距離行列の作成とコスト計算ログ(暫定版)

はじめに

Q＆A形式でのログです。

参考URL

[1] Approximating Wasserstein distances with PyTorch - Daniel Daza
https://dfdazac.github.io/sinkhorn.html

[2] Github Approximating Wasserstein distances with PyTorch
https://github.com/dfdazac/wassdistance

[1]で外観を得て、[2]のソースコードを活用させて頂いた。

Q. AとBは10個のテンソルが集約されたテンソルです。距離行列Cを求めてください。(出力は10×8×8になります。)

Q. また、A,B,Cを使って、sinkhoonアルゴリズムを実装してください。(出力は各テンソルに対応するスカラーが集約された10次元のテンソルになります。)

import torch

from layers import SinkhornDistance

A = torch.randn(5000,8,3)
B = torch.randn(5000,8,3)

sinkhorn = SinkhornDistance(eps=0.1, max_iter=1000)

dist, P, C = sinkhorn(A, B)


print("Cost:",C)
print("Distance:", dist)

Cost: tensor([[[ 2.3851,  4.6197,  6.1333,  ...,  8.1374,  0.0954, 10.4732],
         [ 2.0316,  3.5151,  7.5441,  ..., 16.0305,  5.8873,  1.9347],
         [ 2.7555,  5.4268,  8.6332,  ..., 15.5158, 10.2188,  0.1254],
         ...,
         [ 6.5027,  4.5277,  1.2859,  ...,  9.0828,  8.6347, 17.3766],
         [ 2.9556,  4.4240,  8.7556,  ..., 16.0003,  3.1574,  5.7351],
         [ 4.4386,  9.3323, 11.8327,  ...,  9.0434,  1.1514, 11.5495]],

        [[ 3.9119,  2.5504, 16.9356,  ...,  0.5636,  2.8750,  2.3992],
         [ 0.7460,  1.6460,  6.8128,  ...,  3.4929,  1.6600,  2.0364],
         [ 2.9862,  1.4779,  6.7706,  ...,  4.1522,  0.6850,  5.9932],
         ...,
         [ 1.4956,  1.4795,  5.9848,  ...,  3.4524,  1.0665,  3.1406],
         [ 9.0050, 15.3521,  1.6283,  ..., 19.0885, 14.6266,  8.7985],
         [ 6.1278,  2.8331, 26.2087,  ...,  7.9378,  4.2873, 14.4930]],

        [[ 6.5756,  1.6820,  2.3338,  ...,  0.0749, 12.2661,  4.8055],
         [ 5.5638,  2.3422,  5.7793,  ...,  5.9754,  2.9283,  0.5275],
         [ 8.5439,  2.2426,  2.3104,  ...,  0.5533, 14.1425,  5.5368],
         ...,
         [ 1.7040, 10.5317,  8.9386,  ...,  8.9560,  5.0410, 10.4021],
         [12.1518,  1.4655,  9.7433,  ...,  6.5903, 10.1003,  0.9371],
         [ 1.2980,  7.9191,  7.1047,  ...,  6.7856,  4.4195,  7.9719]],

        ...,

        [[ 0.3512,  6.7879,  1.0256,  ...,  1.1962,  6.1207,  3.1671],
         [ 1.3076, 12.6213,  1.2475,  ...,  4.0746,  9.3167,  5.8667],
         [ 6.8642, 11.7086, 14.6859,  ..., 15.9577,  5.3421,  2.2187],
         ...,
         [ 4.8603, 22.7114,  8.9389,  ..., 11.1563, 15.9644,  4.2383],
         [ 0.5606,  4.8859,  1.8659,  ...,  1.5179,  4.5857,  2.5412],
         [ 9.8800,  5.5608, 13.2379,  ..., 17.1964,  1.4715, 11.7824]],

        [[ 2.8790,  5.0972,  7.1846,  ...,  3.3449, 13.3123,  8.6139],
         [ 5.2767,  1.2954,  5.3868,  ...,  3.5217,  1.5166,  0.1790],
         [ 8.9345,  5.4933, 13.6296,  ...,  8.7750,  7.9416,  1.5231],
         ...,
         [ 7.8432,  7.8624, 15.5582,  ...,  8.9348, 15.2594,  6.0558],
         [ 2.2326,  2.1241,  0.5018,  ...,  0.5880,  5.4888,  7.2493],
         [ 2.3860,  2.9177,  4.9353,  ...,  4.0292,  6.4979,  4.7113]],

        [[ 8.0692,  3.8253,  1.9432,  ...,  7.2783,  3.0775, 12.3579],
         [ 1.9585,  4.3045,  2.8538,  ...,  2.8123, 11.6873, 13.2805],
         [ 4.3140,  5.0203,  6.8623,  ...,  3.9087, 15.3001, 13.2837],
         ...,
         [12.6561,  4.3500,  4.0669,  ..., 10.0359,  1.3259, 11.6383],
         [10.6419,  1.2867,  3.7827,  ...,  4.8285,  1.3232,  3.9288],
         [ 5.9991,  3.8937,  0.7738,  ...,  5.1932,  3.8849, 11.1763]]])

Distance: tensor([2.4712, 1.6183, 1.8624,  ..., 1.9867, 1.9327, 2.3939])

これより、コスト値の集まりが求まった。
(今回は実装ではなく、[2]のソースコードを活用させていただいた。)