code_srt_sgwi_v1 / data /generate_labels.py
natmin322's picture
Upload folder using huggingface_hub
e841b45 verified
"""
instruction 中任务的标签集合作为选项提供,故每个任务应当提供label文件
"""
import os
import json
def label_collect(task_path, task_collect_fun, filter_path=None, max_labels=30):
"""
在任务路径下,为每个数据集,生成一个labels.json
labels.json 包含内容为所有标签的list,以NER为例:
['person', 'location', 'organization']
Args:
task_path: task files path which contains datasets dir
task_collect_fun: task specific label collect function
filter_path: filter datasets with too much labels
Returns:
"""
filter_ds_dirs = []
for dirpath, dirnames, filenames in os.walk(task_path):
for dirname in dirnames:
ds_dir = os.path.join(dirpath, dirname)
labels = []
# 假设全部包含train、dev、test数据集
for ds_type in ['train', 'dev', 'test']:
ds_type_file = os.path.join(ds_dir, ds_type+'.json')
ds_type_labels = task_collect_fun(ds_type_file)
labels += ds_type_labels
labels = list(set(labels))
out_file = os.path.join(ds_dir, 'labels.json')
json.dump(labels, open(out_file, 'w+', encoding='utf-8'), ensure_ascii=False)
print('Finish out {} labels to {}!'.format(len(labels), out_file))
if len(labels) >= max_labels:
filter_dir = os.path.join(filter_path, dirname)
filter_ds_dirs.append([ds_dir, filter_dir])
# filter datasets by mv them to another dirs
for ds_dir, filter_dir in filter_ds_dirs:
cmd = "mv {} {}".format(ds_dir, filter_dir)
os.system(cmd)
print('Move {} to {} cz too much labels!'.format(ds_dir, filter_dir))
def NER_label_collect(file_path):
"""
按照输入文件路径,收集标签,按照list返回。
输入文件为一个样例json列表,返回为标签集合
Args:
file_path: dataset file name
Returns:
[label1, label2]
"""
fi = open(file_path, 'r+', encoding='utf-8')
samples = json.load(fi)
labels = []
for sample in samples:
for entity in sample['entities']:
if entity['type'] not in labels:
labels.append(entity['type'])
return labels
if __name__ == "__main__":
NER_path = '/root/InstructUIE/data/NER_processed'
filter_path = '/root/InstructUIE/data/NER_filter'
label_collect(NER_path, NER_label_collect, filter_path)