def get_data(self):
"""
轨迹比较特殊，原子文件中存储的并不是轨迹而是一个一个点，因此需要先对轨迹进行切割
"""
if self.data is None:
if self.config['cache_dataset'] and os.path.exists(
self.encoder.cache_file_name):
# load cache
f = open(self.encoder.cache_file_name, 'r')
self.data = json.load(f)
self.pad_item = self.data['pad_item']
f.close()
else:
cut_data = self.cutter_filter()
encoded_data = self.encode_traj(cut_data)
self.data = encoded_data
self.pad_item = self.encoder.pad_item
if self.config['cache_dataset']:
if not os.path.exists(self.cache_file_folder):
os.makedirs(self.cache_file_folder)
with open(self.encoder.cache_file_name, 'w') as f:
json.dump(encoded_data, f)
# user 来划，以及按轨迹数来划。
# TODO: 这里可以设一个参数，现在先按照轨迹数来划吧
train_data, eval_data, test_data = self.divide_data()
return generate_dataloader(train_data, eval_data, test_data,
self.encoder.feature_dict,
self.config['batch_size'],
self.config['num_workers'], self.pad_item,
self.encoder.feature_max_len)
def get_data(self):
"""
返回数据的DataLoader，包括训练数据、测试数据、验证数据
Returns:
tuple: tuple contains:
train_dataloader: Dataloader composed of Batch (class) \n
eval_dataloader: Dataloader composed of Batch (class) \n
test_dataloader: Dataloader composed of Batch (class)
"""
# 加载数据集
x_train, y_train, x_val, y_val, x_test, y_test = [], [], [], [], [], []
if self.data is None:
self.data = {}
if self.cache_dataset and os.path.exists(self.cache_file_name):
x_train, y_train, x_val, y_val, x_test, y_test = self._load_cache_train_val_test(
)
else:
x_train, y_train, x_val, y_val, x_test, y_test = self._generate_train_val_test(
)
# 数据归一化
self.feature_dim = x_train.shape[-1]
self.scaler = self._get_scalar(x_train, y_train)
x_train[..., :self.output_dim] = self.scaler.transform(
x_train[..., :self.output_dim])
y_train[..., :self.output_dim] = self.scaler.transform(
y_train[..., :self.output_dim])
x_val[..., :self.output_dim] = self.scaler.transform(
x_val[..., :self.output_dim])
y_val[..., :self.output_dim] = self.scaler.transform(
y_val[..., :self.output_dim])
x_test[..., :self.output_dim] = self.scaler.transform(
x_test[..., :self.output_dim])
y_test[..., :self.output_dim] = self.scaler.transform(
y_test[..., :self.output_dim])
if self.normal_external:
x_train[..., self.output_dim:] = self.scaler.transform(
x_train[..., self.output_dim:])
y_train[..., self.output_dim:] = self.scaler.transform(
y_train[..., self.output_dim:])
x_val[..., self.output_dim:] = self.scaler.transform(
x_val[..., self.output_dim:])
y_val[..., self.output_dim:] = self.scaler.transform(
y_val[..., self.output_dim:])
x_test[..., self.output_dim:] = self.scaler.transform(
x_test[..., self.output_dim:])
y_test[..., self.output_dim:] = self.scaler.transform(
y_test[..., self.output_dim:])
# 把训练集的X和y聚合在一起成为list，测试集验证集同理
# x_train/y_train: (num_samples, input_length, ..., feature_dim)
# train_data(list): train_data[i]是一个元组，由x_train[i]和y_train[i]组成
train_data = list(zip(x_train, y_train))
eval_data = list(zip(x_val, y_val))
test_data = list(zip(x_test, y_test))
# 转Dataloader
self.train_dataloader, self.eval_dataloader, self.test_dataloader = \
generate_dataloader(train_data, eval_data, test_data, self.feature_name,
self.batch_size, self.num_workers, pad_with_last_sample=self.pad_with_last_sample)
return self.train_dataloader, self.eval_dataloader, self.test_dataloader, train_data
def get_data(self):
"""
轨迹比较特殊，原子文件中存储的并不是轨迹而是一个一个点，因此需要先对轨迹进行切割
"""
if self.data is None:
if self.config['cache_dataset'] and os.path.exists(
self.cache_file_name):
# load cache
f = open(self.cache_file_name, 'r')
self.data = json.load(f)
loc_pad = self.data['loc_size'] - 1
tim_pad = self.data['tim_size'] - 1
self.pad_item = {
'current_loc': loc_pad,
'history_loc': loc_pad,
'current_tim': tim_pad,
'history_tim': tim_pad,
'text': np.zeros((self.data['text_size']))
}
f.close()
else:
transformed_data = self.cutter_filter()
# pad parameter
loc_pad = transformed_data['loc_size']
transformed_data['loc_size'] += 1
tim_pad = transformed_data['tim_size']
transformed_data['tim_size'] += 1
self.pad_item = {
'current_loc': loc_pad,
'history_loc': loc_pad,
'current_tim': tim_pad,
'history_tim': tim_pad,
'text': np.zeros((transformed_data['text_size']))
}
