• XNAS
• • Supported Algorithms
  • • Stable
    • Beta
  • Supported Search Spaces
  • • Stable
    • Beta
    • • Cell-based Search Space
      • Chain-structured Search Space
  • Installation
  • Usage Examples
  • Experiment Results
  • • DARTS Search Space
    • Results on CIFAR10
    • TODO
  • BibTex

XNAS

XNAS is an effective, modular and flexible neural architecture search (NAS) codebase, which aims to provide a common framework and baselines for the NAS community.

This project is now supported by PengCheng Lab

Supported Algorithms

Stable

Beta

• DARTS
  • python search/DARTS.py --cfg configs/search/DARTS.yaml
• PCDARTS
  • python search/PDARTS.py --cfg configs/search/PDARTS.yaml
• PDARTS
  • python search/PCDARTS.py --cfg configs/search/PCDARTS.yaml
• SNG
• ASNG
• MDENAS
• DDPNAS
• MIGONAS
• GridSearch
• DrNAS
  • python search/DrNAS/nb201space.py --cfg configs/search/DrNAS/nb201_cifar10_Dirichlet.yaml
  • python search/DrNAS/nb201space.py --cfg configs/search/DrNAS/nb201_cifar100_Dirichlet.yaml
  • python search/DrNAS/DARTSspace.py --cfg configs/search/DrNAS/DARTS_cifar10.yaml
• TENAS
  • python search/TENAS.py --cfg configs/search/TENAS/nb201_cifar10.yaml
• RMINAS
  • ./search/RMINAS/download_weight.sh # prepare weights of teacher models
  • ./python search/RMINAS/RMINAS_nb201.py --cfg configs/search/RMINAS/nb201_cifar10.yaml
  • ./python search/RMINAS/RMINAS_darts.py --cfg configs/search/RMINAS/darts_cifar10.yaml

Supported Search Spaces

Stable

Beta

Cell-based Search Space

• NAS-Bench-1Shot1
• DARTS
• NAS-Bench-201

Chain-structured Search Space

• MobileNetV3
• OFA
• ProxylessNAS

---

Installation

git clone https://git.openi.org.cn/PCL_AutoML/XNAS
cd XNAS
# set root path
export PYTHONPATH=$PYTHONPATH:/Path/to/XNAS

---

Usage Examples

# set gpu devices. Multiple GPUs are under test and may cause errors now.
export CUDA_VISIBLE_DEVICES=0
# unit test example
python tools/test_func/sng_function_optimization.py
# train example
python train/DARTS_train.py --cfg configs/search/darts.yaml
# replace config example
python train/DARTS_train.py --cfg configs/search/darts.yaml OUT_DIR /username/project/XNAS/experiment/darts/test1

---

Experiment Results

DARTS Search Space

On cifar10, the network is trained by using the default training set of pt.darts.

We reimplement several widely used NAS methods including:

