dynamic_embedding table 插入数据报错

[lwmonster]

System information

• OS Platform and Distribution (e.g., Linux Ubuntu 22.04):
• TensorFlow version and how it was installed (source or binary): tensorflow2.8.3 用 pip 安装的
• TensorFlow-Recommenders-Addons version and how it was installed (source or binary): tfra 0.5.0
• Python version: Python3.9.19
• Is GPU used? (yes/no): NO

Describe the bug

在使用 Dynamic Embedding 的时候，创建了 Variable， 这个 Variable 里会有一些 hashtable 用于存储稀疏特征的 embedding。 现在我想用外部数据更新 hashtable 中的 embedding 向量，但是 执行 table.insert 的时候会报错：

ValueError: Operation name: "hashTableGroup_0_mht_1of1_lookup_table_insert/TFRA>CuckooHashTableInsert"
op: "TFRA>CuckooHashTableInsert"
input: "group_0/hashTableGroup_0/hashTableGroup_0_mht_1of1"
input: "hashTableGroup_0_mht_1of1_lookup_table_insert/keys"
input: "hashTableGroup_0_mht_1of1_lookup_table_insert/values"
device: "/job:ps/replica:0/task:0"
attr {
  key: "Tin"
  value {
    type: DT_INT64
  }
}
attr {
  key: "Tout"
  value {
    type: DT_FLOAT
  }
}
attr {
  key: "_class"
  value {
    list {
      s: "loc:@group_0/hashTableGroup_0/hashTableGroup_0_mht_1of1"
    }
  }
}
 is not an element of this graph.

Code to reproduce the issue

我的代码如下：

def _save_merged_ckpt(self,
                          cluster,
                          server,
                          model_config,
                          valid_data_dirs,
                          cluster_spec,
                          task_type,
                          task_index,
                          merged_variables,
                          merged_hashtable,
                          target_ckpt_dir):

        if task_index != 0:
            print('Worker {} is not chief, skip save merged ckpt'.format(task_index))
            return

        if tf.io.gfile.exists(target_ckpt_dir):
            tf.io.gfile.rmtree(target_ckpt_dir)
        tf.io.gfile.makedirs(target_ckpt_dir)

        #首先重置模型图
        tf.reset_default_graph()
        with tf.Graph().as_default():
            eval_ds = DatasetManager.create_dataset(model_config.dataset_type,
                                                    model_config,
                                                    valid_data_dirs,
                                                    training=False,
                                                    is_eval=True)
            inputs = eval_ds.build(batch_size=1,
                                num_epochs=1,
                                num_worker=model_config.num_shard_of_export,
                                worker_idx=task_index)

            with tf.device(
                tf.train.replica_device_setter(worker_device="/job:worker/task:%d" % task_index,
                                            cluster=cluster)):
                # dataset
                inference_iterator = inputs.make_one_shot_iterator()
                sample = inference_iterator.get_next()
                self.model_fn(sample, eval_ds, training=True)

                global_step_assign = tf.assign_add(self._global_step, 1)
                # 获取模型所有参数的值
                all_vars = tf.trainable_variables()
                # 更新dense参数
                assign_ops = []
                for var in all_vars:
                    print("{}'s shape is {}".format(var.name, var.shape))
                    if var.name not in merged_variables:
                        print('var {} not in merged_variables, SKIP it'.format(var.name))
                        continue
                    assign_ops.append(tf.assign(var, merged_variables[var.name]))

                # 更新hashtable参数
                for table in self._hash_tables:
                    if table.name not in merged_hashtable:
                        print('table {} not in merged_hashtable, SKIP it'.format(table.name))
                        continue
                    keys, values = list(merged_hashtable[table.name].keys()), list(merged_hashtable[table.name].values())
                    values = np.array(values)
                    #values_tensor = tf.convert_to_tensor(values, dtype=tf.float32)
                    assign_ops.append(table.insert(keys, values))

            parallelism_conf = self.get_parallelism_conf(task_type, task_index)
            scaffold = self.get_scaffold()
            start = time.time()

            saver = tf.train.Saver()
            with tf.train.MonitoredTrainingSession(master=server.target,
                                                is_chief=(task_index == 0),
                                                scaffold=scaffold,
                                                checkpoint_dir=target_ckpt_dir,
                                                hooks=[],
                                                save_checkpoint_secs=None,
                                                save_checkpoint_steps=None,
                                                save_summaries_steps=None,
                                                save_summaries_secs=None,
                                                config=parallelism_conf) as sess:
                print('Worker {} restore successful....'.format(task_index))

                _, step = sess.run([assign_ops, global_step_assign])
                print('Assign merged_variables to checkpoint done')

我的assign_ops 是在图里的呀，为什么会报CuckooHashTableInsert 这个 OP 不在图里呢？ 哪位大佬帮忙指点一下呀 PS: 这里的 self._hash_tables 是保存的所有 dynamic_embedding_variable 变量的 内部的 tables

Provide a reproducible test case that is the bare minimum necessary to generate the problem.

Other info / logs

Include any logs or source code that would be helpful to diagnose the problem. If including tracebacks, please include the full traceback. Large logs and files should be attached.

[lwmonster]

@MoFHeka @rhdong @jq HELP Please ~

[lwmonster]

问题解决了，但还是不知道为啥...