Reports from the Heart of the Machine
Editor: Chen Ping
After months of closed testing, TorchStudio is now finally in the public beta phase.
The emergence of compilers can be said to have brought great convenience to the majority of researchers. Simply put, a compiler is a program that translates one language (usually a high-level language) into another (usually a low-level language). The main workflows of a modern compiler include source code preprocessor compiler, object code, linker, executables.
The compilers we often hear about or use are Visual Studio, Dev C++, Visual C++ 6.0, GCC, Clang, etc. It is no exaggeration to say that it is the existence of compilers that allows human languages to be transformed into 0s and 1s that computers can understand. Last year's compiler book "The Dragon Book" authors Alfred Vaino Aho and Jeffrey David Ullman also won the 2020 Turing Award.
Today, Robin Lobel, an open source software enthusiast and creator of the audio editing software Steinberg SpectraLayers, released a free and open source PyTorch IDE called TorchStudio, which is now in public beta version 0.9.1. The IDE greatly simplifies the research and training of PyTorch and its ecosystem, and most tasks can be performed intuitively with just a few clicks.
TorchStudio now supports Windows 10 or later, macOS 10.14 or later, Ubuntu 20.04 or later. TorchStudio provides sample tutorials, novices can learn according to the tutorial, it is worth mentioning that torchStudio official website also provides video tutorials intimately, users do not have to worry about not using.
TorchStudio Address: https://torchstudio.ai/getstarted/
What are the features of this IDE? According to the official website, its characteristics can be summarized into three: linking thousands of data sets, thousands of models, visualizing, flexible and scalable.
Datasets: Browse and load thousands of datasets from TorchVision, TorchAudio, or Hugging Face Hub (coming soon), including CIFAR10, CIFAR100, Caltech101, CelebA, and more.
Models: Browse and train thousands of models from TorchVision, TorchAudio, PyTorch Hub (coming soon), or Hugging Face Hub (coming soon), including the familiar AlexNet, DenseNet, GoogLeNet, and more.
Users can also store datasets and train models locally or remotely, locally, or in the cloud. In addition, TorchStudio is compatible with AWS, Azure, Google Cloud, etc.
TorchStudio can also visualize and display samples of datasets on demand, regardless of the number and tensor format of user inputs/outputs:
TorchStudio can also generate statistical reports on the entire data set:
Visualize the model to reveal how internal nodes are connected and how tensor sizes vary between nodes:
Users can also set network parameters and hyperparameters, monitor loss and metric curves in real time, and output predictions:
In addition, TorchStudio compares training models to help users quickly determine the best model:
With TorchStudio, users can also edit or add modules themselves, each of which is customizable, including: datasets, models, analyzers, renderers, losses, metrics, optimizers, and more:
In the face of such an excellent compiler, netizens can't help but sigh: "Very beautiful work!" I won't use any other IDEs except PyCharm, but I'll give this TorchStudio a try!"
If you use PyTorch for research, you can try this full-featured compiler.