multivariate.py 文件源码

def _sample(self, n_samples):
        logits, temperature = self.logits, self.temperature
        if not self.is_reparameterized:
            logits = tf.stop_gradient(logits)
            temperature = tf.stop_gradient(temperature)
        shape = tf.concat([[n_samples], tf.shape(self.logits)], 0)

        uniform = open_interval_standard_uniform(shape, self.dtype)
        # TODO: Add Gumbel distribution
        gumbel = -tf.log(-tf.log(uniform))
        samples = tf.nn.softmax((logits + gumbel) / temperature)

        static_n_samples = n_samples if isinstance(n_samples, int) else None
        samples.set_shape(
            tf.TensorShape([static_n_samples]).concatenate(logits.get_shape()))
        return samples