回答編集履歴
2
修正
test
CHANGED
|
@@ -96,11 +96,19 @@
|
|
|
96
96
|
|
|
97
97
|
import timeit
|
|
98
98
|
|
|
99
|
-
|
|
99
|
+
import numpy as np
|
|
100
100
|
|
|
101
101
|
from numba import jit
|
|
102
102
|
|
|
103
103
|
|
|
104
|
+
|
|
105
|
+
|
|
106
|
+
|
|
107
|
+
np.random.seed(0)
|
|
108
|
+
|
|
109
|
+
X = np.random.randint(0, 9, (20, 15, 13, 10, 10))
|
|
110
|
+
|
|
111
|
+
Y = np.random.randint(0, 9, (9, 10, 10))
|
|
104
112
|
|
|
105
113
|
|
|
106
114
|
|
1
修正
test
CHANGED
|
@@ -75,3 +75,99 @@
|
|
|
75
75
|
|
|
76
76
|
|
|
77
77
|
要素数は変わらないので、メモリ使用量は増えないと思います。
|
|
78
|
+
|
|
79
|
+
|
|
80
|
+
|
|
81
|
+
|
|
82
|
+
|
|
83
|
+
# 追記
|
|
84
|
+
|
|
85
|
+
|
|
86
|
+
|
|
87
|
+
質問のループバージョンのコードを numba で最適化したところ、
|
|
88
|
+
|
|
89
|
+
218ms が 8.45 ms と25倍程度高速化できました。
|
|
90
|
+
|
|
91
|
+
ブロードキャストで計算したバージョンが 8.01 ms なのでほぼ同じぐらいの速度が出るようになりました。
|
|
92
|
+
|
|
93
|
+
|
|
94
|
+
|
|
95
|
+
```python
|
|
96
|
+
|
|
97
|
+
import timeit
|
|
98
|
+
|
|
99
|
+
|
|
100
|
+
|
|
101
|
+
from numba import jit
|
|
102
|
+
|
|
103
|
+
|
|
104
|
+
|
|
105
|
+
|
|
106
|
+
|
|
107
|
+
def calc1(X, Y):
|
|
108
|
+
|
|
109
|
+
# broadcast バージョン
|
|
110
|
+
|
|
111
|
+
X = np.expand_dims(X, axis=3)
|
|
112
|
+
|
|
113
|
+
Z = np.max(X * Y, axis=(-2, -1))
|
|
114
|
+
|
|
115
|
+
|
|
116
|
+
|
|
117
|
+
|
|
118
|
+
|
|
119
|
+
def calc2(X, Y):
|
|
120
|
+
|
|
121
|
+
# for-loop バージョン
|
|
122
|
+
|
|
123
|
+
Z = np.zeros((X.shape[0], X.shape[1], X.shape[2], Y.shape[0]), dtype=X.dtype)
|
|
124
|
+
|
|
125
|
+
for a in range(X.shape[0]):
|
|
126
|
+
|
|
127
|
+
for b in range(X.shape[1]):
|
|
128
|
+
|
|
129
|
+
for c in range(X.shape[2]):
|
|
130
|
+
|
|
131
|
+
for f in range(Y.shape[0]):
|
|
132
|
+
|
|
133
|
+
Z[a, b, c, f] = np.max(X[a, b, c] * Y[f])
|
|
134
|
+
|
|
135
|
+
|
|
136
|
+
|
|
137
|
+
|
|
138
|
+
|
|
139
|
+
@jit(nopython=True)
|
|
140
|
+
|
|
141
|
+
def calc3(X, Y):
|
|
142
|
+
|
|
143
|
+
# for-loop バージョンに numba でコンパイルしたバージョン
|
|
144
|
+
|
|
145
|
+
Z = np.zeros((X.shape[0], X.shape[1], X.shape[2], Y.shape[0]), dtype=X.dtype)
|
|
146
|
+
|
|
147
|
+
for a in range(X.shape[0]):
|
|
148
|
+
|
|
149
|
+
for b in range(X.shape[1]):
|
|
150
|
+
|
|
151
|
+
for c in range(X.shape[2]):
|
|
152
|
+
|
|
153
|
+
for f in range(Y.shape[0]):
|
|
154
|
+
|
|
155
|
+
Z[a, b, c, f] = np.max(X[a, b, c] * Y[f])
|
|
156
|
+
|
|
157
|
+
|
|
158
|
+
|
|
159
|
+
# Jupyter Notebook 上で計測
|
|
160
|
+
|
|
161
|
+
%timeit calc1(X, Y)
|
|
162
|
+
|
|
163
|
+
%timeit calc2(X, Y)
|
|
164
|
+
|
|
165
|
+
%timeit calc3(X, Y)
|
|
166
|
+
|
|
167
|
+
# 8.01 ms ± 239 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
|
|
168
|
+
|
|
169
|
+
# 218 ms ± 32.9 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)
|
|
170
|
+
|
|
171
|
+
# 8.45 ms ± 25.3 µs per loop (mean ± std. dev. of 7 runs, 1 loop each)
|
|
172
|
+
|
|
173
|
+
```
|