Whats the different between Deconvolution groups and deconvolutional layers?

The official post https://devblogs.nvidia.com/parallelforall/tensorrt-3-faster-tensorflow-inference/ says that TensortRT 3.0 support Deconvolution. However when I use TensorRT to accelerate my caffe model, it fails with the message “doesnot support Deconvolution groups”. My deconvolutional layer is:

layer {
  name: "upscore"
  type: "Deconvolution"
  bottom: "score_fr"
  top: "upscore"
  param {
    lr_mult: 0.0
  convolution_param {
    num_output: 21
    bias_term: false
    kernel_size: 63
    group: 21
    stride: 32
    weight_filler {
      type: "bilinear"

I am using the TensorRT 2 on Tx2. Anyone, please tell me what’s wrong?


Group convolution is depthwise convolution.
Check this page for more explanation:

Currently, TensorRT only supports standard convolution layer. (Apply the same convolution to different channel)
Group convolution will be supported in our future release but no concrete schedule yet.

Thanks and sorry for the inconvenience.

Thank you for your answer! But I do need to do some semantic segmentation on my tx2. Is there any previous works or tutorial you can recommend?


Group is kind of model design technique to lower the layers/parameters amount.
Try to add more layer with some residual path.
It should be able to have similar capacity of the group-based model.