Can I export a gpu-trained model to a cpu-only machine for serving/evaluation?

[3rduncle]

I trained a model on a GPU machine.
When l load such model on a cpu-only machine, I get the error below.

Thank you for any suggestions!

[jmchen-g]

@vrv Could you take a look at this please? Do we have such support now? Thanks.

[vrv]

Yup, you can pass tf.ConfigProto(allow_soft_placement=True) to the tf.Session if you really want to ignore device placement directives in the graph. I'm not a huge fan of using this option (I'd rather the graph be rewritten to strip out the device fields explicitly), but that should work.

[vrv]

Also, in the future, StackOverflow is the right place to ask these questions.