-
Notifications
You must be signed in to change notification settings - Fork 8
/
predict_result.py
189 lines (175 loc) · 7.32 KB
/
predict_result.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
import pandas
import torch
import torch.nn as nn
import torch.optim as optim
import torch.nn.functional as F
from torch.utils.data import DataLoader
# 用来评估的数据，目前约有 30k 数据
# 从 2017 年开始，因为这之后就有人介入了
n_test = 14 * 600
epochs = 1
hid = 32
n_batch = 14
# 多头并不提升准确性，只是用来保证（也不能完全保证）每次跑脚本的结果大致差不多
heads = 16
# 去掉最后一部分数据，因为结果可能还没出来
n_trim = 14
matches = sorted([
'挪超', '葡超', '英超', '德乙', '英冠', '日乙', '日职', '瑞超', '欧冠',
'解放者杯', '巴西杯', '欧罗巴', '德甲', '西甲', '意甲', '意大利杯',
'国王杯', '巴甲', '世预赛', '国际赛', '挪威杯', '联合会杯', '欧青赛',
'亚冠', '瑞典杯', '韩职', '金杯赛', '杯赛', '美职足', '日联赛杯',
'俱乐部赛', '荷超杯', '法超杯', '德国杯', '澳超', '意超杯', '社区盾杯',
'葡超杯', '英联赛杯', '西超杯', '欧超杯', '苏超', '荷甲', '法甲',
'苏联赛杯', '法联赛杯', '荷兰杯', '世青赛', '天皇杯', '亚预赛',
'英足总杯', '英甲', '葡联赛杯', '葡萄牙杯', '世俱杯', '苏足总杯',
'法国杯', '非洲杯', '阿甲', '四强赛', '日超杯', '世界杯', '德超杯',
'法乙', '欧预赛', '荷乙', '亚运男足', '东南亚锦', '亚洲杯', '瑞超杯',
'公开赛杯', '英锦标赛', '中北美冠', '英乙', '女世界杯', '美洲杯',
'俱乐部杯', '优胜者杯', '欧青预赛', '欧洲杯', '奥运女足', '奥运男足',
'女四强赛', '阿根廷杯', '圣保罗锦', '非预赛', '亚运女足', '智利甲',
'墨超', '俄超', '阿超杯', '俄超杯', '国冠杯', '墨冠杯', '墨超杯',
'澳杯', '墨西哥杯', '智利杯', '俄罗斯杯', '智超杯', '亚奥赛',
'韩足总杯', '比超杯', '比甲', '比利时杯', '中国杯', '挪超杯',
'女欧洲杯', '欧国联',
], key=len, reverse=True)
m_matches = {m: i for i, m in enumerate(['null'] + matches)}
def match(x):
for m in matches:
if m in x:
return m_matches[m]
n, m = 1, 'null'
for _m in matches:
_n = len(set(_m).intersection(x))
if _n > n:
n, m = _n, _m
return m_matches[m]
csv = pandas.read_csv('csv/total.csv')
# 对赛事与队伍增加注意力，实际效果并不好
# teams = set(csv['host']).union(csv['guest'])
# m_teams = {t: i for i, t in enumerate(teams)}
# hs = torch.from_numpy(csv['host'].map(m_teams.get).to_numpy()).long()
# gs = torch.from_numpy(csv['host'].map(m_teams.get).to_numpy()).long()
# ms = torch.from_numpy(csv['match'].map(match).to_numpy()).long()
odds = torch.from_numpy(csv[csv.columns[8:]].to_numpy()).float()
# 考虑到终赔不一定及时，可以仅使用初赔
# 但加终赔数据效果更好点，所以都在截止之前尽量晚下注
odds = odds[:, [
6 * i + 0 + j for i in range(odds.shape[1] // 6) for j in range(3)]]
d_odds = odds.shape[1]
# 将赔率转换为概率
# for i in range(0, d_odds, 3):
# idx = odds[:, i] != 0.0
# na = odds[idx]
# ret = 1 / (1 / na[:, i] + 1 / na[:, i + 1] + 1 / na[:, i + 2])
# for j in range(3):
# odds[idx, i + j] = ret / na[:, i + j]
# odds = odds - torch.tensor(
# [0.4519, 0.2474, 0.3008]
# ).unsqueeze(0).repeat(1, odds.shape[1] // 3)
odds = odds - 1
scores = torch.from_numpy(csv[['h_score', 'g_score']].to_numpy()).float()
n_size = (odds.shape[0] // 14) * 14
odds = odds[-n_size:]
scores = scores[-n_size:]
vss = [nn.Sequential(
# nn.LayerNorm(d_odds, elementwise_affine=True),
nn.Linear(d_odds, hid),
nn.LeakyReLU(),
nn.Linear(hid, hid),
nn.LeakyReLU(),
nn.Linear(hid, 3),
nn.Softmax(dim=1),
) for _ in range(heads)]
