How do I build an ONNX file that does the image scaling and format conversion? about windows-machine-learning HOT 4 CLOSED

Thanks for posting.

1. The preprocessing model used in the sample was created using LearningModelBuilder which is part of the WinML Experimental API. An example of using LearningModelBuilder to create a model can be found here. Alternatively, you could also use the ONNX python API.

2. Writing a compute shader is possible, but will not be optimized with other shaders in the dml graph. There may be performance gains but it's hard to say.

3. WinML uses DML as a backend. The sample showcases how to use ORT with the DML EP, but ORT can be use with other EPs (CUDA, TensorRT, etc.).

from windows-machine-learning.

Thanks for posting.

1. The preprocessing model used in the sample was created using LearningModelBuilder which is part of the WinML Experimental API. An example of using LearningModelBuilder to create a model can be found here. Alternatively, you could also use the ONNX python API.
2. Writing a compute shader is possible, but will not be optimized with other shaders in the dml graph. There may be performance gains but it's hard to say.
3. WinML uses DML as a backend. The sample showcases how to use ORT with the DML EP, but ORT can be use with other EPs (CUDA, TensorRT, etc.).

Thanks for your reply @nums11!

Would you please confirm for the third bullet point above that if we want to use any providers other than DML, we can't use WinML, only a level down using ORT directly is the way to go? I am asking because for your sample code, the ORT API specifically mentions only DML is supported at this moment. Thanks!

from windows-machine-learning.

Yes, if you want to use a provider other than DML you cannot use WinML and you can use ORT. Here is an example from the ORT docs of using ORT with CUDA as a backend.

from windows-machine-learning.

Yes, if you want to use a provider other than DML you cannot use WinML and you can use ORT. Here is an example from the ORT docs of using ORT with CUDA as a backend.

Thanks for your reply @nums11!

from windows-machine-learning.