictionaries containing:
- word indexes
- tag indexes
"""
def f(x):
return x.lower() if lower else x
data = []
for i in range(len(sentences)):
str_words = sentences[i]
words = [
word_to_id[f(w) if f(w) in word_to_id else '<UNK>']
for w in str_words
]
caps = [cap_feature(w) for w in str_words]
tag_ids = [tag_to_id[t] for t in tags[i]]
data.append({
'str_words': str_words,
'words': words,
'caps': caps,
'tags': tag_ids,
'pos': tag_ids,
})
return data
def preprocess(self, path, draft):
output = []
stopwords = [' ', '\n', '\u3000', '\u202f', '\u2009']
with open(path, 'r', encoding='utf-8') as f:
with open(path, 'r', encoding = 'utf-8') as t:
data = []
for line in t:
data.append(json.loads(line))
t.close()
# pdb.set_trace()
if draft:
data[0]['data'] = data[0]['data'][:1]
for topic in data[0]['data']:
for paragraph in topic['paragraphs']:
context = paragraph['context']
tokens = word_tokenize(context)
for qa in paragraph['qas']:
qid = qa['id']
question = qa['question']
for ans in qa['answers']:
answer = ans['text']
s_idx = ans['answer_start']
e_idx = s_idx + len(answer)
l = 0
s_found = False
for i, t in enumerate(tokens):
while l < len(context):
if context[l] in stopwords:
l += 1
else:
break
if t[0] == '"' and context[l:l + 2] == '\'\'':
t = '\'\'' + t[1:]
elif t == '"' and context[l:l + 2] == '\'\'':
t = '\'\''
l += len(t)
if l > s_idx and s_found == False:
s_idx = i
s_found = True
if l >= e_idx:
e_idx = i
break
output.append(dict([('qid', qid),
('context', context),
('question', question),
('answer', answer),
('start_idx', s_idx),
('end_idx', e_idx)]))
with open('{}l'.format(path), 'w', encoding='utf-8') as f:
for line in output:
json.dump(line, f)
print('', file=f)
def process_file(f, encoding='utf8'):
"""初始化函数
Args:
f[str] : 原始数据文件
encoding[str] : 编码(默认utf8)
"""
data = []
tag = []
for line in f:
x = ''
y = ''
nr_flag = False
nt_flag = False
nr_word = ''
nt_word = ''
#ignore the ID
for pair in line.strip().split()[1:]:
word = pair.split('/')[0]
#split sentence with token '。'
if word == u'。' and len(x) > 0:
lines.append(x+'\n')
labels.append(y.strip()+'\n')
x = ''
y = ''
continue
#process nt words
if pair.startswith('['):
nt_flag = True
nt_word = word[1:]
continue
if nt_flag:
if not pair.endswith(']nt'):
nt_word += word
continue
# process nr tag
if pair.endswith('nr'):
nr_word += word
elif len(nr_word) > 0:
x += nr_word
y += ' B-PER'+' I-PER'*(len(nr_word)-1)
nr_word = ''
if pair.endswith('nt'):
if pair.endswith(']nt'):
word = nt_word+word
nt_flag = False
x += word
y += ' B-ORG'+' I-ORG'*(len(word)-1)
# process ns tag
elif pair.endswith('ns'):
x += word
y += ' B-LOC'+' I-LOC'*(len(word)-1)
else:
x += word
y += ' O'*(len(word))
if len(x) > 0:
data.append(x+'\n')
tag.append(y.strip()+'\n')
return data, tag
def get_test_data():
val_data = pd.read_csv('./Dataset/test.csv')
data = []
for content, label in zip(val_data['content'], val_data['label']):
data.append((content, label))
random.shuffle(data)
contents, True_labels = [],[]
for content, label in data:
contents.append(content)
True_labels.append(label)
return contents, True_labels
def json_to_csv(file, output_file):
data = []
with open(file) as f:
for line in f:
data.append(json.loads(line))
df = pd.DataFrame.from_records(data)[[
'sentence1', 'sentence2', 'gold_label'
]]
df['gold_label'] = df['gold_label'].map(values_dict)
df = df[df['gold_label'].isin([0, 1, 2])]
df.to_csv(output_file, index=False)
def preprocess_data(input_list,
type="train",
letter_width=0.02,
hop_length=400):
data = []
for i in input_list:
data.append(
preprocess_data_single_entry(i,
type=type,
letter_width=letter_width,
hop_length=hop_length))
return data
def add_padding(data, max_length, index_to_pad, index_to_eos):
'''
增加padding
:param data:
:param max_length: 每行最大单词数
:param index_to_pad: pad对应的索引号
:param index_to_eos: 行结束符的索引号
:return:
'''
data = data[:max_length - 1] #减1是为了在最后添加eos这个token索引
padding_len = (max_length - 1) - len(data)
assert padding_len >= 0, 'padding_len = max_length - len(data)'
data.extend([index_to_pad] * padding_len)
data.append(index_to_eos)
return data
def user_builder(df):
'''Given the flattened dataframe, output a user-based collection of data.'''
