41024c61c0737b1beaea8fff8e00a947d6b6ee9b,knowledge_transfer.py,KnowledgeTransferLearner,build_networks,#KnowledgeTransferLearner#,41

Before Change

            loss = -tf.reduce_sum(eligibility)
            self.losses.append(loss)
            optimizer = tf.train.RMSPropOptimizer(learning_rate=self.config["learning_rate"], decay=0.9, epsilon=1e-9)
            self.trainers.append(optimizer.minimize(loss))

        init = tf.global_variables_initializer()

After Change

        self.variation_probs = [tf.nn.softmax(tf.matmul(L1, tf.matmul(knowledge_base, s))) for s in sparse_representations]
        self.optimizer = tf.train.RMSPropOptimizer(learning_rate=self.config["learning_rate"], decay=0.9, epsilon=1e-9)
        net_vars = self.shared_vars + sparse_representations
        self.accum_grads = create_accumulative_gradients_op(net_vars, 1)

        // self.writers = []
        self.losses = []
        for i, probabilities in enumerate(self.variation_probs):
            good_probabilities = tf.reduce_sum(tf.mul(probabilities, tf.one_hot(tf.cast(self.action_taken, tf.int32), self.nA)), reduction_indices=[1])
            eligibility = tf.log(good_probabilities) * self.advantage
            // eligibility = tf.Print(eligibility, [eligibility], first_n=5)
            loss = -tf.reduce_sum(eligibility)
            self.losses.append(loss)
            // writer = tf.summary.FileWriter(self.monitor_dir + "/task" + str(i), self.sess.graph)

        // An add op for every task & its loss
        // add_accumulative_gradients_op(net_vars, accum_grads, loss, identifier)
        self.add_accum_grads = [add_accumulative_gradients_op(
            self.shared_vars + [sparse_representations[i]],
            self.accum_grads,
            loss,
            i)
            for i, loss in enumerate(self.losses)]

        self.apply_gradients = self.optimizer.apply_gradients(
            zip(self.accum_grads, net_vars))
        self.reset_accum_grads = reset_accumulative_gradients_op(net_vars, self.accum_grads, 1)

Instances