Problem inferring conv layer using GIE


I am implementing convolution inference using GIE(TensorRT 1.0). I am inferring feature maps of convolution layer, then convert those into images. I am able to get the results when setHalf2Mode(false) and type is set to DataType::kFLOAT for weights. When setHalf2Mode(true) and type is set to DataType::kHALF for weights, I am getting results(feature map images) at odd index and blank images at even index.

Below is my network implementation.

[Image::float[InputSize]]->addInput(DataType::kFLOAT)->conv1(DataType::kHALF)->activation1(relu, DataType::kHALF)->pool1(DataType::kHALF)->MakeOutput(DataType::kHALF)->[FeatureMaps::float[OutputSize]]

above “” represents simple input data and output result with their types.

What could be going wrong?

attachment of my output.