keras save_model时，错误提示：ValueError: Unable to create dataset (name already exists)

keras 自定义embed layer,如下：

class ConcatClassTokenAddPositionEmbedding(layers.Layer):
    def __init__(self, embed_dims, token_num, name='None'):
        super(ConcatClassTokenAddPositionEmbedding, self).__init__(name=name)
        self.embed_dims = embed_dims
        self.token_num = token_num

    def build(self, input_shape):
        self.class_token = self.add_weight(
                                           shape=[1, 1, self.embed_dims],
                                    initializer=keras.initializers.initializers_v2.Zeros(),
                                           trainable=True,
                                           dtype=tf.float32)
        self.w = self.add_weight(
                                 shape=[1, input_shape[1] + 1, input_shape[2]],
                             initializer=keras.initializers.initializers_v2.HeUniform(),
                                 trainable=True)

在使用model.save('model.h5')保存模型时，崩溃，错误提示如下：

dset_id = h5d.create(parent.id, name, tid, sid, dcpl=dcpl)
  File "h5py\_objects.pyx", line 54, in h5py._objects.with_phil.wrapper
  File "h5py\_objects.pyx", line 55, in h5py._objects.with_phil.wrapper
  File "h5py\h5d.pyx", line 87, in h5py.h5d.create
ValueError: Unable to create dataset (name already exists)

问题原因：模型中有变量重名(name already exists)。具体的是自定义层中定义了两个可训练变量，即self.class_token和 self.w。并且在定义时，self.class_token = self.add_weight(）没有指定name参数，name参数为空时，使用默认名字：Variable.因为self.class_token和slef.w都没有指定name，所以二者名字都是Variable，导致重名。

解决方法：

add_weight中设置不同的name参数即可，即：

 self.class_token = self.add_weight(name='cls_token',
                                           shape=[1, 1, self.embed_dims],
                                           initializer=keras.initializers.initializers_v2.Zeros(),
                                           trainable=True,
                                           dtype=tf.float32)
        print('class_token' + str(self.class_token.shape))
        self.w = self.add_weight(name='pos_embedding',
                                 shape=[1, input_shape[1] + 1, input_shape[2]],
                                 initializer=keras.initializers.initializers_v2.HeUniform(),
                                 trainable=True)

问题解决！