【Tensorflow】用Keras,多GPU训练含有BN层的网络报错

Jul 21, 2019#Tensorflow184

AI-generated summary

The article discusses an error that occurs when using multi_gpu_model in Keras for training a network with BN layers in Tensorflow. The error is caused by the source code and can be resolved by modifying a line in the device.py file. The article provides instructions on how to do this and resume normal training.

如果在 keras 中使用 multi_gpu_model，报错内容：

AttributeError: 'DeviceSpec' object has no attribute 'split'

如果出现这个错误，多半都是源码造成的。

请查看 tensorflow/python/framework/device.py 这个文件，

查找下面这一行：

current_device = DeviceSpec.from_string(node_def.device or "")

将上面这一行注释掉，改成下面这一行：

current_device = node_def.device if isinstance(node_def.device, DeviceSpec) else DeviceSpec.from_string(node_def.device or "")

保存退出，就能正常训练了。