Cannot run model tested on GPU on TX2 - garbage returned.

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

The story:

  1. I have a Jetson TX2 to run a model on (to detect things in images)
  2. I'm given a model, which is converted to UFF.
  3. I use bin2c.py to produce sorta C code to include in my sampleUffMNIST.cpp (heavely patched sample file)
  4. I use sample build setup to produce the binary out of that sampleUffMNIST.cpp (right on that Jetson)
  5. I run that binary and it waits on AF_UNIX socket...
  6. I run the feed.py which grabs the images, feeds it to sampleUffMNIST.cpp over that local socket and
  7. feed.py fetches output from sampleUffMNIST.cpp (which is expected to be a matrix of "probabilities" for a pixel to belong to an object)
  8. the matrix is then applied to the source image and we get the image with everything masked out except of objects found.

I have a mockup in Python3 (just rewritten the same original sampleUffMNIST.cpp) on a PC with GPU and it works fine!
But I have no TensorRT Python binding on Jetson.
Hence, I have to code in C++ (which is not my favorite by any mean) to run the engine.

This doesn’t work. I get garbage on the output.

Here we are.

The details:

  • Jetson TX2 running tegra-ubuntu 4.4.38-tegra
  • tensorrt 4.0.2.0-1+cuda9.0 package on TX2
  • tensorrt 5.0.0.10-1+cuda9.0 package on PC/GPU
  • the source code for mentioned files is available here

Any help would be appreciated!

#2

Few words more: GPU is 1080 (not “Ti”), both FP16/FP32 modes attempted, 1080 prefers FP32, TX2 prefers FP16…

#3

Even more: if I build the same patched sampleUffMNIST.cpp on PC/GPU and run the routine described above… it works!

And fails on Jetson.

Why??

#4

Let’s try the low hanging fruit first. Can you try updating to jetpack 4.1 which contains TRT5 (matching your desktop configuration) and see if results improve?

#5

Well, I’ll try to.

But it’s said that 4.1 is for Xavier only…

#6

Yes, it is.

I mean, no way (“Next” button fails here): https://i.imgur.com/wWja5v4.png (I have no right to use img tag)

This fruit hangs high enough.
wWja5v4.png

#7

Attempt to forcibly upgrade (apt-get install --only-upgrade tensorrt) fails too:

Reading state information... Done
tensorrt is already the newest version (4.0.2.0-1+cuda9.0).
#8

Few more details: an old version model (for 256x256 source images) looks ok for both platforms.
The new one (for 512x512 source image) is not.

I don’t understand how to map this fact to possible reason/fix…