The activation tanh seems can't fit the region of average path length

[char256]

Hi~
I'm reading your code and find that the activation function of mlp is tanh, and the output activation is softplus

pred_fun, loglike_fun, parser = build_mlp(layer_specs, output_activation=softplus)

def build_mlp(layer_sizes, activation=np.tanh, output_activation=lambda x: x):
......
    def predict(weights, X):
        cur_X = copy(X.T)
        for layer in range(len(layer_sizes) - 1):
            cur_W = parser.get(weights, ('weights', layer))
            cur_B = parser.get(weights, ('biases', layer))
            cur_Z = np.dot(cur_X, cur_W) + cur_B
            cur_X = activation(cur_Z)
        return output_activation(cur_Z.T)

    def log_likelihood(weights, X, y):
        y_hat = predict(weights, X)
        return mse(y.T, y_hat.T)

the output of tanh ranges from -1 to 1, so after the output activation(softplus), the final output is also no greater than 1, but isn't average path length can sometimes be greater than 1?