Hai Xuan Pham

Interesting! Hope you succeed in building CNTK with CUDA 11 and maybe newer Python version too.

@delzac RTX 2000 series did not support CUDA 9.x - worked with 10.0. So my guess RTX 3000 will require more recent CUDA version. I still hold my (little) hope...

Don't buy new if not required! Find a couple of 2nd hand Tesla P100 or Quadro P6000 cards with 16GB RAM each.

AlexNet is small, thus little footprint. Inception-v3 is a much bigger network, which requires more memory for more hidden feature maps. That's normal behavior really.

Can you elaborate more about your setting? Does every spectrogram have the same dimensions or does it vary from sample to sample?

Out of topic: I would suggest you use Python for training, and C# for inference only. C# wrapper does not expose all primitive functions in C++, and Python API contains...

Gotcha! There's no API reference for C++, have to look up the header. https://github.com/microsoft/CNTK/blob/release/latest/Source/CNTKv2LibraryDll/API/CNTKLibrary.h Line 4624: there is "sequential" argument which indicates that the convolution is to be applied on...

It would be clear if you provided us with a diagram of your model. That way we would know exactly which functions are required and how they are called.

I believe yes, using C++ API where you have direct memory access. Unfortunately the C++ API is not documented like the Python API.

I don't have experience with CUDA in C#, unfortunately. But you may find the way from this [regression example](https://github.com/microsoft/CNTK/blob/release/latest/Examples/TrainingCSharp/Common/LogisticRegression.cs). Look at GenerateValueData() Basically, you pass memory to Value variables, that...