順序データを入力し、勝敗データに変換してからqNEIを求めるプログラムをかいているのですが、どのように順序データを入力していけばいいかわかりません。

前提・実現したいこと

該当のソースコード

python
1ソースコード
2import os
3
4import numpy as np
5import matplotlib.pyplot as plt
6from scipy.stats import zscore
7
8from botorch.models.pairwise_gp import PairwiseGP, PairwiseLaplaceMarginalLogLikelihood
9from botorch.fit import fit_gpytorch_model
10from botorch.acquisition import qNoisyExpectedImprovement
11from botorch.optim import optimize_acqf
12import torch
13import itertools
14from create_function import utility_3D, generate_train_x, generate_comparisons
15from normalization_max_min import norm_max_min
16
17IMAGE_DIR = "./accuracy/qNEI1_qNEI2/image_3D"
18if not os.path.exists(IMAGE_DIR):
19    # ディレクトリが存在しない場合、ディレクトリを作成する
20    os.makedirs(IMAGE_DIR)
21
22
23def observation_max_points(model, results):
24    observation_point = model.posterior(results).mean.tolist()
25    next_x_index = observation_point.index(max(observation_point))
26    next_x = results[next_x_index]
27
28    return next_x
29
30
31def plot_3d(x, y, image_name):
32    plt.clf()
33    y_2D_12 = y.mean(axis=2)
34    plt.contourf(x, x, y_2D_12)
35    plt.savefig(image_name + "_12.png")
36    plt.clf()
37    y_2D_13 = y.mean(axis=1)
38    plt.contourf(x, x, y_2D_13)
39    plt.savefig(image_name + "_13.png")
40    plt.clf()
41    y_2D_23 = y.mean(axis=0)
42    plt.contourf(x, x, y_2D_23)
43    plt.savefig(image_name + "_23.png")
44
45
46def main(m, loop, noise, bounds, test_f_name, acq_session=10):
47    mse_list = []
48    points = torch.linspace(0, 1, 11)
49    points_0, points_1, points_2 = torch.meshgrid(points, points, points)
50    points_mesh = torch.stack((points_0, points_1, points_2), dim=-1)
51    val_y = utility_3D(points_mesh).reshape(11, 11, 11).numpy()
52
53    # 検証用の関数(2次元だけ取り出してあとは平均値)
54    image_name = "./accuracy/qNEI1_qNEI2/image_3D/" + test_f_name + "/val_2D"
55    plot_3d(points, val_y, image_name=image_name)
56
57    for i in range(loop):
58        # 1回目の試行
59        train_X = generate_train_x(n=2, dim=3)
60        train_y = utility_3D(train_X)
61        train_comp = generate_comparisons(train_X[0], train_X[1])
62
63        for j in range(m):
64            if j < m - (acq_session+1):
65                model = PairwiseGP(train_X, train_comp)
66                mll = PairwiseLaplaceMarginalLogLikelihood(model)
67                mll = fit_gpytorch_model(mll)
68
69                pred_y = norm_max_min(model.posterior(points_mesh).mean.squeeze().detach().numpy(),
70                                      max_=val_y.max(), min_=val_y.min())
71
72                mse = np.square(val_y - pred_y).mean()
73                mse_list.append(mse)
74                print("MSE", i, j, ":", mse)
75
76                # 予測平均をplot(2次元だけ取り出してあとは平均値)
77                image_name = "./accuracy/qNEI1_qNEI2/image_3D/" + test_f_name + "/pred_y_mean_2D_" + str(
78                    i + 1) + "_" + str(j + 1)
79                plot_3d(points, pred_y, image_name=image_name)
80
81                train_X_another = generate_train_x(n=2)
82                train_X = torch.cat((train_X, train_X_another), dim=0)
83                train_comp_another = generate_comparisons(train_X_another[0], train_X_another[1])
84                train_comp_another = train_comp_another + (2 * j + 1)
85                train_comp = torch.cat((train_comp, train_comp_another), dim=0)
86                train_y = torch.cat((train_y, utility_3D(train_X_another)), dim=0)
87                image_name = "./accuracy/qNEI1_qNEI2/image_3D/" + test_f_name + "/pred_y_mean_2D_" + str(
