| PyTorch | |
|---|---|
 | |
| Original authors |
|
| Developer | Meta AI |
| Initial release | September 2016; 9 years ago (2016-09)[1] |
| Stable release | |
| Written in | |
| Operating system | |
| Platform | IA-32,x86-64,ARM64 |
| Available in | English |
| Type | Library fordeep learning |
| License | BSD-3[3] |
| Website | pytorch |
| Repository | github |
| Part of a series on |
| Machine learning anddata mining |
|---|
Learning with humans |
Model diagnostics |
PyTorch is anopen-sourcedeep learninglibrary, originally developed byMeta Platforms and currently developed with support from theLinux Foundation. The successor toTorch, PyTorch provides ahigh-levelAPI that builds upon optimised, low-level implementations of deep learning algorithms and architectures, such as theTransformer, orSGD. Notably, this API simplifies modeltraining and inference to a few lines of code. PyTorch allows for automaticparallelization of training and, internally, implementsCUDA bindings that speed training further byleveraging GPU resources.
PyTorch utilises thetensor as a fundamentaldata type, similarly toNumPy. Training is facilitated by areversed automatic differentiation system, Autograd, that constructs adirected acyclic graph of the operations (and their arguments) executed by a model during its forward pass. With aloss,backpropagation is then undertaken.[4]
As of 2025[update], PyTorch remains one of the most popular deep learning libraries, alongside others such asTensorFlow andKeras.[5] A number of commercial deep learning architectures are built on top of PyTorch, includingChatGPT,[6]Tesla Autopilot,[7]Uber's Pyro,[8]Hugging Face's Transformers,[9][10] and Catalyst.[11][12]
In 2001, Torch was written and released under aGPL. It was a machine-learning library written in C++ and CUDA, supporting methods including neural networks,support vector machines (SVM),hidden Markov models, etc.[13][14][15] Around 2010, it was rewritten to by Ronan Collobert, Clement Farabet and Koray Kavuckuoglu. This was known as Torch7 or LuaTorch. This was written so that the backend was inC and the frontend was inLua.[16] In mid-2016, some developers refactored it to decouple the frontend and the backend, with strong influence from torch-autograd andChainer. In turn, torch-autograd was influenced byHIPS/autograd. Development on Torch7 ceased in 2018 and was subsumed by the PyTorch project.[17][18]
Meta (formerly known as Facebook) operates both PyTorch and Convolutional Architecture for Fast Feature Embedding (Caffe2), but models defined by the two frameworks were mutually incompatible. TheOpen Neural Network Exchange (ONNX) project was created by Meta andMicrosoft in September 2017 for converting models between frameworks. Caffe2 was merged into PyTorch at the end of March 2018.[19] In September 2022, Meta announced that PyTorch would be governed by the independent PyTorch Foundation, a newly created subsidiary of theLinux Foundation.[20]
PyTorch 2.0 was released on 15 March 2023, introducingTorchDynamo, a Python-levelcompiler that makes code run up to two times faster, along with significant improvements in training and inference performance across majorcloud platforms.[21][22]
PyTorch defines a class called Tensor (torch.Tensor) to store and operate on homogeneous multidimensional rectangular arrays of numbers. PyTorch Tensors are similar toNumPy Arrays, but can also be operated on by aCUDA-capableNVIDIAGPU. PyTorch has also been developing support for other GPU platforms, for example, AMD'sROCm[23] and Apple'sMetal Framework.[24]
PyTorch supports various sub-types of Tensors.[25]
The meaning of the word "tensor" in machine learning is only superficially related to its original meaning in mathematics or physics as a certain kind of object inlinear algebra. Tensors in PyTorch are simply multi-dimensional arrays.
PyTorch defines a module called nn (torch.nn) to describe neural networks and to support training. This module offers a comprehensive collection of building blocks for neural networks, including various layers and activation functions, enabling the construction of complex models. Networks are built by inheriting from thetorch.nn module and defining the sequence of operations in theforward() function.
The following program shows the low-level functionality of the library with a simple example.
importtorchdtype=torch.floatdevice=torch.device("cpu")# Execute all calculations on the CPU# device = torch.device("cuda:0") # Executes all calculations on the GPU# Create a tensor and fill it with random numbersa=torch.randn(2,3,device=device,dtype=dtype)print(a)# Output: tensor([[-1.1884, 0.8498, -1.7129],# [-0.8816, 0.1944, 0.5847]])b=torch.randn(2,3,device=device,dtype=dtype)print(b)# Output: tensor([[ 0.7178, -0.8453, -1.3403],# [ 1.3262, 1.1512, -1.7070]])print(a*b)# Output: tensor([[-0.8530, -0.7183, 2.58],# [-1.1692, 0.2238, -0.9981]])print(a.sum())# Output: tensor(-2.1540)print(a[1,2])# Output of the element in the third column of the second row (zero-based)# Output: tensor(0.5847)print(a.max())# Output: tensor(0.8498)
The following code block defines a neural network with linear layers using thenn module.
fromtorchimportnn# Import the nn sub-module from PyTorchclassNeuralNetwork(nn.Module):# Neural networks are defined as classesdef__init__(self):# Layers and variables are defined in the __init__ methodsuper().__init__()# Must be in every network.self.flatten=nn.Flatten()# Construct a flattening layer.self.linear_relu_stack=nn.Sequential(# Construct a stack of layers.nn.Linear(28*28,512),# Linear Layers have an input and output shapenn.ReLU(),# ReLU is one of many activation functions provided by nnnn.Linear(512,512),nn.ReLU(),nn.Linear(512,10),)defforward(self,x):# This function defines the forward pass.x=self.flatten(x)logits=self.linear_relu_stack(x)returnlogits
{{citation}}: CS1 maint: work parameter with ISBN (link)| Open source | |
|---|---|
| Proprietary |
|
Differentiable computing | |
|---|---|
| General | |
| Hardware | |
| Software libraries | |
