无法在tf.distribute.MirroredStrategy中使用嵌入层

Question

我试图在tensorflow版本2.4.1上使用嵌入层并行化一个模型。但是它给我带来了以下错误：

InvalidArgumentError: Cannot assign a device for operation sequential/emb_layer/embedding_lookup/ReadVariableOp: Could not satisfy explicit device specification '' because the node {{colocation_node sequential/emb_layer/embedding_lookup/ReadVariableOp}} was colocated with a group of nodes that required incompatible device '/job:localhost/replica:0/task:0/device:GPU:0'. All available devices [/job:localhost/replica:0/task:0/device:CPU:0, /job:localhost/replica:0/task:0/device:XLA_CPU:0, /job:localhost/replica:0/task:0/device:XLA_GPU:0, /job:localhost/replica:0/task:0/device:GPU:0]. 
Colocation Debug Info:
Colocation group had the following types and supported devices: 
Root Member(assigned_device_name_index_=2 requested_device_name_='/job:localhost/replica:0/task:0/device:GPU:0' assigned_device_name_='/job:localhost/replica:0/task:0/device:GPU:0' resource_device_name_='/job:localhost/replica:0/task:0/device:GPU:0' supported_device_types_=[CPU] possible_devices_=[]
GatherV2: GPU CPU XLA_CPU XLA_GPU 
Cast: GPU CPU XLA_CPU XLA_GPU 
Const: GPU CPU XLA_CPU XLA_GPU 
ResourceSparseApplyAdagradV2: CPU 
_Arg: GPU CPU XLA_CPU XLA_GPU 
ReadVariableOp: GPU CPU XLA_CPU XLA_GPU 

Colocation members, user-requested devices, and framework assigned devices, if any:
  sequential_emb_layer_embedding_lookup_readvariableop_resource (_Arg)  framework assigned device=/job:localhost/replica:0/task:0/device:GPU:0
  adagrad_adagrad_update_update_0_resourcesparseapplyadagradv2_accum (_Arg)  framework assigned device=/job:localhost/replica:0/task:0/device:GPU:0
  sequential/emb_layer/embedding_lookup/ReadVariableOp (ReadVariableOp) 
  sequential/emb_layer/embedding_lookup/axis (Const) 
  sequential/emb_layer/embedding_lookup (GatherV2) 
  gradient_tape/sequential/emb_layer/embedding_lookup/Shape (Const) 
  gradient_tape/sequential/emb_layer/embedding_lookup/Cast (Cast) 
  Adagrad/Adagrad/update/update_0/ResourceSparseApplyAdagradV2 (ResourceSparseApplyAdagradV2) /job:localhost/replica:0/task:0/device:GPU:0

     [[{{node sequential/emb_layer/embedding_lookup/ReadVariableOp}}]] [Op:__inference_train_function_631]

将模型简化为基本模型，使其可复制：

import tensorflow as tf
central_storage_strategy = tf.distribute.MirroredStrategy()
with central_storage_strategy.scope():
  user_model = tf.keras.Sequential([
       tf.keras.layers.Embedding(10, 2, name = "emb_layer")
     ])
user_model.compile(optimizer=tf.keras.optimizers.Adagrad(0.1), loss="mse")
user_model.fit([1],[[1,2]], epochs=3) 

任何帮助都将不胜感激。谢谢！

Answer 1

所以我终于想出了问题，如果有人在找答案的话。

到目前为止，Tensorflow还没有完全实现Adagrad优化器的GPU。ResourceSparseApplyAdagradV2操作在GPU上产生误差，这是嵌入层的一个重要组成部分。因此，它不能用于数据并行策略的嵌入层。使用Adam或Adam可以很好地工作。