4 ms·
I wonder if and if not why not ML uses this to speed up training
by nwoli 3y ago
I wonder if and if not why not ML uses this to speed up training
- emef 3y agoThe model weights (the thing being updated by the training process) stay loaded in gpu memory during training (the slow part). This could be useful to serialize the model weights to disk when checkpointing or completed, but it's a drop in the bucket compared to the rest of the time spent training.
- nwoli 3y agoI meant it more for the image data