Practical overview
When it fits
Use it to study or run the documented DARTS research workflow when you can work with its older specified Python and PyTorch stack and the relevant datasets. Account for separate full-sized training after proxy-model search.
Who it is for
- Researchers studying differentiable or neural architecture search.
- Developers seeking a research-code baseline for convolutional image-classification or recurrent language-model architecture searches.
- Practitioners who want to evaluate the repository’s documented pretrained models.
Documented uses
- Search for convolutional cells for image classification on CIFAR-10 or ImageNet.
- Search for recurrent cells for language modeling on Penn Treebank or WikiText-2.
- Train a selected architecture from scratch and evaluate it.
- Evaluate documented pretrained CIFAR-10, ImageNet or PTB models.
- Visualize learned cells, with Graphviz installed.
Requirements & limits
The repository is public and licensed under Apache-2.0. No account or pricing information is documented.
- The documented dependencies are Python >= 3.5.5, PyTorch 0.3.1 and torchvision 0.2.0; the README warns PyTorch 0.4 is unsupported and can cause out-of-memory errors.
- The README says a single GPU is sufficient for the described workflow.
- Proxy-search validation is not final performance; selected architectures require full-sized training from scratch.
- Search outcomes can differ by seed, so the README recommends repeated searches and selection based on validation after brief retraining.
- CIFAR-10 can be downloaded through torchvision, while ImageNet requires manual acquisition; the README points to external instructions for ImageNet and PTB/WT2 datasets.
Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.
Run an architecture search on CIFAR-10 or Penn Treebank
Use the documented second-order search scripts to search for convolutional cells on CIFAR-10 or recurrent cells on Penn Treebank.
Before you start
- Python >= 3.5.5, PyTorch == 0.3.1 and torchvision == 0.2.0.
- A GPU; the README says a single GPU is sufficient.
- The relevant dataset: CIFAR-10 for the CNN search or Penn Treebank (PTB) for the RNN search.
Steps
For CIFAR-10 convolutional cells, run
cd cnn && python train_search.py --unrolled.For PTB recurrent cells, run
cd rnn && python train_search.py --unrolled.Train the selected architecture from scratch using the full-sized-model workflow; do not treat search-stage validation performance as final performance.
Expected outcome
A searched cell or architecture to carry forward to full-sized training and evaluation.
Guide limitations
- Different runs can converge to different local minima; the README recommends repeating searches with different seeds and selecting cells based on validation after brief from-scratch training.
- The search-stage validation result does not indicate the final architecture’s performance.
- The README warns that PyTorch 0.4 is unsupported and can lead to out-of-memory errors.
Evaluate a pretrained CIFAR-10 model
Run the repository’s documented test command for its pretrained CIFAR-10 model.
Before you start
- Python >= 3.5.5, PyTorch == 0.3.1 and torchvision == 0.2.0.
- The CIFAR-10 pretrained checkpoint,
cifar10_model.pt. - A GPU; the README describes the implementation as requiring a single GPU.
Steps
Obtain the
cifar10_model.ptcheckpoint linked from the repository README.From the repository, run
cd cnn && python test.py --auxiliary --model_path cifar10_model.pt.
Expected outcome
The repository reports an expected test error rate of 2.63% for the checkpoint, with 3.3M model parameters.
Guide limitations
- The stated result is the README’s expected result, not a guarantee for another environment or run.
- The documented PyTorch 0.4 warning applies: that version is unsupported and can lead to out-of-memory errors.
Sources & review
Source record details
- Source availability
- Active
- Sources checked
- 16 October 2026
- Original website title
- quark0/darts · Official homepage title · retained in its source language