I’d like to take a yolov4 model that has been deployed on a live camera system, then train it more using say 1000 new images (+ augmented images) for better local performance. The training would ideally occur on an NX or AGX ideally during a 12-hour “down-time” period.
As I cant use TLT due to using a custom model, any recommendations for resources that would tell me how to use transfer learning to improve the model on the edge?
To do transferring learning, do I have to have access to the full original dataset? or can i just continue training on the pre-trained model?
Usually, we recommend users to use PyTorch to retrain the model on Jetson.
You can find some ideas below:
The dataset issue depends on if you need the original class output.
If yes, please train the model with the original dataset (at least a subset of it).
If all the output classes are custom, you can just retrain it with the new dataset directly.