Team PyTorch has recently released the latest version of PyTorch 1.7, with many changes included in the package.
Significant highlights of the python package are:
- It officially supports CUDA 11 with binaries available at www.pytorch.org.
- It supports NumPy compatible Fast Fourier transforms (FFT) via torch.fft. (Beta)
- It supports Nvidia A100 generation GPUs and native TF32 format. (Prototype)
- It supports distributed training on Windows. (Prototype)
- New updates are introduced to profiling and performance for remote procedure call (RPC), TorchScript, and Stack traces in the autograd profiler. (Stable)
- In touch audio, it has added support for speech rec (wav2letter), text to speech (WaveRNN), and source separation (ConvTasNet). (Stable)
- In torchvision, it now supports Tensor inputs, batch computation, GPU, and TorchScript.
- The native image I/O for JPEG and PNG formats are also added in PyTorch 1.7. (Stable)
PyTorch is a widely used, open-source deep learning platform used for writing neural network layers in Python. Developers worldwide use it for a smooth transition from research to production.
It provides two high-level features:
- Tensor computation (like NumPy) with strong GPU acceleration
- Deep neural networks built on a tape-based auto-grad system
It is used mainly:
- As a replacement for NumPy to utilize the power of GPUs
- As a Deep Learning platform to gain maximum strength and speed
PyTorch has delivered many applications:
- Image classification (MNIST) using Convents.
- Variational Auto-Encoders.
- World-level language modeling using LSTM RNNs
- Generative Adversarial Networks (DCGAN)
- Superresolution using an efficient sub-pixel convolutional neural network
The latest version, PyTorch 1.7.0 is available for installation on the website: www.pytorch.org.