why /DML:0 instead of /GPU:0 ? about tensorflow-directml HOT 4 CLOSED

Yeah, it's a discussion we had and there isn't a perfect solution. A few reasons:

• Some third-party scripts use the presence of a "GPU" device to mean the presence of CUDA and use CUDA-specific operators (e.g. CuDNNLSTM) that don't even have CPU implementations.
• Some internal TF runtime code uses "GPU" to mean CUDA and do CUDA-specific optimizations.
• Using "GPU" would prevent DML from coexisting with other backends. This may be more important when we move to TF2 if things transition to a more modular device backend.

from tensorflow-directml.

(Firstly, thanks for this excellent project!)
This issue is about the cost of migrating form existing TensorFlow project to TensorFlow-DirectML.

Related question: When migrating form existing TF project, is it as easy as changing the device placement from /GPU:0 to /DML:0? Is there any other code which needs to be modified?
And, is there any "Brief Migration Guideline"?

(Currently I don't own a suitable graphic card for DirectML, so sorry that I can't test it by myself. But I'm sure that many developers are also wondering about the migration cost.)

from tensorflow-directml.

DirectML should work in most DirectX12-capable graphics cards. I want to also point out that for an existing script that doesn't explicitly request a /GPU device and that it doesn't have a strong dependency on any explicit cuDNN kernel, and simply opt for the default behavior, TensorFlow-DirectML defaults to /DML. There were some active discussion in the TensorFlow community at one point about overloading /GPU for a different backend in the context of TF2 pluggable device. But I believe the decision was to not mess with what people already use and risk breaking their expectation.

from tensorflow-directml.

Thanks for detailed reply!

from tensorflow-directml.