我以分布式模式运行代码，并且我的代码在异步模式下运行良好；但是代码以同步模式运行不成功。
opt = tf.train.MomentumOptimizer(learning_rate=lr_placeholder, momentum=0.9) opt=tf.train.SyncReplicasOptimizer(opt,replicas_to_aggregate=len(worker_hosts),total_num_replicas=len(worker_hosts),use_locking=True)
train_op = opt.minimize(full_loss, global_step=global_step)
val_op = validation_op(validation_step, vali_top1_error, vali_loss)
sync_replicas_hook = opt.make_session_run_hook(True)
init=tf.global_variables_initializer()
with training.MonitoredTrainingSession(master=server.target, is_chief=True,hooks=[sync_replicas_hook]) as sess:
追溯（最近的最新通话）：文件"/usr/local/python3/lib/python3.6/site-packages/tensorflow/python/client/session.py"，第1行292，在_DO_CALL返回fn（*args）文件"/usr/local/python3/lib/python3.6/site-packages/tensorflow/python/client/session.py"，第1行277，在_run_fn选项中，feed_dict，fetch_list，target_list，run_metadata）文件"/usr/local/python3/lib/python3.6/site-packages/tensorflow/python/client/session.py"，第1行367，在_call_tf_sessionrun run_metadata中tensorflow.python.framework.errors_impl.invalidargumenterror：nodedef缺少attr'readuction_type'从OP句柄：Ref（String）;attr = dtype：type，washe = [dt_float，dt_double，dt_int32，dt_uint8，dt_int16，...，dt_uint16，dt_complex128，dt_half，dt_uint32，dt_uint32，dt_uint64];attr = shape：shape;attr = container：string，default =";attr = shared_name：string，default =";attr = reduction_type：string，default =" mean"，washe = ["平均"，" sum"];is_stateful = true>;nodedef：{{node sync_replicas/conditdatoral_accumulator}}} = porysticationalAccumulator_class = [" loc：@sync_replicas/setglobalstep"]，container ="，dtype = dttype = dt_float，shape = [3,3,3,3,3,3,16]，/sharde/name vcretCons：0/Grad_accum"，_device ="/job：ps/epplica：0/任务：0/设备：CPU：0"在处理上述例外时，发生了另一个例外：
tensorflow.python.framework.errors_impl.invalidargumenterror：nodedef缺少attr'redauction_type'从OP句柄：Ref（String）;attr = dtype：type，washe = [dt_float，dt_double，dt_int32，dt_uint8，dt_int16，...，dt_uint16，dt_complex128，dt_half，dt_uint32，dt_uint32，dt_uint64];attr = shape：shape;attr = container：string，default =";attr = shared_name：string，default =";attr = reduction_type：string，default =" mean"，washe = ["平均"，" sum"];is_stateful = true>;nodedef：{{node sync_replicas/conditdatoral_accumulator}}} = porysticationalAccumulator_class = [" loc：@sync_replicas/setglobalstep"]，container ="，dtype = dttype = dt_float，shape = [3,3,3,3,3,3,16]，/sharde/name vcretCons：0/Grad_accum"，_device ="/job：ps/epplica：0/任务：0/设备：CPU：0"