Shay Aharon

ddp logic changed

1

When gpu_mode is AUTO- gpu_mode is set to DDP if multiple devices are available.

Feature/sg 757 resume for spots

4

Added support for resuming from a remote ckpt stored by the SG logger during training (meaning when sg_logger_params.save_checkpoints_remot=True). For the base sg logger this is still problematic, as we dont...

Feature/sg 850 support torch 2.0

1

- Fixed some failing tests in build after moving it to run on 3.8. - Moved build to run on 3.8 (minimum python version requirement for torch 2.0) - Fixed...

Bug/sg 861 decouple qat from train from config

1

This PR decouples QAT/PTQ from train_from_config. The goal is to let users launch PTQ/QAT using Python instead of CLI + configs, while using as much defaults as possible. Note that...

- Made the download of the weights optional (and set to true for backward compatibility), Weights are downlaoded only whwn preatrined_weights-"platform/XXX" (thisi is what @ofrimasad and I agreed upon). -...

Also fixes earlier bug that always loaded the Trainer.best_metric from the "acc" entry that actually holds latest metric_to_watch rather then best.