self.data = transformed_data
if self.config['cache_dataset']:
if not os.path.exists(self.cache_file_folder):
os.makedirs(self.cache_file_folder)
with open(self.cache_file_name, 'w') as f:
json.dump(transformed_data, f)
# 切完轨迹之后，就是做 batch 了
# 划分训练集、测试集、验证集：有两种方式按 user 来划，以及按轨迹数来划。
# TODO: 这里可以设一个参数，现在先按照轨迹数来划吧
train_data, eval_data, test_data = self.gen_input()
return generate_dataloader(train_data, eval_data, test_data,
self.feature_name,
self.config['batch_size'],
self.config['num_workers'], self.pad_item,
self.pad_max_len)
def get_data(self):
"""
获取数据，数据归一化，之后返回训练集、测试集、验证集对应的DataLoader
:return:
train_dataloader (pytorch.DataLoader)
eval_dataloader (pytorch.DataLoader)
test_dataloader (pytorch.DataLoader)
all the dataloaders are composed of Batch (class)
"""
# 加载数据集
x_train, y_train, x_val, y_val, x_test, y_test = [], [], [], [], [], []
ext_x_train, ext_y_train, ext_x_test, ext_y_test, ext_x_val, ext_y_val = [], [], [], [], [], []
if self.data is None:
self.data = {}
if self.cache_dataset and os.path.exists(self.cache_file_name):
x_train, y_train, x_val, y_val, x_test, y_test, \
ext_x_train, ext_y_train, ext_x_test, ext_y_test, ext_x_val, ext_y_val \
= self._load_cache_train_val_test()
else:
x_train, y_train, x_val, y_val, x_test, y_test, \
ext_x_train, ext_y_train, ext_x_test, ext_y_test, ext_x_val, ext_y_val \
= self._generate_train_val_test()
# 数据归一化
self.feature_dim = x_train.shape[-1]
self.scaler = self._get_scalar(x_train, y_train)
x_train = self.scaler.transform(x_train)
y_train = self.scaler.transform(y_train)
x_val = self.scaler.transform(x_val)
y_val = self.scaler.transform(y_val)
x_test = self.scaler.transform(x_test)
y_test = self.scaler.transform(y_test)
if self.normal_external:
ext_x_train = self.scaler.transform(ext_x_train)
ext_y_train = self.scaler.transform(ext_y_train)
ext_x_val = self.scaler.transform(ext_x_val)
ext_y_val = self.scaler.transform(ext_y_val)
ext_x_test = self.scaler.transform(ext_x_test)
ext_y_test = self.scaler.transform(ext_y_test)
# 把训练集的X和y聚合在一起成为list，测试集验证集同理
# x_train/y_train: (num_samples, input_length, ..., feature_dim)
# train_data(list): train_data[i]是一个元组，由x_train[i]和y_train[i]组成
train_data = list(zip(x_train, y_train, ext_x_train, ext_y_train))
eval_data = list(zip(x_val, y_val, ext_x_val, ext_y_val))
test_data = list(zip(x_test, y_test, ext_x_test, ext_y_test))
# 转Dataloader
self.train_dataloader, self.eval_dataloader, self.test_dataloader = \
generate_dataloader(train_data, eval_data, test_data, self.feature_name,
self.batch_size, self.num_workers, pad_with_last_sample=self.pad_with_last_sample)
return self.train_dataloader, self.eval_dataloader, self.test_dataloader
def get_data(self):
"""
获取切割后的数据集
Returns:
"""
if self.data is None:
if self.config['cache_dataset_load'] and os.path.exists(
self.encoder.cache_file_name):
# 加载cache
f = open(self.encoder.cache_file_name, 'r')
self.data = json.load(f)
f.close()
else:
dyna = pd.read_csv(
os.path.join(self.data_path, '{}.dyna'.format(
self.config['dataset']))).values.T
geo = pd.read_csv(
os.path.join(self.data_path, '{}.geo'.format(
self.config['dataset']))).values.T
self.geo = geo
sequence, sequence_user, sequence_time, sequence_distance, id_to_geo = self.build_sequence(
dyna)
top_n = self.pop_n(sequence, self.num_sample)
train_set = (sequence, sequence_user, sequence_time,
sequence_distance)
final_train_set, final_eval_test, final_test_set = \
self.load_data(
train_set,
self.num_sample,
top_n,
id_to_geo,
True)
self.data = {
"train": final_train_set,
"eval": final_eval_test,
"test": final_test_set
}
# 先不管这块
if self.config['cache_dataset_save']:
if not os.path.exists(self.cache_file_folder):
os.makedirs(self.cache_file_folder)
with open(self.encoder.cache_file_name, 'w') as f:
json.dump(self.data, f)
print("finish getting data!")
# 这里没传self.pad_item
return generate_dataloader(train_data=self.data["train"],
eval_data=self.data["eval"],
test_data=self.data["test"],
feature_name={
'loc': 'float',
'tim': 'float',
'dis': 'float',
'uid': 'float',
'loc_neg': 'float',
'tim_neg': 'float',
'dis_neg': 'float',
'target': 'int'
},
batch_size=self.config['batch_size'],
num_workers=self.config['num_workers'],
shuffle=False)
