ccbe291ec9859698a57147b1033d503a7739dfa9,tensorforce/core/distributions/categorical.py,Categorical,create_tf_operations,#Categorical#Any#Any#,54

Before Change

    def create_tf_operations(self, x, deterministic):
        flat_size = util.prod(self.shape) * self.num_actions
        bias = [prob for _ in range(util.prod(self.shape)) for prob in self.probabilities]
        logits = layers["linear"](x=x, size=flat_size, bias=bias)
        logits = tf.reshape(tensor=logits, shape=((-1,) + self.shape + (self.num_actions,)))
        self.probabilities = tf.nn.softmax(logits=logits)

After Change

        self.logits = tf.reshape(tensor=self.logits, shape=shape)

        // linearly shift logits for numerical stability
        self.logits -= tf.reduce_max(input_tensor=self.logits, axis=-1, keep_dims=True)

        // softmax for corresponding probabilities
        self.probabilities = tf.nn.softmax(logits=self.logits, dim=-1)

Instances