88                    i + 1) + "_" + str(j + 1)
89                plot_3d(points, pred_y, image_name=image_name)
90
91            else:
92                qNEI = qNoisyExpectedImprovement(
93                    model=model,
94                    X_baseline=train_X.float()
95                )
96
97                next_X, acq_val = optimize_acqf(
98                    acq_function=qNEI,
99                    bounds=bounds,
100                    q=2,
101                    num_restarts=5,
102                    raw_samples=256
103                )
104
105                next_X = torch.cat((next_X[0].reshape(-1, 3), next_X[1].reshape(-1, 3)), dim=0).type(torch.float64)
106                train_X = torch.cat((train_X, next_X), dim=0)
107
108                if utility_3D(next_X[0]) > utility_3D(next_X[1]):
109                    preference = torch.Tensor([len(train_X) - 2, len(train_X) - 1]).long().reshape(-1, 2)
110                    train_comp = torch.cat((train_comp, preference), dim=0)
111                else:
112                    preference = torch.Tensor([len(train_X) - 1, len(train_X) - 2]).long().reshape(-1, 2)
113                    train_comp = torch.cat((train_comp, preference), dim=0)
114
115                model = PairwiseGP(train_X, train_comp)
116                mll = PairwiseLaplaceMarginalLogLikelihood(model)
117                mll = fit_gpytorch_model(mll)
118                pred_y = norm_max_min(model.posterior(points_mesh).mean.squeeze().detach().numpy()
119                                      , max_=val_y.max(), min_=val_y.min())
120
121                mse = np.square(val_y - pred_y).mean()
122                mse_list.append(mse)
123                print("MSE", i, j, ":", mse)
124
125                # 予測平均をplot(2次元だけ取り出してあとは平均値)
126                image_name = "./accuracy/qNEI1_qNEI2/image_3D/" + test_f_name + "/pred_y_mean_2D_" + str(
127                    i + 1) + "_" + str(j + 1)
128                plot_3d(points, pred_y, image_name=image_name)
129
130    number_of_comparisons = []
131    for k in range(m):
132        number_of_comparisons.append(k)
133
134    mse_list = np.array(mse_list).reshape(loop, -1)
135    rmse_list = np.sqrt(mse_list)
136    rmse_list_mean = rmse_list.mean(axis=0)
137    rmse_list_sd = rmse_list.std(axis=0)
138
139    plt.clf()
140    plt.ylim(0, 1)
141    plt.plot(number_of_comparisons, rmse_list_mean)
142    plt.fill_between(number_of_comparisons,
143                     rmse_list_mean - rmse_list_sd,
144                     rmse_list_mean + rmse_list_sd,
145                     alpha=0.5
146                     )
147    plt.savefig("./accuracy/qNEI1_qNEI2/image_3D/" + test_f_name + "/rmse_mean.pdf")
148

試したこと

if utility_3D(next_X[0]) > utility_3D(next_X[1]):
preference = torch.Tensor([len(train_X) - 2, len(train_X) - 1]).long().reshape(-1, 2)
train_comp = torch.cat((train_comp, preference), dim=0)
else:
preference = torch.Tensor([len(train_X) - 1, len(train_X) - 2]).long().reshape(-1, 2)
train_comp = torch.cat((train_comp, preference), dim=0)

おそらくこの部分をいじるのかと思うのですが...

補足情報（FW/ツールのバージョンなど）

import itertools
import torch
import numpy as np

num = np.array(["a","b","c","d","e","f","g","h","i","j","k","l","m","n","o"])
score = np.array([70, 90 ,80 ,50 ,30, 40, 55, 85, 95, 20, 10, 45, 35, 25, 100])
l = list(zip(num, score))

for i in itertools.combinations(l,2):
win = i[1][1] > i[0][1]
lose = not win
print(i[win][0], i[lose][0]) #勝ち負け

このプログラムのように順序データを勝敗データに変換したいと考えています。