This release brings support for PyTorch 1.12, Torchvision 0.13, Apple Silicon, Metal acceleration, Fedora, new features and fix many issues (see full changelog below for details).

If your mac supports it, it is strongly recommended to make Metal Acceleration the default device for training and inference (Menu > Settings). If you can’t see it, click Reset Python and relaunch Torchstudio to reinstall a fresh Python environment.

Download TorchStudio 0.9.8

Full Changelog

  • add support for PyTorch 1.12 and future versions of PyTorch
  • add support for Apple Silicon
  • add support for Metal acceleration (new default device for inference and training, if not change it in Menu > Settings)
  • add RPM installer for Fedora 32 or higher and other RPM-based Linux distributions
  • add support for the new optical_flow torchvision model category
  • add Early Stopping feature (in the Model Tab, Hyperparameters)
  • add Restore Best Epoch feature (in the Model Tab, Hyperparameters)
  • add an Add Packages… button in Menu > Settings > Local Environment to add conda package to the local python environment
  • add an Open Terminal… button in Menu > Settings > Local Enviroment to open a python terminal using the local environment
  • python install and basic packages install now much faster and stable
  • local python environment details are now displayed in Menu > Settings > Local Environment
  • signals renderer module improved with more parameters (scaling)
  • all parameters now have tooltips
  • fix unstable ssh connections with some servers
  • fix TorchStudio not launching with some Linux environments
  • fix some urls not being displayed properly for some models and datasets documentation
  • fix rendering issues when displaying signals with many channels
  • fix shuffling setting not resetting when reloading the dataset
  • fix training state sometime not properly updated when training is stopped
  • TorchStudio scripts are now under MIT license

Subscribe to TorchStudio channels for live updates: