VAR AutoML vs AutoML Edge training time

[melvinclanceydu]

I have been training a video action recognition model with ~400 videos and ~1000 annotations using AutoML successfully, each taking between 1-2 hours. However, I started a training pipeline using AutoML edge (TFLite) on the same data, and so far it has taken over 23 hours (and counting). Can there be such a large disparity in training time between these two methods? 

I'm wondering if I should just wait it out or start another training pipeline - I'm working on a personal project and am wary of the node-hour costs. 

[melvinclanceydu · Accepted solution]

I was able to train on generic/nvidia jetson successfully. Looks like coral training may have had some issues.