Missing TensorBoard support

There is very little mention of TensorBoard in the TLT 3.0 documentation. A search turns up two instances, in the config files of Facial Landmarks Estimation and Gaze Detection. I’ve noticed that the EfficientNet classification model outputs tfevents files to the location set by --results_dir. I am currently using FasterRCNN, but don’t see a similar output.

Typically I prefer to evaluate models together in a visualization like the one shown below. Is TensorBoard supported by TLT for this purpose? Does TLT support a different method to visualize the loss score over time, and to compare the performance of trained models?



From your reply in the linked post, I understand that TensorBoard is not supported within the TLT container, but tfevents files are still generated by DetectNet. Is it possible to generate tfevents from other models in TLT 3.0?

For other object detection models, the tfevents is not available yet.
One tip, you can save the training log and filter the loss which printed in the log.