获取特征图的均值和减少通道

Question

我有一个大小为 1 x 5 x 5 x 32 的特征图，我想将特征图通道从 1 x 5 x 5 x 32 减少到 1 x 5 x 5 x 1（从 32 个通道减少到1 个通道）。唯一的通道是原来32个通道的平均值，怎么办？

下面是我的特征图输出：

        convOutputs:  tf.Tensor(
        [[[[ 0.21890974  0.2801039  -0.02241993 -0.02482897 -0.02028599
            -0.00421949 -0.03049909  0.12293314 -0.00455996  0.16527294
             0.25661796  0.08303508 -0.00884647  0.18990402  0.17974378
            -0.00271507 -0.01113621 -0.02265824 -0.01211608 -0.02467087
            -0.01819963 -0.01684237  0.26299486 -0.01183819 -0.01383558
            -0.01530357 -0.02485864 -0.00031154  0.21875194 -0.0014731
            -0.02675386 -0.00685479]
           [ 0.21890974  0.2801039  -0.02241993 -0.02482897 -0.02028599
            -0.00421949 -0.03049909  0.12293314 -0.00455996  0.16527294
             0.25661796  0.08303508 -0.00884647  0.18990402  0.17974378
            -0.00271507 -0.01113621 -0.02265824 -0.01211608 -0.02467087
            -0.01819963 -0.01684237  0.26299486 -0.01183819 -0.01383558
            -0.01530357 -0.02485864 -0.00031154  0.21875194 -0.0014731
            -0.02675386 -0.00685479]
          ...
           [ 0.21890974  0.2801039  -0.02241993 -0.02482897 -0.02028599
            -0.00421949 -0.03049909  0.12293314 -0.00455996  0.16527294
             0.25661796  0.08303508 -0.00884647  0.18990402  0.17974378
            -0.00271507 -0.01113621 -0.02265824 -0.01211608 -0.02467087
            -0.01819963 -0.01684237  0.26299486 -0.01183819 -0.01383558
            -0.01530357 -0.02485864 -0.00031154  0.21875194 -0.0014731
            -0.02675386 -0.00685479]
           [ 0.21890974  0.2801039  -0.02241993 -0.02482897 -0.02028599
            -0.00421949 -0.03049909  0.12293314 -0.00455996  0.16527294
             0.25661796  0.08303508 -0.00884647  0.18990402  0.17974378
            -0.00271507 -0.01113621 -0.02265824 -0.01211608 -0.02467087
            -0.01819963 -0.01684237  0.26299486 -0.01183819 -0.01383558
            -0.01530357 -0.02485864 -0.00031154  0.21875194 -0.0014731
            -0.02675386 -0.00685479]]]], shape=(1, 5, 5, 32), dtype=float32)

Answer 1

尝试：

import tensorflow as tf

x = tf.random.normal((1, 5, 5, 32))
x = tf.reduce_mean(x, axis=-1, keepdims=True)
print(x.shape)
# (1, 5, 5, 1)

获取特征图的均值和减少通道

Get mean and reduce channels of feature maps

python

deep-learning

tensorflow