Current behavior
WinMLRunner rejects 5D tensors, with the message "Input feature input shape is too large. GPU path only accepts tensor dimensions <= 4 : 5", even though DirectML supports 5D tensors (and will likely support more than that too).
https://github.com/microsoft/Windows-Machine-Learning/blob/master/Tools/WinMLRunner/src/Run.cpp#L296
Expected behavior
WinMLRunner should just let any dimension count through - let the API decide if it is supported or not.
Minimal reproduction of the problem with instructions
Any model with 5D bindings.
WinMLRunner.exe -gpu -model unet3d.channels_first.keras2onnx.onnx
(source is models.zip from here: #215)
Environment
Windows Build Number: 18363.900
App min and target version: Current master, commit 8ca3e52 on 2019 Oct 15
OS Version (Server, IoT Core, Desktop, etc): Desktop
Graphics Driver version:
DxDiag:
WinMLTools specific:
- Source training framework: (e.g. CoreML, Scikit-learn, …)
- WinMLTools version
Visual Studio
Current behavior
WinMLRunner rejects 5D tensors, with the message "Input feature input shape is too large. GPU path only accepts tensor dimensions <= 4 : 5", even though DirectML supports 5D tensors (and will likely support more than that too).
https://github.com/microsoft/Windows-Machine-Learning/blob/master/Tools/WinMLRunner/src/Run.cpp#L296
Expected behavior
WinMLRunner should just let any dimension count through - let the API decide if it is supported or not.
Minimal reproduction of the problem with instructions
Any model with 5D bindings.
WinMLRunner.exe -gpu -model unet3d.channels_first.keras2onnx.onnx(source is models.zip from here: #215)
Environment
Windows Build Number: 18363.900
App min and target version: Current master, commit 8ca3e52 on 2019 Oct 15
OS Version (Server, IoT Core, Desktop, etc): Desktop
Graphics Driver version:
DxDiag:
WinMLTools specific:
- Source training framework: (e.g. CoreML, Scikit-learn, …)
- WinMLTools version
Visual Studio