PyTorch 1.7 has been released. This version adds many new APIs, including support for NumPy-compatible FFT operations, tools for performance analysis, and support for Distributed Data Parallel (DDP) and remote-based A major update of distributed training for remote procedure call (RPC).
In addition, some functions have stabilized, including custom C++ classes, memory analyzers, extensions through similar custom tensor objects, user asynchronous functions in RPC, and many other functions in torch, such as Per-RPC timeout, DDP dynamic bucketing and RRef helper.
Some update highlights are as follows:
- CUDA 11 is officially supported, and the binary files can be downloaded from PyTorch.org .
- Update and add analysis and performance of RPC, TorchScript and Stack traces in autograd analyzer
- (Beta) Support NumPy-compatible Fast Fourier Transform (FFT) through torch.fft
- (Prototype) Support Nvidia's new generation A100 GPU and native TF32 format
- (Prototype) Now supports distributed training on Windows
- torchvision
- (Stable) Transformation now supports Tensor input, batch calculation, GPU and TorchScript
- (Stable) Native image I/O for JPEG and PNG formats
- (Beta) New video reader API
- torchaudio
- (Stable) Added support for voice recording (wav2letter), text to speech (WaveRNN) and source separation (ConvTasNet)
It is worth noting that starting from PyTorch 1.6 , the state of the function will be divided into three types, namely stable, beta and prototype.
The full release notes can be found here .