We use tf.keras.layers.Conv2D time to build layer convolution weights initialization method generally used method is the default value of this function, i.e. 'glorot_uniform' .
Its source code to make an explanation:
'''
It draws samples from a uniform distribution within [-limit, limit]
where `limit` is `sqrt(6 / (fan_in + fan_out))`
where `fan_in` is the number of input units in the weight tensor
and `fan_out` is the number of output units in the weight tensor.
'''
That is, uniform distribution of this range initialization method using [-limit, limit] , wherein,
Here
It represents the number of input neurons,
It indicates the number of output neurons, namely:
Assume that the input of a network28 * 28 * 1data, the shape of the convolution kernel is a3 * 3convolution core has channels32th (i.e., the number of output channels with a32th), then the timelimitis: