NetworkRT doesn't support 2D input and output layers #175

New Issue

2020-12-15T23:49:01+01:00

nightduck commented

2020-12-15 23:49:01 +01:00

(Migrated from github.com)

The layers used in the dataDim_t struct is naturally 5 dimensional. NetworkRT reduces them to 3 by filling dimensions in with a 1. Eg 1x224x224x3x1. However, calling getBindingDimensions on a layer will return a struct that pads it with zeros, due to a quirk of a TensorRT call. dataDim_t has a method that will return the total number of elements in it, but if a dimension is zero, this total is always zero. So all 2D layers look empty.

I'm submitted a pull request so the dataDim_t struct can be initialized in a way to correct these zeros into ones.

The layers used in the [dataDim_t](https://github.com/ceccocats/tkDNN/blob/b8855b9599e52a51b371e99255063cd6f00fecd7/include/tkDNN/Network.h#L17) struct is naturally 5 dimensional. NetworkRT reduces them to 3 by filling dimensions in with a 1. Eg 1x224x224x3x1. However, calling getBindingDimensions on a layer will return a struct that pads it with zeros, due to a quirk of a TensorRT call. dataDim_t has a method that will return the total number of elements in it, but if a dimension is zero, this total is always zero. So all 2D layers look empty. I'm submitted a pull request so the dataDim_t struct can be initialized in a way to correct these zeros into ones.

mive93 commented

2021-01-04 16:12:35 +01:00

(Migrated from github.com)

Thank you, we'll look into that in the next days

This repo is archived. You cannot comment on issues.

1 Participants

Due Date

No due date set.

Dependencies

No dependencies set.

Reference: mmr/tkDNN#175