No gradients provided for any variable

最新推荐文章于 2024-02-20 08:03:04 发布

li128ve980

最新推荐文章于 2024-02-20 08:03:04 发布

阅读量430

点赞数

文章标签： tensorflow python 深度学习

本文链接：https://blog.csdn.net/li128ve980/article/details/126000693

版权

        with tf.GradientTape() as tape:
            q_eval_parameters = self.q_parameter(states, training=True)
            # q_values = tf.squeeze(self.q_eval(np.concatenate([states, q_eval_parameters], 2)))
            q_values = tf.squeeze(self.q_eval(tf.keras.layers.concatenate([states, q_eval_parameters], 2)))
            q_values = tf.gather_nd(params=q_values, indices=enum_actions)
            loss_parameter = -sum(q_values)
        grads_parameter = tape.gradient(loss_parameter, self.q_parameter.trainable_variables)
        self.optimizer.apply_gradients(zip(grads_parameter, self.q_parameter.trainable_variables))

错误原因就是np.concat()对向量进行的拼接，而不是tf.keras.layers.concatenate()

修改后错误解决

# q_values = tf.squeeze(self.q_eval(np.concatenate([states, q_eval_parameters], 2)))


q_values = tf.squeeze(self.q_eval(tf.keras.layers.concatenate([states, q_eval_parameters], 2)))