darts
Automates neural architecture search for convolutional and recurrent networks using gradient descent.
Find high-performing CNN and RNN architectures efficiently on a single GPU with continuous relaxation.
DARTS is an implementation of differentiable neural architecture search for convolutional and recurrent neural networks. It replaces black-box search algorithms with gradient descent, optimizing continuous architecture relaxations. Deep learning researchers should use it to experiment with automated neural design on a single GPU.
What it is
DARTS converts discrete neural cell search spaces into continuous representations, allowing continuous validation loss optimization through gradient descent. The repository contains code to run search experiments for CNN cells on CIFAR-10 and RNN cells on Penn Treebank. It also includes evaluation scripts to train discovered cell genotypes from scratch and visualizers to generate architecture diagrams.
Who it's for
This tool suits deep learning researchers investigating Neural Architecture Search (NAS) and AutoML algorithms using PyTorch. It is not intended for production ML engineers who need modern, maintained, or plug-and-play Python packages.
Key features
Continuous Architecture Relaxation
Formulates discrete cell choices as continuous parameters to allow direct optimization via gradient descent.
CNN and RNN Search Spaces
Provides dedicated search modules for convolutional image classification cells and recurrent language modeling cells.
Single GPU Search Pipeline
Executes architecture search using smaller proxy models to keep compute requirements down to one GPU.
Genotype Graph Visualization
Converts learned architectural genotypes into visually legible directed graphs via Graphviz integration.
Up and running in 3 steps
Our honest verdict
What's great
- Searches complex neural cell topologies using standard gradient descent on a single GPU.
- Supports both vision (convolutional) and language (recurrent) architecture search tasks.
- Includes exact reference code for training derived genotypes from scratch on standard datasets.
- Provides built-in Graphviz scripts to visualize learned neural cell structures.
Where it falls short
- Requires legacy dependencies, specifically Python 3.5.5 and PyTorch 0.3.1, causing modern CUDA compatibility issues.
- Requires running multiple search seeds to avoid poor local minima in the continuous search space.
- Requires training candidate genotypes from scratch across 10 independent runs to evaluate accuracy properly.
Pricing & license
DARTS is free software released under the Apache-2.0 license. The only cost involved is the cost of running GPU compute resources to execute search and evaluation scripts.
Open remote AI-training roles
874 open remote AI-training roles, updated November 15, 2026; most pay $20–$170 an hour. Zekai lists them free; the application goes to the employer’s platform.
Alternatives
These reviewed AI tools do the same job instantly — from our scored directory.
Use it, or have us set it up
FAQ
Can I run DARTS with modern versions of PyTorch?
The original repository specifically requires PyTorch 0.3.1. Running version 0.4 or higher will result in out-of-memory errors without code modifications.
Do I need a massive GPU cluster to run the search?
No, the continuous relaxation approach allows architecture search to run on a single GPU using small proxy networks.
