Implementing TAO's classification model as SGIE in deepstream-app

Accelerated Computing Intelligent Video Analytics DeepStream SDK
,
1

Please provide complete information as applicable to your setup.

• Hardware Platform (GPU) RTX4070
• DeepStream Version 7.1
• TensorRT Version 10.3.0.26
• Issue Type( questions, new requirements, bugs)
I am implementing SGIE in deepstream-app. SGIE is a classification model with two classes and trained with TAO library.
SGIE has two classes, but the label shows the first class only. The second class is not shown.
Where to update?
My config files are attached.
config_infer_primary_sasco.txt (3.5 KB)
config_infer_secondary_1_sasco.txt (3.0 KB)
sasco_main.txt (5.6 KB)

3

Could you attach posture_labels.txt and the complete DeepStream log?

For one classifier output with two classes, the label file must contain both labels on the same line:

class0;class1

Do not put the two labels on separate lines. DeepStream treats each line as a different output attribute, while labels separated by semicolons are treated as classes of the same output. This can cause class 0 to display while class 1 has no label.

Also, num-detected-classes is only for detector models and can be removed. Please remove parse-classifier-func-name as well unless you provide your own custom-lib-path.

For debugging, temporarily set:

  classifier-async-mode=0
  classifier-threshold=0

If class 1 still never appears, please provide the ONNX output name and dimensions, the TAO training/export configuration, and the preprocessing parameters. The RGB/BGR format, scale and offsets must match those used during training.

4

Thanks for the reply. I’ll update later.

5

posture_labels.txt has two lines. Now it works.