avg_pool2d count_include_pad flag not set correctly in TensorRT wrapper #32
Description
In tensorrt_executor.cc, the AddPooling function does not call nvinfer1::IPoolingLayer::setAverageCountExcludesPadding, which is necessary to set count_include_pad in avg_pool2d. Thus, AverageCountExcludesPadding is true by default. For models such as Inception V3 which using average pooling (for multiple frameworks, including MXNet and PyTorch), AverageCountExcludesPadding should be set to false via nvinfer1::IPoolingLayer::setAverageCountExcludesPadding since count_include_pad is set to true for Inception V3. The result is that Inception V3 produces incorrect outputs for multiple frameworks. @reminisce