Thousands of Datasets
Browse and load thousands of datasets from TorchVision, TorchAudio or Hugging Face Hub (upcoming).
Thousands of Models
Browse and train thousands of models from TorchVision, TorchAudio, PyTorch Hub (upcoming) or Hugging Face Hub (upcoming).
Local & Remote Servers
Store datasets and train models locally or remotely, on premise or in the cloud. TorchStudio is compatible with AWS, Azure, Google Cloud and many more.
Display datasets samples on demand, no matter the number of inputs/outputs and the tensor formats.
Format & Analyze Datasets
Redefine samples usage, balancing and shuffling on the fly. Produce statistical reports about the entire dataset.
Show model graph, revealing how internal nodes are connected and how tensor size change between nodes.
Prepare & Monitor Trainings
Set parameters and hyperparameters, monitor your loss and metric curves and output predictions in real time.
The Dashboard shows all your trainings at once and how their performance compare to each other, helping you quickly identify the best model.
Edit or add your own modules. Every components are customizable: dataset, model, analyzer, renderer, loss, metric, optimizer, learning rate scheduler.
Open Project Format
Reuse TorchStudio projects externally : they are simple zip files containing all code, data and parameters in Python and PyTorch compatible formats.
Open Source Code
All core routines under the hood are freely accessible and editable Python files. Tweak to your own needs, or connect TorchStudio to other systems.