Open Source Python Neural Network Libraries
Sort By:
Python Neural Network Libraries
View 51 business solutionsBrowse free open source Python Neural Network Libraries and projects below. Use the toggles on the left to filter open source Python Neural Network Libraries by OS, license, language, programming language, and project status.
-
MongoDB Atlas runs apps anywhereMongoDB Atlas gives you the freedom to build and run modern applications anywhere—across AWS, Azure, and Google Cloud. With global availability in over 115 regions, Atlas lets you deploy close to your users, meet compliance needs, and scale with confidence across any geography.
-
Custom VMs From 1 to 96 vCPUs With 99.95% UptimeLive migration and automatic failover keep workloads online through maintenance. One free e2-micro VM every month.
-
1
Fast Artificial Neural Network Library is a free open source neural network library, which implements multilayer artificial neural networks in C with support for both fully connected and sparsely connected networks. Cross-platform execution in both fixed and floating point are supported. It includes a framework for easy handling of training data sets. It is easy to use, versatile, well documented, and fast. Bindings to more than 15 programming languages are available. An easy to read introduction article and a reference manual accompanies the library with examples and recommendations on how to use the library. Several graphical user interfaces are also available for the library.
-
2
Fairseq
Facebook AI Research Sequence-to-Sequence Toolkit written in Python
Fairseq(-py) is a sequence modeling toolkit that allows researchers and developers to train custom models for translation, summarization, language modeling and other text generation tasks. We provide reference implementations of various sequence modeling papers. Recent work by Microsoft and Google has shown that data parallel training can be made significantly more efficient by sharding the model parameters and optimizer state across data parallel workers. These ideas are encapsulated in the new FullyShardedDataParallel (FSDP) wrapper provided by fairscale. Fairseq can be extended through user-supplied plug-ins. Models define the neural network architecture and encapsulate all of the learnable parameters. Criterions compute the loss function given the model outputs and targets. Tasks store dictionaries and provide helpers for loading/iterating over Datasets, initializing the Model/Criterion and calculating the loss. -
3
spaCy
Industrial-strength Natural Language Processing (NLP)
spaCy is a library built on the very latest research for advanced Natural Language Processing (NLP) in Python and Cython. Since its inception it was designed to be used for real world applications-- for building real products and gathering real insights. It comes with pretrained statistical models and word vectors, convolutional neural network models, easy deep learning integration and so much more. spaCy is the fastest syntactic parser in the world according to independent benchmarks, with an accuracy within 1% of the best available. It's blazing fast, easy to install and comes with a simple and productive API. -
4
AIMET
AIMET is a library that provides advanced quantization and compression
Qualcomm Innovation Center (QuIC) is at the forefront of enabling low-power inference at the edge through its pioneering model-efficiency research. QuIC has a mission to help migrate the ecosystem toward fixed-point inference. With this goal, QuIC presents the AI Model Efficiency Toolkit (AIMET) - a library that provides advanced quantization and compression techniques for trained neural network models. AIMET enables neural networks to run more efficiently on fixed-point AI hardware accelerators. Quantized inference is significantly faster than floating point inference. For example, models that we’ve run on the Qualcomm® Hexagon™ DSP rather than on the Qualcomm® Kryo™ CPU have resulted in a 5x to 15x speedup. Plus, an 8-bit model also has a 4x smaller memory footprint relative to a 32-bit model. However, often when quantizing a machine learning model (e.g., from 32-bit floating point to an 8-bit fixed point value), the model accuracy is sacrificed. -
Try Google Cloud Risk-Free With $300 in CreditUse your credit across every product. Compute, storage, AI, analytics. When it runs out, 20+ products stay free. You only pay when you choose to.
