---

Automated Deep Learning (AutoDL-Projects) is an open source, lightweight, but useful project for researchers. This project implemented several neural architecture search (NAS) and hyper-parameter optimization (HPO) algorithms.

Who should consider using AutoDL-Projects

• Beginners who want to try different AutoDL algorithms
• Engineers who want to try AutoDL to investigate whether AutoDL works on your projects
• Researchers who want to easily implement and experiement new AutoDL algorithms.

Why should we use AutoDL-Projects

• Simple library dependencies
• All algorithms are in the same codebase
• Active maintenance

AutoDL-Projects Capabilities

At the moment, this project provides the following algorithms and scripts to run them. Please see the details in the link provided in the description column.

Type	ABBRV	Algorithms	Description
NAS	TAS	Network Pruning via Transformable Architecture Search	NIPS-2019-TAS.md
	DARTS	DARTS: Differentiable Architecture Search	ICLR-2019-DARTS.md
	GDAS	Searching for A Robust Neural Architecture in Four GPU Hours	CVPR-2019-GDAS.md
	SETN	One-Shot Neural Architecture Search via Self-Evaluated Template Network	ICCV-2019-SETN.md
	NAS-Bench-201	NAS-Bench-201: Extending the Scope of Reproducible Neural Architecture Search	NAS-Bench-201.md
	...	ENAS / REA / REINFORCE / BOHB	NAS-Bench-201.md
HPO	HPO-CG	Hyperparameter optimization with approximate gradient	coming soon
Basic	ResNet	Deep Learning-based Image Classification	BASELINE.md

History of this repo

At first, this repo is GDAS, which is used to reproduce results in Searching for A Robust Neural Architecture in Four GPU Hours. After that, more functions and more NAS algorithms are continuely added in this repo. After it supports more than five algorithms, it is upgraded from GDAS to NAS-Projects. Now, since both HPO and NAS are supported in this repo, it is upgraded from NAS-Projects to AutoDL-Projects.

Requirements and Preparation

Please install Python>=3.6 and PyTorch>=1.3.0. (You could also run this project in lower versions of Python and PyTorch, but may have bugs). Some visualization codes may require opencv.

CIFAR and ImageNet should be downloaded and extracted into $TORCH_HOME. Some methods use knowledge distillation (KD), which require pre-trained models. Please download these models from Google Drive (or train by yourself) and save into .latent-data.

Experiments

Results on CIFAR10

Run on NAS-Projects

Model	Error(%)	Search cost(GPU days)	Genotypes	Val_acc of arch	Search-seed
DARTS-V1	3.00±0.14	0.25	Paper	-	-
DARTS-V1	-	0.26	Searched 1	89.24	90627
DARTS-V1	-	0.26	Searched 2	89.11	77047
DARTS-V1	3.56±0.04	0.26	Searched 3	89.27	90451
DARTS-V1	-	0.26	Searched 4	89.10	2189
DARTS-V2	2.76±0.09	1	Paper	-	-
DARTS-V2	-	1.32	Searched 1	89.13	52665
DARTS-V2	-	1.33	Searched 2	89.16	68513
DARTS-V2		1.43	Searched 3	89.20	33376
DARTS-V2	-	1.33	Searched 4	89.17	74480

The command to search architectures.

CUDA_VISIBLE_DEVICES=0 bash ./scripts-search/DARTS1V-search-NASNet-space.sh cifar10 -1
CUDA_VISIBLE_DEVICES=0 bash ./scripts-search/DARTS2V-search-NASNet-space.sh cifar10 -1

The command to train.

CUDA_VISIBLE_DEVICES=0 bash ./scripts/nas-infer-train.sh cifar10  DARTS_V1 96 -1
CUDA_VISIBLE_DEVICES=0 bash ./scripts/nas-infer-train.sh cifar10  DARTS_V2 96 -1

Model	Error(%)	Search cost(GPU days)	Genotypes	Val_acc of arch	Search-seed
GDAS	2.93	0.21	Paper	-	-
GDAS	2.82(Reproduced)	-	Paper	-	-
GDAS	-	0.19	Searched 1	78.26	11110
GDAS	-	0.19	Searched 2	78.23	21945
GDAS	-	0.19	Searched 3	78.03	39472
GDAS	2.89±0.05	0.19	Searched 4	78.60	49342

The command to search architectures.

CUDA_VISIBLE_DEVICES=0 bash ./scripts-search/GDAS-search-NASNet-space.sh cifar10 -1

The command to train.

CUDA_VISIBLE_DEVICES=0 bash ./scripts/nas-infer-train.sh cifar10  GDAS_V1 96 -1

Model	Error(%)	Search cost(GPU days)	Genotypes	Val_acc of arch	Search-seed
DARTS+	2.50±0.11	0.4	Paper	-	-
DARTS+	-	0.42	Searched 1	84.06	67841
DARTS+	-	0.56	Searched 2	85.41	68221
DARTS+	2.97±0.07	0.51	Searched 3	85.78	99594
DARTS+	-	0.49	Searched 4	84.75	22713
DARTS+(V1)	-	0.14	Searched 1	86.02	38979
DARTS+(V1)	-	0.16	Searched 2	85.56	57053
DARTS+(V1)	3.19±0.00	0.17	Searched 3	87.20	34853
DARTS+(V1)	-	0.14	Searched 4	85.28	21462

The command to search and train are same as DARTS.

