Wrong result in TensorRT, but it seems something is working correctly

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

Description

I am working on the tensorRT conversion for the onnx model of semantic segmentation on NVIDIA Xavier AGX dev kit.

I got the same result with the onnx model (not exact same with the pytorch model, but almost similar).

However, the result of the tensorRT model is weird. The output should place within 0-27, because it’s the output of argmax and there are 28 labels for segments. The result of the tensorRT model is almost 0. I use np.unique for the outputs to compare the values in ‘labels of segment’.

input image

image
image858×483 39.8 KB

output of the onnx model

image
image773×389 2.91 KB

labels of segment: [ 0 1 3 4 6 9 10 11 15 16 19 22]

output of the tensorRT model
labels of segment: [0.0e+00 1.4e-45 4.2e-45 5.6e-45 8.4e-45 1.3e-44 1.4e-44 1.5e-44 2.1e-44
2.2e-44 2.7e-44 3.1e-44]

Very interestingly, if I multiply the output of tensorRT model by 5e+45 and save as png. I could see this image.

image
image773×383 2.89 KB

What’s wrong with this? Please help me.

Environment

TensorRT Version: 8.0.1.6
GPU Type: tegra194
Nvidia Driver Version: NVIDIA Jetson AGX Xavier 16GB, Jetpack 4.6 [L44 32.6.1]
CUDA Version: 10.2.300
CUDNN Version: 8.2.1.32
Operating System + Version: Ubuntu 18.04
Python Version (if applicable): 3.6.9
TensorFlow Version (if applicable):
PyTorch Version (if applicable): 1.9
Baremetal or Container (if container which image + tag):

Relevant Files

You can download the model and scripts here

Steps To Reproduce

  • run the onnx model
    python3 run_onnx.py

  • convert the onnx model to the tensorrt model
    /usr/src/tensorrt/bin/trtexec --onnx=model.onnx --saveEngine=model.trt --explicitBatch --fp16

  • run the trt model
    python3 run_trt.py

#2

Hi,
Request you to share the ONNX model and the script if not shared already so that we can assist you better.
Alongside you can try few things:

  1. validating your model with the below snippet

check_model.py

import sys
import onnx
filename = yourONNXmodel
model = onnx.load(filename)
onnx.checker.check_model(model).
2) Try running your model with trtexec command.

In case you are still facing issue, request you to share the trtexec “”–verbose"" log for further debugging
Thanks!

#5

I already shared the onnx model in the relevant files. You can download the model and scripts with the link.

#6

Could you please grant access to the issue repro.

#7

I gave you the access. Please try again.

#8

Hi,

We could reproduce the same behavior.
Please allow us some time to work on this.

Thank you.

#9

Good to hear you can reproduce the issue. Look forward to the update. Thank you.

#10

Hi,

The problem here is in the script - it allocated a host buffer of type np.float32, rather than the datatype actually used in the engine output. The following patch fixes the problem for us.

diff --git a/run_trt.py b/run_trt.py
index add0030..5c645c0 100644
--- a/run_trt.py
+++ b/run_trt.py
@@ -23,7 +23,7 @@ with open('model.trt', 'rb') as f:
 # create buffer
 for binding in engine:
     size = trt.volume(engine.get_binding_shape(binding)) * engine.max_batch_size
-    host_mem = cuda.pagelocked_empty(shape=[size],dtype=np.float32)
+    host_mem = cuda.pagelocked_empty(shape=[size],dtype=trt.nptype(engine.get_binding_dtype(binding)))
     cuda_mem = cuda.mem_alloc(host_mem.nbytes)

Thank you.

#11

Thank you for the reply. I resolved the issue. For the output, the type was supposed to be np.int32. I appreciated it.

1 Like