-
5
I3D models trained on Kinetics
Convolutional neural network model for video classification
Kinetics-I3D, developed by Google DeepMind, provides trained models and implementation code for the Inflated 3D ConvNet (I3D) architecture introduced in the paper “Quo Vadis, Action Recognition? A New Model and the Kinetics Dataset” (CVPR 2017). The I3D model extends the 2D convolutional structure of Inception-v1 into 3D, allowing it to capture spatial and temporal information from videos for action recognition. This repository includes pretrained I3D models on the Kinetics dataset, with both RGB and optical flow input streams. The models have achieved state-of-the-art results on benchmark datasets such as UCF101 and HMDB51, and also won first place in the CVPR 2017 Charades Challenge. The project provides TensorFlow and Sonnet-based implementations, pretrained checkpoints, and example scripts for evaluating or fine-tuning models. It also offers sample data, including preprocessed video frames and optical flow arrays, to demonstrate how to run inference and visualize outputs. -
6
SVoice (Speech Voice Separation)
We provide a PyTorch implementation of the paper Voice Separation
SVoice is a PyTorch-based implementation of Facebook Research’s study on speaker voice separation as described in the paper “Voice Separation with an Unknown Number of Multiple Speakers.” This project presents a deep learning framework capable of separating mixed audio sequences where several people speak simultaneously, without prior knowledge of how many speakers are present. The model employs gated neural networks with recurrent processing blocks that disentangle voices over multiple computational steps, while maintaining speaker consistency across output channels. Separate models are trained for different speaker counts, and the largest-capacity model dynamically determines the actual number of speakers in a mixture. The repository includes all necessary scripts for training, dataset preparation, distributed training, evaluation, and audio separation. -
7
DeepXDE
A library for scientific machine learning & physics-informed learning
DeepXDE is a library for scientific machine learning and physics-informed learning. DeepXDE includes the following algorithms. Physics-informed neural network (PINN). Solving different problems. Solving forward/inverse ordinary/partial differential equations (ODEs/PDEs) [SIAM Rev.] Solving forward/inverse integro-differential equations (IDEs) [SIAM Rev.] fPINN: solving forward/inverse fractional PDEs (fPDEs) [SIAM J. Sci. Comput.] NN-arbitrary polynomial chaos (NN-aPC): solving forward/inverse stochastic PDEs (sPDEs) [J. Comput. Phys.] PINN with hard constraints (hPINN): solving inverse design/topology optimization [SIAM J. Sci. Comput.] Residual-based adaptive sampling [SIAM Rev., arXiv] Gradient-enhanced PINN (gPINN) [Comput. Methods Appl. Mech. Eng.] PINN with multi-scale Fourier features [Comput. Methods Appl. Mech. Eng.] -
8
FairChem
FAIR Chemistry's library of machine learning methods for chemistry
FAIRChem is a unified library for machine learning in chemistry and materials, consolidating data, pretrained models, demos, and application code into a single, versioned toolkit. Version 2 modernizes the stack with a cleaner core package and breaking changes relative to V1, focusing on simpler installs and a stable API surface for production and research. The centerpiece models (e.g., UMA variants) plug directly into the ASE ecosystem via a FAIRChem calculator, so users can run relaxations, molecular dynamics, spin-state energetics, and surface catalysis workflows with the same pretrained network by switching a task flag. Tasks span heterogeneous domains—catalysis (OC20-style), inorganic materials (OMat), molecules (OMol), MOFs (ODAC), and molecular crystals (OMC)—allowing one model family to serve many simulations. The README provides quick paths for pulling models (e.g., via Hugging Face access), then running energy/force predictions on GPU or CPU. -
9
Neural Network Intelligence
AutoML toolkit for automate machine learning lifecycle
Neural Network Intelligence is an open source AutoML toolkit for automate machine learning lifecycle, including feature engineering, neural architecture search, model compression and hyper-parameter tuning. NNI (Neural Network Intelligence) is a lightweight but powerful toolkit to help users automate feature engineering, neural architecture search, hyperparameter tuning and model compression. The tool manages automated machine learning (AutoML) experiments, dispatches and runs experiments' trial jobs generated by tuning algorithms to search the best neural architecture and/or hyper-parameters in different training environments like Local Machine, Remote Servers, OpenPAI, Kubeflow, FrameworkController on K8S (AKS etc.) DLWorkspace (aka. DLTS) AML (Azure Machine Learning) and other cloud options. NNI provides CommandLine Tool as well as an user friendly WebUI to manage training experiements. -
Gemini 3 and 200+ AI Models on One PlatformBuild generative AI apps with Vertex AI. Switch between models without switching platforms.