• Darts, paper, official_code

Results on CIFAR10

Method	Trial	params(M)	search(hrs)	train(hrs)	Top1	Flops(M)	download	Search Top1	Search Space
darts	1	4.39	21	39	96.97	689.335	-	90.32	cell-based
darts	2	4.25	26.36	39	97.31	680.073	-	90.47	cell-based
darts	3	4.450	27.63	39.7	97.32	708.468	-	90.09	cell-based
darts	4	4.467	21.48	48	97.39	717.454	-	90.52	cell-based
paper	-	3.3	96	-	97.24	-	-	-	cell-based
sng	1	3.042	2.5	33.45	96.87	506.002	-	87.52	cell-based
sng	2	2.477	3.0	26.62	96.73	397.068	-	87.81	cell-based
sng	3	2.087	3.0	21.75	96.56	339.201	-	87.00	cell-based
sng	4	3.230	2.5	27.47	97.30	509.071	-	88.51	cell-based
asng	1	2.001	2.5	18.98	96.61	330.575	-	85.78	cell-based
asng	2	2.749	2.5	25.66	96.48	450.153	-	87.47	cell-based
asng	3	2.991	2.5	27.88	97.31	476.695	-	85.52	cell-based
asng	4	2.189	2.5	23.88	96.55	350.647	-	86.41	cell-based
MIGO	1	3.266	1.5	28.20	97.35	531.217	-	84.75	cell-based
MIGO	2	3.274	1.5	26.33	97.41	523.973	-	84.61	cell-based
MIGO	3	2.848	1.5	25.91	97.36	451.480	-	84.89	cell-based
MIGO	4	2.749	1.5	30.19	97.28	439.619	-	84.44	cell-based
pcdarts(official)	1	4.052	3.61	41.28	97.20	638.823		85.296	cell-based
pcdarts(official)	2	3.247	3.6	27.96	97.23	512.444		84.552	cell-based
pcdarts(official)	3	4.368	3.63	38.68	97.25	688.561		84.792	cell-based
pcdarts(official)	4	4.148	3.16	34.58	97.49	649.108		85.280	cell-based
xnas-pcdarts	1	3.779	3.46	31.9	97.55	595.498		85.192	cell-based
xnas-pcdarts	2	3.641	3.03	33.08	96.92	573.933		85.036	cell-based
xnas-pcdarts	3	4.536	3.02	41.18	97.37	722.790		84.592	cell-based
xnas-pcdarts	4	3.143	3.48	32.26	97.25	505.227		85.088	cell-based
pdarts(official)	1	4.052	3.50	-	97.41	555.270		-	cell-based
pdarts(official)	2	3.247	3.31	-	97.25	529.419		-	cell-based
pdarts(official)	3	4.368	3.39	-	97.25	545.732		-	cell-based
pdarts(official)	4	4.148	4.08	-	97.29	642.555		-	cell-based
dynamic_ASNG	1	2.901	0.0	31.77	96.86	465.193	-	78.65	cell-based
dynamic_ASNG	2	2.208	0.0	18.00	96.78	351.145	-	79.2	cell-based
dynamic_ASNG	3	2.365	0.0	19.93	96.20	387.364	-	79.87	cell-based
dynamic_ASNG	4	3.466	0.0	31.35	97.11	565.058	-	79.87	cell-based
dynamic_SNG	1	2.245	0.0	23.98	96.28	352.693	-	78.95	cell-based
dynamic_SNG	2	2.927	0.0	24.13	96.87	473.156	-	78.07	cell-based
dynamic_SNG	3	2.724	0.0	28.07	97.45	442.826	-	77.68	cell-based
dynamic_SNG	4	3.323	0.0	31.85	96.65	528.784	-	79.78	cell-based
RMINAS	-	-	1.92	31.9	97.36	-	-	-	cell-based

TODO

1. data parallel support
2. fix Nvidia DALI backend support
3. add **kwargs for space_builder
4. test code for imagenet

BibTex

@inproceedings{zheng2022rminas,
  title={Neural Architecture Search with Representation Mutual Information},
  author={Xiawu Zheng, Xiang Fei, Lei Zhang, Chenglin Wu, Fei Chao, Jianzhuang Liu, Wei Zeng, Yonghong Tian, Rongrong Ji},
  journal={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year={2022}
}
@article{zheng2021migo,
  title={MIGO-NAS: Towards fast and generalizable neural architecture search},
  author={Zheng, Xiawu and Ji, Rongrong and Chen, Yuhang and Wang, Qiang and Zhang, Baochang and Chen, Jie and Ye, Qixiang and Huang, Feiyue and Tian, Yonghong},
  journal={IEEE Transactions on Pattern Analysis and Machine Intelligence},
  year={2021},
  publisher={IEEE}
}
@inproceedings{zheng2020rethinking,
  title={Rethinking performance estimation in neural architecture search},
  author={Zheng, Xiawu and Ji, Rongrong and Wang, Qiang and Ye, Qixiang and Li, Zhenguo and Tian, Yonghong and Tian, Qi},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={11356--11365},
  year={2020}
}