-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathload_data.py
313 lines (273 loc) · 12.5 KB
/
load_data.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
# 数据：
# open close high low quantity volumn
import torch
from torch.utils.data import Dataset, DataLoader, BatchSampler, SequentialSampler
import numpy as np
import json
import os
import copy
# 自定义数据集类
class StocksDataset(Dataset):
def __init__(self,
prices,
indexes,
label_name=["000133.XSHG"],
data_start_from=1158,
data_end_at=3000,
num1=500,
num2=100,
use_rand=False,
embedding_config={},
only_test=False,
only_need_result=False,
):
time_slices = [0, -1, -2, -3, -4, -6, -8, -10, -12, -14,
-17, -20, -23, -26, -29, -33, -37, -41, -45, -49,
-54, -59, -64, -69, -74, -80, -86, -92, -98, -104, -111, -118, -125, -132, -142, -154]
time_slices_label = [1, 2, 3, 5, 8, 16, 28, 44, 64]
self.history_len = len(time_slices)
self.predict_len = len(time_slices_label)
self.data_start_from = data_start_from
self.data_end_at = data_end_at
self.label_name = label_name
self.label_data = None
self.label_mean = None
self.label_std = None
self.num1 = num1
self.num2 = num2
self.bianhao_code_dict = {}
self.code_bianhao_dict = {}
self.key_indexes_list = []
self.use_rand = use_rand
self.embedding_config = embedding_config
# 每支股票在data_block中的行数到 embedding 中的编号
self.position_to_embedding_num = []
self.only_need_result = only_need_result
if only_need_result:
self.time_range = -time_slices[-1] - 1
else:
self.time_range = time_slices_label[-1] - time_slices[-1]
self.time_slices = np.array(time_slices)
self.time_slices_label = np.array(time_slices_label)
prices_and_indexes = copy.copy(prices)
prices_and_indexes.update(indexes)
self.prices_and_indexes_data_block = self.get_data_block(prices_and_indexes)
self.position_to_embedding_num = np.array(self.position_to_embedding_num)
# mean = np.nanmean(self.prices_data_block, axis=1)
# mean = np.expand_dims(mean, 1)
# std = np.nanstd(self.prices_data_block, axis=1)
# std = np.expand_dims(std, 1)
# self.prices_data_block = (self.prices_data_block - mean) / std
# self.indexes_data_block = self.get_data_block(indexes)
# mean = np.nanmean(self.indexes_data_block, axis=1)
# mean = np.expand_dims(mean, 1)
# std = np.nanstd(self.indexes_data_block, axis=1)
# std = np.expand_dims(std, 1)
# self.indexes_data_block = (self.indexes_data_block - mean) / std
could_train_len = self.prices_and_indexes_data_block.shape[1] - self.time_range - 1
print("could_train_len: ", could_train_len)
self.data = list(range(could_train_len))
if self.num1 == 0:
self.num1 = len(self.prices_and_indexes_data_block)
# if self.num2 == 0:
# self.num2 = len(self.indexes_data_block)
def get_data_block(self, prices):
'''
将数据转化为npy的array形式
prices: pd式数据
'''
filtered_st_prices = {}
# 过滤当前已停牌股票
for key, value in prices.items():
if value.iloc[-1].isnull().sum() > 0:
continue
value = value[self.data_start_from: self.data_end_at]
filtered_st_prices[key] = value
filtered_st_prices_list = []
# means = {}
# stds = {}
ct = 0
for key1, value in filtered_st_prices.items():
print(ct, key1)
if key1 not in self.embedding_config:
print(key1, " 没有记录 embedding 编号，错误")
self.bianhao_code_dict[ct] = key1
self.code_bianhao_dict[key1] = ct
value_tmp = np.expand_dims(value.values, axis=0)
filtered_st_prices_list.append(value_tmp)
