array slicing for convolutional block-CodePudding

I get confused by a array slicing notation.

def hypernetwork(self, inputs):
    x = self.fc(inputs)

    return x[..., :self.channels], x[..., self.channels:]

what is the return? what does ..., means? the self.channels is defined as the number of channels of input. I think x is just the input feature block. Below is the relevant code for self.fc and self.channels

def build(self, input_shape):
    self.channels = input_shape[0][-1]  # input_shape: [x, z].

    self.fc = KL.Dense(
        int(2 * self.channels),
        kernel_initializer=self.init,
        kernel_regularizer=tf.keras.regularizers.l2(
            l=self.wt_decay,
        ),
        bias_regularizer=tf.keras.regularizers.l2(
            l=self.wt_decay,
        ),
    )

CodePudding user response：

... means all elements of all dimensions until you start explicitely referencing, which you do with the :self.channels. So in conclusion if x is e.g. a 10x4x6 array and self.channels is 4, the output will be one 10x4x4 array and one 10x4x2 array. If x is 10x6 and self.channels is 2, you'll get a 10x2 and a 10x4 array. You're splitting the array along the last dimension.

CodePudding user response：

You are referring to all the dimensions except the last one (which you are slicing) when using .... It is equivalent to the other notation x[:, :, :channels]:

import tensorflow as tf
tf.random.set_seed(111)
channels = 2
x = tf.random.normal((1, 2, 3))
print(x)
print(x[..., :channels], x[:, :, :channels]) # Equivalent
print(x[..., channels:], x[:, :, channels:]) # Equivalent

tf.Tensor(
[[[ 0.7558127   1.5447265   1.6315602 ]
  [-0.19868968  0.08828261  0.01711658]]], shape=(1, 2, 3), dtype=float32)
tf.Tensor(
[[[ 0.7558127   1.5447265 ]
  [-0.19868968  0.08828261]]], shape=(1, 2, 2), dtype=float32) tf.Tensor(
[[[ 0.7558127   1.5447265 ]
  [-0.19868968  0.08828261]]], shape=(1, 2, 2), dtype=float32)
tf.Tensor(
[[[1.6315602 ]
  [0.01711658]]], shape=(1, 2, 1), dtype=float32) tf.Tensor(
[[[1.6315602 ]
  [0.01711658]]], shape=(1, 2, 1), dtype=float32)