Implement Average Pooling by using cuDNN 9.5

AI & Data Science Deep Learning (Training & Inference) cuDNN
1

Hello, my name is Alex. I need to implement Average Pooling on Tensor object in C++ by using cuDNN 9.5. Function cudnnPoolingForward and related descriptor functions are deprecated. Thanks in advance.

3

Here are some potential solutions to implement Average Pooling on a Tensor object in C++ using cuDNN 9.5, considering that cudnnPoolingForward and related descriptor functions are deprecated:

  1. Transition to the PyTorch Quantization Library:

    • Utilize the PyTorch Quantization Library, which includes classes like QuantAvgPool2d, QuantAvgPool3d, and QuantAdaptiveAvgPool1d, suitable for average pooling operations.

  2. Utilize TensorRT for Average Pooling:

    • Explore the TensorRT C++ API for implementing Average Pooling. This allows you to create and configure layers for average pooling without relying on deprecated cuDNN functions.

  3. Explore Alternative Libraries:

    • Consider using other deep learning libraries that provide optimized implementations for Average Pooling operations compatible with cuDNN 9.5 or newer.

  4. Consult Official Documentation and Forums:

    • Refer to the official cuDNN and TensorRT documentation for updates or replacements for the deprecated pooling functions. Engaging with community forums can also provide valuable insights.

  5. Custom Implementation:

    • If necessary, implement a custom Average Pooling operation using lower-level CUDA programming or custom kernel implementations, offering a solution tailored to your specific needs.

For specific guidance and detailed instructions, consult the official documentation of the relevant libraries and explore community resources.