# if value_tmp.shape[1] < 3:
# import time
# time.sleep(100)
self.position_to_embedding_num.append(self.embedding_config[key1])
# if key1 in self.label_name:
# self.key_indexes[key1] = ct
ct += 1
# self.label_data = filtered_st_prices_list[-1]
# self.label_mean = tmp_mean
# self.label_std = tmp_std
# print("key indexes: ", self.key_indexes)
prices_with_nan = np.concatenate(filtered_st_prices_list, axis=0)
prices_with_nan = prices_with_nan[:, :, :]
# prices_with_nan = torch.tensor(prices_with_nan)
# prices_with_nan = torch.transpose(prices_with_nan, 0, 1)
return prices_with_nan
def get_non_nan_data_row(self, data_block, idx, bias, num, must_have_indexes=[]):
'''
data_block: [stock, time, feature]
idx: time index
num: train stocks num
idx表示时间，训练哪一天的股票
从idx对应的日期开始，往后 self.time_slices[-1] 天的数据作为训练数据
将这段时间里不包含nan的股票选出来组成 [stock, time, feature] 的数据块，其中time的长度与self.time_slices有关
'''
# print(idx)
# num = 100
# bias = 0
# must_have_indexes=[]
# data_block = self.prices_and_indexes_data_block[:, 1200:, :]
# idx = 0
stocks_num = data_block.shape[0]
idx = idx - self.time_slices[-1]
# tmp = data_block[:, self.time_slices + idx, :]
# 将 time_slices 每两个相邻的日期的数据取出，取平均值
slice_all = []
slice_tmp = np.expand_dims(data_block[:, idx, :], axis=1)
slice_all.append(slice_tmp)
for i in range(self.history_len - 1):
# start = self.time_slices[i]
# end = self.time_slices[i + 1]
slice_tmp = np.mean(data_block[:, self.time_slices[i + 1] + idx: self.time_slices[i] + idx, :], axis=1)
slice_tmp = np.expand_dims(slice_tmp, axis=1)
slice_all.append(slice_tmp)
tmp = np.concatenate(slice_all, axis=1)
tmp2 = copy.copy(tmp)
# 将数据做随机化扰动，避免过拟合
if self.use_rand:
rand_vars = np.linspace(0.01, 0.08, self.history_len)
stocks_num, time_range, feature_size = tmp.shape
rands = []
for var in rand_vars:
rands.append(np.random.normal(1, var, (stocks_num, 1, feature_size)))
rand_array = np.concatenate(rands, axis=1)
tmp = tmp * rand_array
# tmp = item[4][0].numpy()
# open、close、high、low、volume 数据取变化率情况，避免过拟合
jizhun = tmp[:, 0, :].copy()
jizhun = np.expand_dims(jizhun, axis=1)
jizhun[:, :, -1] = 1
tmp = (tmp / jizhun) - 1
tmp = tmp[:, 1:, :]
# print("tmp1: ", tmp)
# money 数据取对数，反应市值情况
tmp[:, :, -1] = np.log10(tmp[:, :, -1])
# tmp.size = [stock, time, price], 对特征做归一化
tmp_mean = np.mean(tmp, axis=0)
tmp_mean = np.expand_dims(tmp_mean, axis=0)
tmp_std = np.std(tmp, axis=0)
tmp_std = np.expand_dims(tmp_std, axis=0)
tmp = (tmp - tmp_mean) / tmp_std
# print("tmp2: ", tmp)
# # tmp = raw_prices[0].numpy()
# jizhun = tmp[:, 0, :].copy()
# jizhun = np.expand_dims(jizhun, axis=1)
# tmp = (tmp / jizhun) - 1
# tmp = tmp[:, 1:, :]
# # stock, time, price
# tmp_mean = np.mean(tmp, axis=0)
# tmp_mean = np.expand_dims(tmp_mean, axis=0)
# tmp_std = np.std(tmp, axis=0)
# tmp_std = np.expand_dims(tmp_std, axis=0)
# tmp = (tmp - tmp_mean) / tmp_std
# 取出不包含nan数据的股票
stocks_indexes = np.array(list(range(stocks_num)))
stock_not_nan_indexes = ~np.isnan(tmp).any(axis=2).any(axis=1)
# np.save("tmp.npy", tmp)
# print("tmp: ", tmp)
stock_not_nan_indexes = stocks_indexes[stock_not_nan_indexes]
# print(stock_not_nan_indexes)
# print(idx)
# print(num)
if num > 0:
stock_final_indexes = np.random.choice(stock_not_nan_indexes, num, replace=False)
else:
stock_final_indexes = stock_not_nan_indexes
if must_have_indexes:
stock_final_indexes = [item for item in stock_final_indexes if item not in must_have_indexes]
stock_final_indexes = np.concatenate([must_have_indexes, stock_final_indexes])[: num]
# print(stock_final_indexes)
stock_final_indexes.sort()
stocks_final = tmp[stock_final_indexes, :, :]
# print("stock_not_nan_indexes: ", stock_not_nan_indexes)
# print("tmp3: ", stocks_final)
# stocks_final = torch.tensor(stocks_final, dtype=torch.float32).cuda()
# stock_final_indexes += bias
embedding_bianhao = self.position_to_embedding_num[stock_final_indexes]
return stocks_final, stock_final_indexes, embedding_bianhao, tmp2
def get_label(self, data_block, idx, must_have_indexes=[0]):
# if self.only_need_result:
# stocks_final = np.zeros([len(must_have_indexes), 7])
# return stocks_final
idx = idx - self.time_slices[-1]
slice_all = []
for i in range(self.predict_len - 1):
# start = self.time_slices_label[i]
# end = self.time_slices_label[i + 1]
slice_tmp = np.mean(data_block[:, self.time_slices_label[i] + idx: self.time_slices_label[i + 1] + idx, :], axis=1)
if self.only_need_result:
# print("111", slice_tmp.shape)
# print(slice_tmp)
if np.isnan(slice_tmp).sum() > 0:
# print(222)
slice_tmp = data_block[:, -1]
# print(slice_tmp)
# print(333)
# print(slice_tmp)
slice_tmp = np.expand_dims(slice_tmp, axis=1)
slice_all.append(slice_tmp)
tmp = np.concatenate(slice_all, axis=1)
# print("tmp ", tmp)
tmp = tmp[must_have_indexes]
label_raw = copy.copy(tmp)
jizhun = tmp[:, 0, :].copy()
jizhun = np.expand_dims(jizhun, axis=1)
tmp = (tmp / jizhun) - 1
tmp = tmp[:, 1:, :]
# print("tmp2 ", tmp)
# stock, time, price
tmp_mean = np.mean(tmp, axis=0)
# print("tmp_mean ", tmp_mean)
tmp_mean = np.expand_dims(tmp_mean, axis=0)
tmp_std = np.std(tmp, axis=0)
tmp_std[tmp_std == 0] = 1
tmp_std = np.expand_dims(tmp_std, axis=0)
tmp = (tmp - tmp_mean) / tmp_std
# print("tmp3 ", tmp)
stocks_final = tmp[:, :, 0]
return stocks_final, label_raw
def __len__(self):
return len(self.data)
def __getitem__(self, idx):
stock_prices, stock_indexes, embedding_bianhao, stock_prices_raw = self.get_non_nan_data_row(self.prices_and_indexes_data_block, idx, bias=0, num=self.num1)
# index_prices, index_indexes = self.get_non_nan_data_row(self.indexes_data_block, idx, bias=10000, num=self.num2, must_have_indexes=[])
label_prices, label_raw = self.get_label(self.prices_and_indexes_data_block, idx, must_have_indexes=stock_indexes)
# normal
sample1 = stock_prices
# sample1 = np.concatenate([stock_prices[:, :, :], index_prices[:, :, :]], axis=0)
# rand = np.random.normal(1, 0.02, size=sample1.shape)
# sample1 = sample1 * rand
# cheat
# sample1 = np.concatenate([stock_prices, index_prices], axis=0)
# sample1 = index_prices[0: 2, :, :]
# sample1 = np.expand_dims(sample1, axis=0)
# jizhun = sample1[:, 0, :].copy()
# jizhun = np.expand_dims(jizhun, axis=1)
# jizhun[:, :, :4] = 1
# print(jizhun)
# sample1 = (sample1 / jizhun) - 1
# sample2 = np.concatenate([stock_indexes, index_indexes], axis=0)
sample2 = np.array(embedding_bianhao)
# sample2 = np.array([0, 0])
# label = index_prices[0, self.history_len:, 0] / index_prices[0, 0, 0]
sample1 = torch.tensor(sample1, dtype=torch.float32)
sample1 = sample1[:, :, :6]
return sample1, sample2, label_prices, idx, stock_prices_raw, stock_indexes, label_raw