Accessing model activations

I have trained a couple of models on my data. Now I want to use access the model activations to apply active learning and various other feature based methods. How does one take a trained TLT model and generate the activations from an arbitrary layer in the model?

Accessing model activations is not supported for end user.

That’s unfortunate because that really degrades the utility of TLT for our purposes. Will this feature be made available in an upcoming version? Also, is it possible to get the bounding box scores from TLT object detection models?

I will sync with internal team for your request.
The bbox scores are available when run inference, for example, running “tlt detectnet_v2 inference”, “tlt ssd inference”, etc.

1 Like

End user can freeze blocks during training. Can you evaluate if it can meet your requirement “accessing model activations”?

Is there documentation for doing so?

You can search “freeze_blocks” in tlt user guide.
For example, in detectnet_v2,
https://docs.nvidia.com/metropolis/TLT/tlt-user-guide/text/object_detection/detectnet_v2.html#model-config

freeze_blocks: This parameter defines which blocks may be frozen from the instantiated feature extractor template, and is different for different feature extractor templates.