data = []
user_ids = df.user_id.unique()
for i in tnrange(len(user_ids)):
user = user_ids[i]
sub_df = df.loc[df['user_id'] == user].reset_index(drop=True)
country = sub_df['country_id'][0]
join_date = sub_df['joining_date'][0]
sorted_df = sub_df.sort_values('unix_timestamp').reset_index(drop=True)
history = sorted_df['cities'].tolist()
timestamps = sorted_df['unix_timestamp'].tolist()
history = [(timestamps[j], history[j]) for j in range(len(sorted_df))]
data.append((user, country, join_date, history))
return data
def data_process(src_path, tgt_path, src_tokenizer, trg_tokenizer, src_vocab,
trg_vocab):
src_iter = iter(io.open(src_path, encoding="utf8"))
trg_iter = iter(io.open(tgt_path, encoding="utf8"))
data = []
for raw_src, raw_trg in zip(src_iter, trg_iter):
raw_src, raw_trg = preprocess(raw_src), preprocess(raw_trg)
# print(f"src:{raw_src}, target: {raw_trg}")
src_tensor = torch.tensor(
[src_vocab[token] for token in src_tokenizer(raw_src)],
dtype=torch.long)
trg_tensor = torch.tensor(
[trg_vocab[token] for token in trg_tokenizer(raw_trg)],
dtype=torch.long)
data.append((src_tensor, trg_tensor))
# print(f"data: {data}")
return data
def evaluate(model, data_iter, tipe, epoch, fold):
model.eval()
corrects, avg_loss = 0, 0
data_iter.sort_key = lambda x: len(x.text)
for batch in data_iter:
text_numerical, target = batch.text, batch.label
if args.cuda:
text_numerical, target = text_numerical.cuda(), target.cuda()
text_numerical.data.t_()
target.data.sub_(1)
forward = model(text_numerical)
loss = F.cross_entropy(forward, target, size_average=False)
avg_loss += loss.data[0]
corrects += (torch.max(forward, 1)[1].view(
target.size()).data == target.data).sum()
size = len(data_iter.dataset)
avg_loss = avg_loss / size
accuracy = 100.0 * corrects / size
# print('Avg Loss = {}'.format(avg_loss))
output['fold_{}'.format(fold + 1)]['epoch_{}'.format(
epoch)][tipe]['avg_loss'] = avg_loss.item()
if tipe == 'testing':
if args.cuda:
data.append(accuracy.item())
else:
data.append(accuracy)
return target.data, torch.max(forward, 1)[1].view(target.size()).data
type=int,
default=1,
help='number of ngrams to be selected in each batch [default: 10]')
parser.add_argument('-threshold',
type=float,
default=0.9,
help='threshold for selecting ngram [default: 0.9]')
args = parser.parse_args()
# load up data
data = []
with open('/mnt/storage01/milliet/data/ngrams/clean-ngrams-score-9500.csv',
'r') as csvfile:
lines = csvfile.readlines()
for line in lines:
data.append(line.split('\sep'))
threshold = 0.9999
bests = list(elem for elem in data if float(elem[-1]) > threshold)
print("N-grams with score > " + str(threshold) + ": " + str(len(bests)))
if len(bests) > 0:
args.summary_dir = os.path.join(
args.save_dir,
datetime.datetime.now().strftime('%Y-%m-%d_%H-%M-%S') + "threshold_" +
str(threshold) + "_model_dict_" + args.state_dict)
writer = SummaryWriter(log_dir=args.summary_dir)
x_data = list(
re.sub("\s+", ",", elem[1][2:-2]).split(',') for elem in bests)
def augmentation(data):
augmented = []
aug_df = pd.DataFrame({'summary': [], 'rating': []})
reps = []
for index, game in data.iterrows():
genre = game['rating']
if (genre == 'Dislike'):
s = 38
tok = tokenize_word(game['summary'])
elif (genre == 'Acclaim'):
s = 14
tok = tokenize_word(game['summary'])
else:
s = 3
tok = tokenize_sent(game['summary'])
# if(genre in ('Turn-based strategy (TBS)')):
# s = 120
# tok = tokenize_word(game['summary'])
# elif(genre in ('Tactical')):
# s = 110
# tok = tokenize_word(game['summary'])
# elif(genre in ('Quiz/Trivia')):
# s = 100
# tok = tokenize_word(game['summary'])
# elif(genre in ('Music')):
# s = 45
# tok = tokenize_word(game['summary'])
# elif(genre in ('Arcade')):
# s = 17
# tok = tokenize_word(game['summary'])
# elif(genre in ("Hack and slash/Beat 'em up")):
# s = 21
# tok = tokenize_word(game['summary'])
# elif(genre in ('Racing')):
# s = 20
# tok = tokenize_word(game['summary'])
# elif(genre in ('Sport')):
# s = 23
# tok = tokenize_word(game['summary'])
# elif(genre in ('Indie')):
# s = 18
# tok = tokenize_sent(game['summary'])
# elif(genre in ('Real Time Strategy (RTS)')):
# s = 12
# tok = tokenize_sent(game['summary'])
# elif(genre in ('Strategy')):
# s = 1
# tok = tokenize_sent(game['summary'])
# elif(genre in ('Fighting')):
# s = 7
# tok = tokenize_sent(game['summary'])
# elif(genre in ('Point-and-click')):
# s = 6
# tok = tokenize_sent(game['summary'])
# elif(genre in ('Simulator')):
# s = 7
# tok = tokenize_sent(game['summary'])
# elif(genre in ('Adventure')):
# s = 3
# tok = tokenize_sent(game['summary'])
# elif(genre in ('Platform')):
# s = 2
# tok = tokenize_sent(game['summary'])
# elif(genre in ('Puzzle')):
# s = 4
# tok = tokenize_sent(game['summary'])
# else:
# s = 0
# tok = tokenize_sent(game['summary'])
shuffled = [tok]
#print(ng_rev)
for i in range(s):
#generate 11 new reviews
shuffled.append(shuffle_tokenized(shuffled[-1]))
for k in shuffled:
'''create new review by joining the shuffled sentences'''
s = ' '
new_game = s.join(k)
if new_game not in augmented:
augmented.append(new_game)
aug_df = aug_df.append(
pd.DataFrame({
'summary': [new_game],
'rating': [genre]
}))
else:
reps.append(new_game)
return data.append(aug_df)
