Tensorflow占位符，用于tf.Variable

Question

The code below throws a "no gradients" error.

self.x1 = tf.placeholder(tf.float64)
self.x2 = tf.placeholder(tf.float64)
self.x3 = tf.placeholder(tf.float64)

self.cos1_denom = tf.norm(self.x1, axis=0) * tf.norm(self.x2, axis=0)
self.cos1 = tf.matmul(self.x1, self.x2, transpose_b=True) / self.cos1_denom
self.cos2_denom = tf.norm(self.x1, axis=0) * tf.norm(self.x2, axis=0)
self.cos2 = tf.matmul(self.x1, self.x3, transpose_b=True) / self.cos2_denom
self.loss = tf.reduce_mean(self.cos2) - tf.reduce_mean(self.cos1)
self.optimizer = tf.train.AdamOptimizer(learning_rate=self.eta).minimize(self.loss)

I believe it's because my loss function depends on placeholders instead of Variables, but in my training function I pass in a Variable value for the placeholder.

Is there a way to create a placeholder for a Variable?

Answer 1

I think the problem is that there are no trainable variables in your model. Back propogation is for trainable parameters to adjust according to the loss. Unfortunately there is none here