TensorRT4 ConvTranspose layer VERY SLOW inference speed.

I’m experiencing slow inference speed for ConvTranspose Layer (deconv) in TensorRT4, as is shown in the jpg URL.

The network architecture is based on resnet-18 layers, with 32x downsampling and 8x upsampling(three deconv layers), where the three deconv layers take almost 1/2 time of the inference time. It doesn’t seem right when compared to Conv operations, as both layers are both convolution ops.