-
10
Stanza
Stanford NLP Python library for many human languages
Stanza is a collection of accurate and efficient tools for the linguistic analysis of many human languages. Starting from raw text to syntactic analysis and entity recognition, Stanza brings state-of-the-art NLP models to languages of your choosing. Stanza is a Python natural language analysis package. It contains tools, which can be used in a pipeline, to convert a string containing human language text into lists of sentences and words, to generate base forms of those words, their parts of speech and morphological features, to give a syntactic structure dependency parse, and to recognize named entities. The toolkit is designed to be parallel among more than 70 languages, using the Universal Dependencies formalism. Stanza is built with highly accurate neural network components that also enable efficient training and evaluation with your own annotated data. -
11
captcha_break
Identification codes
This project will use Keras to build a deep convolutional neural network to identify the captcha verification code. It is recommended to use a graphics card to run the project. The following visualization codes are jupyter notebookall done in . If you want to write a python script, you can run it normally with a little modification. Of course, you can also remove these visualization codes. captcha is a library written in python to generate verification codes. It supports image verification codes and voice verification codes. We use its function of generating image verification codes. First, we set our verification code format to numbers and capital letters, and generate a string of verification codes. It is well known that tensorflow occupies all video memory by default, which is not conducive to us conducting multiple experiments at the same time, so we can use the following code when tensorflow uses the video memory it needs instead of directly occupying all video memory. -
12
AdaNet
Fast and flexible AutoML with learning guarantees
AdaNet is a TensorFlow framework for fast and flexible AutoML with learning guarantees. AdaNet is a lightweight TensorFlow-based framework for automatically learning high-quality models with minimal expert intervention. AdaNet builds on recent AutoML efforts to be fast and flexible while providing learning guarantees. Importantly, AdaNet provides a general framework for not only learning a neural network architecture but also for learning to the ensemble to obtain even better models. At each iteration, it measures the ensemble loss for each candidate, and selects the best one to move onto the next iteration. Adaptive neural architecture search and ensemble learning in a single train call. Regression, binary and multi-class classification, and multi-head task support. A tf.estimator.Estimator API for training, evaluation, prediction, and serving models. -
13
Alpa
Training and serving large-scale neural networks
Alpa is a system for training and serving large-scale neural networks. Scaling neural networks to hundreds of billions of parameters has enabled dramatic breakthroughs such as GPT-3, but training and serving these large-scale neural networks require complicated distributed system techniques. Alpa aims to automate large-scale distributed training and serving with just a few lines of code. -
14
cnn-for-image-retrieval is a research-oriented project that demonstrates the use of convolutional neural networks (CNNs) for image retrieval tasks. The repository provides implementations of CNN-based methods to extract feature representations from images and use them for similarity-based retrieval. It focuses on applying deep learning techniques to improve upon traditional handcrafted descriptors by learning features directly from data. The code includes training and evaluation scripts that can be adapted for custom datasets, making it useful for experimenting with retrieval systems in computer vision. By leveraging CNN architectures, the project showcases how learned embeddings can capture semantic similarity across varied images. This resource serves as both an educational reference and a foundation for further exploration in image retrieval research.
-
15
Compare GAN
Compare GAN code
compare_gan is a research codebase that standardizes how Generative Adversarial Networks are trained and evaluated so results are comparable across papers and datasets. It offers reference implementations for popular GAN architectures and losses, plus a consistent training harness to remove confounding differences in optimization or preprocessing. The library’s evaluation suite includes widely used metrics and diagnostics that quantify sample quality, diversity, and mode coverage. With configuration-driven experiments, you can sweep hyperparameters, run ablations, and log results at scale. The goal is to turn GAN experimentation into a disciplined, repeatable process rather than a patchwork of scripts. It also provides baselines strong enough to serve as starting points for new ideas without re-implementing the world. -
16
CoreNet
CoreNet: A library for training deep neural networks
CoreNet is Apple’s internal deep learning framework for distributed neural network training, designed for high scalability, low-latency communication, and strong hardware efficiency. It focuses on enabling large-scale model training across clusters of GPUs and accelerators by optimizing data flow and parallelism strategies. CoreNet provides abstractions for data, tensor, and pipeline parallelism, allowing models to scale without code duplication or heavy manual configuration. Its distributed runtime manages synchronization, load balancing, and mixed-precision computation to maximize throughput while minimizing communication bottlenecks. CoreNet integrates tightly with Apple’s proprietary ML stack and hardware, serving as the foundation for research in computer vision, language models, and multimodal systems within Apple AI. The framework includes monitoring tools, fault tolerance mechanisms, and efficient checkpointing for massive training runs. -
17
Differentiable Neural Computer
A TensorFlow implementation of the Differentiable Neural Computer
The Differentiable Neural Computer (DNC), developed by Google DeepMind, is a neural network architecture augmented with dynamic external memory, enabling it to learn algorithms and solve complex reasoning tasks. Published in Nature in 2016 under the paper “Hybrid computing using a neural network with dynamic external memory,” the DNC combines the pattern recognition power of neural networks with a memory module that can be written to and read from in a differentiable way. This allows the model to learn how to store and retrieve information across long time horizons, much like a traditional computer. The architecture consists of modular components including an access module for managing memory operations, a controller (often an LSTM or feedforward network) for issuing read/write commands, and submodules for temporal linkage and memory allocation tracking. -
18
ffnet is a fast and easy-to-use feed-forward neural network training solution for python. Many nice features are implemented: arbitrary network connectivity, automatic data normalization, very efficient training tools, network export to fortran code. Now ffnet has also a GUI called ffnetui.
