chainer.functions.gumbel_softmax

chainer.functions.gumbel_softmax(log_pi, tau=0.1, axis=1)

Gumbel-Softmax sampling function.

This function draws samples $y_i$ from Gumbel-Softmax distribution,

$$y_i = {\exp((g_i + \log\pi_i)/\tau) \over \sum_{j}\exp((g_j + \log\pi_j)/\tau)},$$

where $\tau$ is a temperature parameter and $g_i$ s are samples drawn from Gumbel distribution $Gumbel(0, 1)$

See Categorical Reparameterization with Gumbel-Softmax.

Parameters

• log_pi (Variable or N-dimensional array) – Input variable representing pre-normalized log-probability $\log\pi$.

• tau (float or Variable or N-dimensional array) – Input variable representing temperature $\tau$.

Returns

Output variable.

Return type

Variable