NoisyNetを使用した強化学習モデルについて、テスト時はノイズは含まれていないのでしょうか。

Noisy Netを実装したDQNで強化学習を行っています。
学習済みモデルの評価を行っていると、Inputは同じデータなのに対して異なるQ値のリストをOutputしていました。
Noisy Netは評価時にもノイズが乗った行動をとってしまうのでしょうか？もしくは評価時は特別な処理を入れる必要があるのでしょうか。

python
1class NoisyDense(Layer):
2    def __init__(self, 
3                units=7,
4                sigma_init=0.02,
5                activation=None,
6                use_bias=True,
7                kernel_initializer='glorot_uniform',
8                bias_initializer='zeros',
9                kernel_regularizer=None,
10                bias_regularizer=None,
11                activity_regularizer=None,
12                kernel_constraint=None,
13                bias_constraint=None,
14                **kwargs):
15        if 'input_shape' not in kwargs and 'input_dim' in kwargs:
16            kwargs['input_shape'] = (kwargs.pop('input_dim'),)
17        super(NoisyDense, self).__init__(**kwargs)
18        self.units = units
19        self.sigma_init = sigma_init
20        self.activation = activations.get(activation)
21        self.use_bias = use_bias
22        self.kernel_initializer = initializers.get(kernel_initializer)
23        self.bias_initializer = initializers.get(bias_initializer)
24        self.kernel_regularizer = regularizers.get(kernel_regularizer)
25        self.bias_regularizer = regularizers.get(bias_regularizer)
26        self.activity_regularizer = regularizers.get(activity_regularizer)
27        self.kernel_constraint = constraints.get(kernel_constraint)
28        self.bias_constraint = constraints.get(bias_constraint)
29
30    def build(self, input_shape):
31        assert len(input_shape) >= 2
32        self.input_dim = input_shape[-1]
33
34        self.kernel_shape = tf.constant((self.input_dim, self.units))
35        self.bias_shape = tf.constant((self.units, ))
36
37        self.kernel = self.add_weight(shape=(self.input_dim, self.units),
38                                      initializer=self.kernel_initializer,
39                                      name='kernel',
40                                      regularizer=self.kernel_regularizer,
41                                      constraint=self.kernel_constraint)
42
43        self.sigma_kernel = self.add_weight(shape=(self.input_dim, self.units),
44                                      initializer=initializers.Constant(value=self.sigma_init),
45                                      name='sigma_kernel'
46                                      )
47
48        if self.use_bias:
49            self.bias = self.add_weight(shape=(self.units,),
50                                        initializer=self.bias_initializer,
51                                        name='bias',
52                                        regularizer=self.bias_regularizer,
53                                        constraint=self.bias_constraint)
54            self.sigma_bias = self.add_weight(shape=(self.units,),
55                                        initializer=initializers.Constant(value=self.sigma_init),
56                                        name='sigma_bias')
57        else:
58            self.bias = None
59            self.epsilon_bias = None
60
61        self.epsilon_kernel = K.zeros(shape=(self.input_dim, self.units))
62        self.epsilon_bias = K.zeros(shape=(self.units,))
63
64        self.sample_noise()
65        super(NoisyDense, self).build(input_shape)
66
67    def call(self, X):
68        perturbation = self.sigma_kernel * K.random_uniform(shape=self.kernel_shape)
69        perturbed_kernel = self.kernel + perturbation
70        output = K.dot(X, perturbed_kernel)
71        if self.use_bias:
72            bias_perturbation = self.sigma_bias * K.random_uniform(shape=self.bias_shape)
73            perturbed_bias = self.bias + bias_perturbation
74            output = K.bias_add(output, perturbed_bias)
75        if self.activation is not None:
76            output = self.activation(output)
77        return output
78
79    def compute_output_shape(self, input_shape):
80        assert input_shape and len(input_shape) >= 2
81        assert input_shape[-1]
82        output_shape = list(input_shape)
83        output_shape[-1] = self.units
84        return tuple(output_shape)
85
86    def sample_noise(self):
87        K.set_value(self.epsilon_kernel, np.random.normal(0, 1, (self.input_dim, self.units)))
88        K.set_value(self.epsilon_bias, np.random.normal(0, 1, (self.units,)))
89
90    def remove_noise(self):
91        K.set_value(self.epsilon_kernel, np.zeros(shape=(self.input_dim, self.units)))
92        K.set_value(self.epsilon_bias, np.zeros(shape=self.units,))