search

microsoft / Windows-Machine-Learning

Samples and Tools for Windows ML.
https://docs.microsoft.com/en-us/windows/ai/
MIT License
1.03k stars 446 forks source link

WinMLRunner.exe. Evaluating ..[FAILED] #379

Open yahuuu opened 3 years ago

yahuuu commented 3 years ago

I convert the pt model to the onnx model and verify that the onnx model has good output accuracy. But my onnx model cannot evaluateing with GPU. I only saved the weights in the model.

error like: Binding (device = CPU, iteration = 9, inputBinding = CPU, inputDataType = Tensor, deviceCreationLocation = WinML)...[SUCCESS] Evaluating (device = CPU, iteration = 9, inputBinding = CPU, inputDataType = Tensor, deviceCreationLocation = WinML)...[SUCCESS] ... etc ... Binding (device = GPU, iteration = 1, inputBinding = CPU, inputDataType = Tensor, deviceCreationLocation = WinML)...[SUCCESS] [FAILED] Evaluating (device = GPU, iteration = 1, inputBinding = CPU, inputDataType = Tensor, deviceCreationLocation = WinML)...[FAILED]

command: WinMLRunner.exe -model D:\projects\PYTHON\u2net_matting_verify\u2net_matting_verify\utils\test.onnx -iterations 9 -perf -GPU

environment: WinMLRunner.exe v1.0.1.0 x64 executable win10 20H2 19042.685, Intel(R) Core(TM) i7-8700 CPU, RAM 32.0 GB x64 GPU: Intel UHD Graphics 630 GPU Driver Version: 25.20.100.6473

ryanlai2 commented 3 years ago

Hi @yahuuu, thanks for reaching out. May I ask if you can share your model with us for reproduction purposes?

ryanlai2 commented 3 years ago

Can you also share your ONNX operator set version of the model? Are you using inbox or NuGet version of WinML? If using NuGet, which version are you using?

yahuuu commented 3 years ago

Hi @ryanlai2, This is the pytorch model I'm using.
This is a pre-trained model released by the MODNet community. I referenced the official ONNX conversion code, pytorch -> onnx in Ubuntu16.04, run WinMLRunner.exe in Win10.

ubuntu version : ) pytorch==1.3.0 (gpu version) torchvision==0.4.1 WinMLRunner.exe v1.0.1.0 && v1.2.2 x64 CUDA Version: 10.1 NVIDIA-SMI 418.39 Tesla K40m

my code : ) “”“ torch.onnx.export(modnet, example, export_onnx_file, opset_version = 10, export_params=True, do_constant_folding = True, input_names = ["input"], output_names = ["output"],
dynamic_axes = {"input": {0: "batch_size"}, "output": {0: "batch_size"} }) ”“” If I use pytorch==1.7.1(cpu version), WinMLRunner won't even output the results under the CPU.

yahuuu commented 3 years ago

opset_version = 12, WinMLRunner v1.2.2.zip Waiting for your answer.