categorical.py 文件源码

def tf_parameterize(self, x):
        # Flat logits
        logits = self.logits.apply(x=x)

        # Reshape logits to action shape
        shape = (-1,) + self.shape + (self.num_actions,)
        logits = tf.reshape(tensor=logits, shape=shape)

        # !!!
        state_value = tf.reduce_logsumexp(input_tensor=logits, axis=-1)

        # Softmax for corresponding probabilities
        probabilities = tf.nn.softmax(logits=logits, dim=-1)

        # Min epsilon probability for numerical stability
        probabilities = tf.maximum(x=probabilities, y=util.epsilon)

        # "Normalized" logits
        logits = tf.log(x=probabilities)

        return logits, probabilities, state_value