This release brings lots of performance, stability and UI improvements.

It also adds a new GenericLoader dataset, which can handle most common image, audio and numpy tensors datasets formats, for either classification, segmentation or regression.

Read the new Loading Custom Datasets tutorial to learn how to use it.

Download TorchStudio 0.9.14

Full Changelog

  • new GenericLoader dataset, which can load several kinds of image, audio and numpy tensors datasets for either classification, segmentation or regression
  • TorchStudio projects now use much less RAM
  • transfers with local and remote servers is now much faster
  • new weights and state transfers from local and remote servers is now asynchronous
  • only best weights and state are transferred to speed up training and optimize memory use
  • dataset can now be cached on local and remote trainign server to speed up the start of new trainings
  • environment installer now compatible with the newest conda installers
  • Patience value can now be defined (in number of epochs) when Early Stopping is checked
  • channel colors can now be defined for Bitmap, Spectrogram and Volume renderers
  • default threshold value for accuracy metric set to 0.01
  • new best weights indicator in the loss and metric plots (both in Model tabs and Dashboard panel)
  • loss indicator now plotted on a square root scale for easier readibility
  • model color indicator added to the parameters plot (in the Dashboard panel)
  • tool tips improved
  • fix an issue when loading remote dataset
  • fix an issue where inference could stop working
  • fix an issue where tensor display and analysis could stop working
  • fix an issue where tcp decoding could break
  • fix project opening when a dataset is loaded from a remote server
  • fix metrics not properly displaying epochs beyond 100

Subscribe to TorchStudio channels for live updates: