He (Kaiming) initialization is designed for networks using ReLU activations. Since ReLU zeros out negative values (roughly half the distribution), He initialization uses a larger variance than Xavier to compensate, depending only on the fan-in.
Given a raw weight matrix W with values in [0, 1] and fan_in (number of input units), scale the weights to He uniform initialization.
Return a matrix of four-decimal floats with the same shape as W.
Input: W = [[0.5, 0.5]], fan_in = 2
Output: [[0.0, 0.0]]
Explanation: A raw midpoint of 0.5 maps to the center of the symmetric He range.
Input: W = [[0], [1]], fan_in = 2
Output: [[-1.7321], [1.7321]]
Compute the symmetric limit from fan_in only.
Map each raw value from [0, 1] into the interval from negative limit to positive limit.
Accepts: array
Accepts: number
He (Kaiming) initialization is designed for networks using ReLU activations. Since ReLU zeros out negative values (roughly half the distribution), He initialization uses a larger variance than Xavier to compensate, depending only on the fan-in.
Given a raw weight matrix W with values in [0, 1] and fan_in (number of input units), scale the weights to He uniform initialization.
Return a matrix of four-decimal floats with the same shape as W.
Input: W = [[0.5, 0.5]], fan_in = 2
Output: [[0.0, 0.0]]
Explanation: A raw midpoint of 0.5 maps to the center of the symmetric He range.
Input: W = [[0], [1]], fan_in = 2
Output: [[-1.7321], [1.7321]]
Compute the symmetric limit from fan_in only.
Map each raw value from [0, 1] into the interval from negative limit to positive limit.
Accepts: array
Accepts: number