# scales = [nn.Parameter(4 * torch.ones(1, 2)) for _ in range(heads)]
# gatts = [nn.Parameter(torch.rand(1, d_odds)) for _ in range(heads)]
# matt = nn.Parameter(torch.ones(ms.shape[0], d_odds))
# tatt = nn.Parameter(torch.ones(hs.shape[0], d_odds))
opt = optim.Adam([p for vs in vss for p in vs.parameters()])
def predict(perm, i=None):
h = odds[perm]
if i is not None:
pred = vss[i](h)
else:
h = [h for i in range(heads)]
# h = [h[i] * torch.softmax(gatts[i], dim=1) for i in range(heads)]
# h = h * torch.softmax(matt[ms[perm]], dim=1)
# h = h * torch.softmax(tatt[hs[perm]] + tatt[gs[perm]], dim=1)
pred = torch.cat([
vss[i](h[i]).unsqueeze(0) for i in range(heads)
], dim=0).mean(dim=0)
real = scores[perm]
result = 1 - (real[:, 0] - real[:, 1]).sign().long()
loss = F.cross_entropy(
pred, result,
# 减少胜的预测增加平的预测
# 使预测结果的分布与实际数据一致，但总体准确性降低
# 策略中会降低下注频率但相对收益升高
# TODO: 另外我估计这种方式在计算赔率时收益会比较高
# weight=1/torch.tensor([0.4519, 0.2474, 0.3008])
)
return pred, loss
for vs in vss:
vs.train()
for perm in DataLoader(
range(n_size - n_test - n_trim),
batch_size=n_batch,
shuffle=False):
opt.zero_grad()
_, loss = predict(perm, i=torch.randint(0, heads, (1, )).item())
loss.backward()
opt.step()
total_loss = 0.0
total_pred = []
for perm in DataLoader(
range(n_size - n_test - n_trim, n_size),
batch_size=n_batch,
shuffle=False):
for vs in vss:
vs.eval()
with torch.no_grad():
pred, _ = predict(perm)
total_pred.append(pred)
for vs in vss:
vs.train()
opt.zero_grad()
_, loss = predict(perm)
total_loss += loss.item()
loss.backward()
opt.step()
print('mean loss:', total_loss / n_test)
total_pred = torch.cat(total_pred, dim=0)
pandas.DataFrame(total_pred.numpy()).to_csv('pred/results.csv', index=False)
pred = total_pred[:-n_trim]
pred_val, pred_idx = pred.max(dim=1)
real = scores[-n_test - n_trim:-n_trim]
result = 1 - (real[:, 0] - real[:, 1]).sign().long()
print('[real] win: %f, draw: %f, lose: %f' % (
(result == 0).sum() / n_test,
(result == 1).sum() / n_test,
(result == 2).sum() / n_test))
print('[pred] win: %f, draw: %f, lose: %f' % (
(pred_idx == 0).sum() / n_test,
(pred_idx == 1).sum() / n_test,
(pred_idx == 2).sum() / n_test))
print('[pred] win: %f, draw: %f, lose: %f' % (
pred[:, 0].mean(), pred[:, 1].mean(), pred[:, 2].mean()))
print('[pred-acc] win: %.2f%%, draw: %.2f%%, lose: %.2f%%' % (
100 * ((pred_idx == 0) & (result == 0)).sum() / (pred_idx == 0).sum(),
100 * ((pred_idx == 1) & (result == 1)).sum() / (pred_idx == 1).sum(),
100 * ((pred_idx == 2) & (result == 2)).sum() / (pred_idx == 2).sum(),
))
print('[pred-rec] win: %.2f%%, draw: %.2f%%, lose: %.2f%%' % (
100 * ((pred_idx == 0) & (result == 0)).sum() / (result == 0).sum(),
100 * ((pred_idx == 1) & (result == 1)).sum() / (result == 1).sum(),
100 * ((pred_idx == 2) & (result == 2)).sum() / (result == 2).sum(),
))
print('total acc: %.2f%%' % (100 * (pred_idx == result).sum() / n_test))
for conf in list(range(0, 90, 10)) + list(range(90, 100)):
flt = pred_val >= (conf / 100.0)
flt = flt & (pred_val < (conf + 10) / 100.0)
print('%.2f%% acc: %.2f%% in %d' % (
conf,
100 * (pred_idx[flt] == result[flt]).sum() / flt.sum(),
flt.sum(),
))