我的神经网络（多输入，多输出）给出 NaN 值

Question

我目前正在尝试训练我的神经网络并在给定点后体验 NaN 值。

输入：状态和动作

输出：状态值和动作概率

输入已先验标准化

状态大小 = (87,)

action_size = (864,)

我无法理解为什么 NaN 值会在一段时间后出现。

需要输入！

    def normalize_data(self, input_state, legal_actions):
        scaler = MinMaxScaler()
        input_state = np.array(input_state).reshape(-1, 1)
        input_state = scaler.fit_transform(input_state)

        legal_actions = np.array(legal_actions).reshape(-1, 1)
        legal_actions = scaler.fit_transform(legal_actions)
        input_state = input_state.ravel()
        legal_actions = legal_actions.ravel()
        return input_state, legal_actions

def build_network(self):
    # incorporating Residual tower
    state_input = Input(shape=self.state_size, name='Feature_Vector')
    leg_actions = Input(shape=self.action_size, name='Legal Actions')

    t = concatenate([state_input, leg_actions])
    t = BatchNormalization()(t) 
    t = Dense(2340)(t) 
    t = Activation('relu')(t)
    t = Dropout(0.0)(t)

    a = leg_actions

    actions_prob = Dense(self.action_size[0],activation='softmax')(self.policy_head(t))
    action_probs_masked = Multiply()([actions_prob, a])
    layer = Lambda(lambda x1: x1 / K.sum(x1, axis=1)[:, None])
    action_logits_masked = layer(action_probs_masked)  # probability
    self.pi = action_logits_masked
    self.v = Dense(1, activation='sigmoid', name='v')(self.value_head(t))


    self.model = Model(inputs=[state_input, leg_actions], outputs=[self.v, self.pi])
    plot_model(self.model, to_file='residual_tower_Productionline.png', show_shapes=True, show_layer_names=True)
    self.model.compile(loss=[self.mse, 'categorical_crossentropy'],optimizer=tf.keras.optimizers.Adam(0.001), metrics=['mae', 'accuracy'])
    print(self.model.summary())


    def value_head(self, input):
        x = Dense(100, name='Dense_valuehead')(input)enter image description here
        layer = tf.keras.layers.ReLU()
        x = layer(x)

        return x

    def policy_head(self, input):
        print('policy head')
        y = Dense(2160, name='Dense_policyhead')(input)
        layer = tf.keras.layers.ReLU()
        y = layer(y)

        return y