GlobalAveragePooling3D
类keras.layers.GlobalAveragePooling3D(data_format=None, keepdims=False, **kwargs)
针对 3D 数据的全局平均池化操作。
参数
"channels_last"
或 "channels_first"
。输入张量的维度顺序。"channels_last"
对应形状为 (batch, spatial_dim1, spatial_dim2, spatial_dim3, channels)
的输入,而 "channels_first"
对应形状为 (batch, channels, spatial_dim1, spatial_dim2, spatial_dim3)
的输入。默认使用 Keras 配置文件 ~/.keras/keras.json
中 image_data_format
的值。如果您从未设置过,则默认为 "channels_last"
。keepdims
为 False
(默认值),则张量的空间维度会被移除,从而降低张量的秩。如果 keepdims
为 True
,则空间维度将被保留,但长度为 1。其行为与 tf.reduce_mean
或 np.mean
相同。输入形状
data_format='channels_last'
:形状为 (batch_size, spatial_dim1, spatial_dim2, spatial_dim3, channels)
的 5D 张量。data_format='channels_first'
:形状为 (batch_size, channels, spatial_dim1, spatial_dim2, spatial_dim3)
的 5D 张量。输出形状
keepdims=False
:形状为 (batch_size, channels)
的 2D 张量。keepdims=True
:- 如果 data_format="channels_last"
:形状为 (batch_size, 1, 1, 1, channels)
的 5D 张量。- 如果 data_format="channels_first"
:形状为 (batch_size, channels, 1, 1, 1)
的 5D 张量。示例
>>> x = np.random.rand(2, 4, 5, 4, 3)
>>> y = keras.layers.GlobalAveragePooling3D()(x)
>>> y.shape
(2, 3)