Framework học sâu mã nguồn mở được dùng rộng rãi, do PyTorch Foundation quản lý
Nền tảng phát triển AI pytorch.org
What is PyTorch
PyTorch is an open-source deep learning library, originally developed by Meta (formerly Facebook) and governed since September 2022 by the PyTorch Foundation under the Linux Foundation. It uses the tensor as its basic data type, computes gradients automatically with Autograd and speeds up training on GPUs, and it is one of the most popular deep learning frameworks. Systems built on PyTorch include ChatGPT, Tesla Autopilot and Hugging Face Transformers.
What it can do
- Build and train neural networks, defining, training and running models in a few lines of code
- GPU acceleration through CUDA and other backends, with parallel training
- Dynamic computation graphs: code runs the way Python does, which makes debugging straightforward
- Compilation: PyTorch 2.0, released in March 2023, added the TorchDynamo compiler, which can make code run up to twice as fast
- A rich ecosystem: the PyTorch Foundation also hosts projects such as vLLM, DeepSpeed, Ray, ExecuTorch and Safetensors
Pricing
Free and open source.
Who it's for
- Students and researchers working on deep learning
- AI engineers training and fine-tuning large models
- Developers getting started with neural networks
Good to know
- Install the build that matches your GPU and CUDA version; the official homepage has an install command generator.
- PyTorch is fairly low level, so if you only want to use existing models, higher-level libraries such as Hugging Face Transformers are easier.
Similar tools
Compare TensorFlow, JAX and Keras; in China, Baidu's PaddlePaddle is an alternative.