三维全局平均池化层
GlobalAveragePooling3D class
keras.layers.GlobalAveragePooling3D(data_format=None, keepdims=False, **kwargs)
3D 数据的全球平均池化操作.
参数:
data_format: 字符串,可以是 "channels_last" 或 "channels_first".
输入中维度的顺序."channels_last" 对应于输入形状为
(batch, spatial_dim1, spatial_dim2, spatial_dim3, channels)
而 "channels_first" 对应于输入形状为
(batch, channels, spatial_dim1, spatial_dim2, spatial_dim3).
它默认为在 ~/.keras/keras.json 中的 image_data_format 值.
如果你从未设置过它,那么它将是 "channels_last".
keepdims: 一个布尔值,是否保留时间维度.
如果 keepdims 为 False（默认）,张量的秩会因空间维度而减少.
如果 keepdims 为 True,空间维度将以长度 1 保留.
其行为与 tf.reduce_mean 或 np.mean 相同.
输入形状:
- 如果
data_format='channels_last': 5D 张量,形状为:(batch_size, spatial_dim1, spatial_dim2, spatial_dim3, channels) - 如果
data_format='channels_first': 5D 张量,形状为:(batch_size, channels, spatial_dim1, spatial_dim2, spatial_dim3)
输出形状:
- 如果
keepdims=False: 2D 张量,形状为(batch_size, channels). - 如果
keepdims=True: - 如果data_format="channels_last": 5D 张量,形状为(batch_size, 1, 1, 1, channels)- 如果data_format="channels_first": 5D 张量,形状为(batch_size, channels, 1, 1, 1)
示例:
>>> x = np.random.rand(2, 4, 5, 4, 3)
>>> y = keras.layers.GlobalAveragePooling3D()(x)
>>> y.shape
(2, 3)