Disabling shuffling parse uff

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

Hi,

I added placeholder nodes in graph surgeon while parsing to uff, their input shapes need to be taken as is in TRT. However, TRT inserts shuffling operations right after them. How do I disable the adding of shuffling operations? I register these nodes as inputs and tried every combination of inputFormat from NCHW and NHWC. I can’t make the automatically added shuffle operations go away.

Fan

2

I should add that I have 5 inputs, one of them is a natural input from my Tensorflow model, the 4 others I added through graph surgeon. Registering the first one, the natural one works as expected: when I specify my input format is already in CHW, TRT doesn’t add a shuffle layer. For the 4 other ones, no matter what I specify as input format kNCHW or kNHWC, it adds a shuffle layer after it and the shuffle operation is CHW->HWC ie it takes the first dimension and puts it at the end.

3

Did you find any solution to this? I’m having a similar problem with shuffle layers being added

4

I have the same problem too. Does anyone solve it ?