-
19
Flax
Flax is a neural network library for JAX
Flax is a flexible neural-network library for JAX that embraces functional programming while offering ergonomic module abstractions. Its design separates pure computation from state by threading parameter collections and RNGs explicitly, enabling reproducibility, transformation, and easy experimentation with JAX transforms like jit, pmap, and vmap. Modules define parameterized computations, but initialization and application remain side-effect free, which pairs naturally with JAX’s staging and compilation model. Flax emphasizes composability: optimizers, training loops, and checkpointing are provided as examples or utilities rather than monolithic frameworks, encouraging research-friendly customization. The library is widely used in vision, language, and reinforcement learning, often serving as a thin layer atop NumPy-like JAX primitives. Tutorials and examples show patterns for multi-host training, mixed precision, and advanced input pipelines that scale from laptops to TPUs. -
20
Forecasting Best Practices
Time Series Forecasting Best Practices & Examples
Time series forecasting is one of the most important topics in data science. Almost every business needs to predict the future in order to make better decisions and allocate resources more effectively. This repository provides examples and best practice guidelines for building forecasting solutions. The goal of this repository is to build a comprehensive set of tools and examples that leverage recent advances in forecasting algorithms to build solutions and operationalize them. Rather than creating implementations from scratch, we draw from existing state-of-the-art libraries and build additional utilities around processing and featuring the data, optimizing and evaluating models, and scaling up to the cloud. The examples and best practices are provided as Python Jupyter notebooks and R markdown files and a library of utility functions. -
21
Haiku
JAX-based neural network library
Haiku is a library built on top of JAX designed to provide simple, composable abstractions for machine learning research. Haiku is a simple neural network library for JAX that enables users to use familiar object-oriented programming models while allowing full access to JAX’s pure function transformations. Haiku is designed to make the common things we do such as managing model parameters and other model state simpler and similar in spirit to the Sonnet library that has been widely used across DeepMind. It preserves Sonnet’s module-based programming model for state management while retaining access to JAX’s function transformations. Haiku can be expected to compose with other libraries and work well with the rest of JAX. Similar to Sonnet modules, Haiku modules are Python objects that hold references to their own parameters, other modules, and methods that apply functions on user inputs. -
22
Imagen - Pytorch
Implementation of Imagen, Google's Text-to-Image Neural Network
Implementation of Imagen, Google's Text-to-Image Neural Network that beats DALL-E2, in Pytorch. It is the new SOTA for text-to-image synthesis. Architecturally, it is actually much simpler than DALL-E2. It consists of a cascading DDPM conditioned on text embeddings from a large pre-trained T5 model (attention network). It also contains dynamic clipping for improved classifier-free guidance, noise level conditioning, and a memory-efficient unit design. It appears neither CLIP nor prior network is needed after all. And so research continues. For simpler training, you can directly supply text strings instead of precomputing text encodings. (Although for scaling purposes, you will definitely want to precompute the textual embeddings + mask) -
23
Jraph
A Graph Neural Network Library in Jax
Jraph (pronounced “giraffe”) is a lightweight JAX library developed by Google DeepMind for building and experimenting with graph neural networks (GNNs). It provides an efficient and flexible framework for representing, manipulating, and training models on graph-structured data. The core of Jraph is the GraphsTuple data structure, which enables users to define graphs with arbitrary node, edge, and global attributes, and to batch variable-sized graphs efficiently for JAX’s just-in-time compilation. The library includes a comprehensive set of utilities for batching, padding, masking, and partitioning graph data, making it ideal for distributed and large-scale GNN experiments. Jraph also comes with a model zoo—a collection of forkable reference implementations of common message-passing GNN architectures, such as Graph Networks, Graph Convolutional Networks, and Graph Attention Networks. -
24
Lightweight' GAN
Implementation of 'lightweight' GAN, proposed in ICLR 2021
Implementation of 'lightweight' GAN proposed in ICLR 2021, in Pytorch. The main contribution of the paper is a skip-layer excitation in the generator, paired with autoencoding self-supervised learning in the discriminator. Quoting the one-line summary "converge on single gpu with few hours' training, on 1024 resolution sub-hundred images". Augmentation is essential for Lightweight GAN to work effectively in a low data setting. You can test and see how your images will be augmented before they pass into a neural network (if you use augmentation). The general recommendation is to use suitable augs for your data and as many as possible, then after some time of training disable the most destructive (for image) augs. You can turn on automatic mixed precision with one flag --amp. You should expect it to be 33% faster and save up to 40% memory. Aim is an open-source experiment tracker that logs your training runs, and enables a beautiful UI to compare them. -
25
MMDeploy
OpenMMLab Model Deployment Framework
MMDeploy is an open-source deep learning model deployment toolset. It is a part of the OpenMMLab project. Models can be exported and run in several backends, and more will be compatible. All kinds of modules in the SDK can be extended, such as Transform for image processing, Net for Neural Network inference, Module for postprocessing and so on. Install and build your target backend. ONNX Runtime is a cross-platform inference and training accelerator compatible with many popular ML/DNN frameworks. Please read getting_started for the basic usage of MMDeploy.