Run on quark0/darts

Model	Error(%)	Search cost(GPU days)	Genotypes	Val_acc of arch	Search-seed
DARTS-V1	3.00±0.14	0.25	Paper	-	-
DARTS-V1	-	0.38	Searched 1	88.97	2
DARTS-V1	-	0.35	Searched 2	89.08	444
DARTS-V1	2.99±0.04	0.35	Searched 3	89.31	666
DARTS-V1	-	0.34	Searched 4	89.30	999
DARTS-V2	2.76±0.09	1	Paper	-	-
DARTS-V2	-	0.83	Searched 1	89.23	2
DARTS-V2	-	1.04	Searched 2	89.17	555
DARTS-V2	3.02±0.16	0.79	Searched 3	89.34	777
DARTS-V2	-	0.79	Searched 4	89.02	888

The command to search architectures.

python train_search.py     # for conv cells on CIFAR-10(DARTS-V1)
python train_search.py --unrolled     # for conv cells on CIFAR-10(DARTS-V2)

The command to train.

python train.py --auxiliary --cutout            # CIFAR-10

Model	Error(%)	Search cost(GPU days)	Genotypes	Val_acc of arch	Search-seed
DARTS+	2.50±0.11	0.4	Paper	-	-
DARTS+	-	0.58	Searched 1	85.64	233
DARTS+	-	0.52	Searched 2	85.99	694
DARTS+	2.97±0.07	0.73	Searched 3	88.48	1113
DARTS+	-	0.54	Searched 4	86.38	3763
DARTS+(V1)	-	0.40	Searched 1	85.64	38
DARTS+(V1)	-	0.61	Searched 2	85.99	2981
DARTS+(V1)	2.69±0.06	0.31	Searched 3	86.39	8632
DARTS+(V1)	-	0.23	Searched 4	83.94	3659

The command to search and train the DARTS+ are same as DARTS.

P.S. The searched architecture in DARTS+(V1) is out of memory when training by the quark0/darts. So I train it by the NAS-Projects code. The command is

CUDA_VISIBLE_DEVICES=0 bash ./scripts/nas-infer-train.sh cifar10  DARTS_V1 96 -1

Run on DenseNAS

Model	Error(%)	Search cost(GPU days)	Search Space	Params(M)	Val_acc of arch	latency	Search-seed
DenseNAS	13.15	0.75	mbv2	2.05	58.44	3.66	2
DenseNAS	11.51	0.73	Darts	0.5	21.04	5.68	2

Comparison

Model	Error(%)	Search cost(GPU days)	Genotypes	Params(M)	Train cost()GPU days
DARTS-V1	3.00±0.14	0.25	Paper	3.3	-
DARTS-V2	2.76±0.09	1	Paper	3.3	-
GDAS	2.93	0.21	Paper	3.4	-
NASP	2.83±0.09	0.1	Paper	3.3	-
NASP(12ops)	2.44±0.04	0.2	Paper	7.4	-
DARTS-V1	2.99±0.04	0.35	Our Searched (quark0/darts)	3.20	1.71
DARTS-V1	3.56±0.04	0.26	Our Searched (NAS-Projects)	1.73	1.33
DARTS-V2	3.02±0.16	0.79	Our Searched (quark0/darts)	3.17	1.46
DARTS-V2		1.43	Our Searched (NAS-Projects)
GDAS	2.89±0.05	0.19	Our Searched	3.77	2.25
NASP	3.34±0.07	0.25	Our Searched	2.40	1.29
NASP(12ops)	2.44±0.04	0.25	Our Searched	8.27	1.5

Citation

If you find that this project helps your research, please consider citing some of the following papers:

@inproceedings{dong2020nasbench201,
  title     = {NAS-Bench-201: Extending the Scope of Reproducible Neural Architecture Search},
  author    = {Dong, Xuanyi and Yang, Yi},
  booktitle = {International Conference on Learning Representations (ICLR)},
  url       = {https://openreview.net/forum?id=HJxyZkBKDr},
  year      = {2020}
}
@inproceedings{dong2019tas,
  title     = {Network Pruning via Transformable Architecture Search},
  author    = {Dong, Xuanyi and Yang, Yi},
  booktitle = {Neural Information Processing Systems (NeurIPS)},
  year      = {2019}
}
@inproceedings{dong2019one,
  title     = {One-Shot Neural Architecture Search via Self-Evaluated Template Network},
  author    = {Dong, Xuanyi and Yang, Yi},
  booktitle = {Proceedings of the IEEE International Conference on Computer Vision (ICCV)},
  pages     = {3681--3690},
  year      = {2019}
}
@inproceedings{dong2019search,
  title     = {Searching for A Robust Neural Architecture in Four GPU Hours},
  author    = {Dong, Xuanyi and Yang, Yi},
  booktitle = {Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
  pages     = {1761--1770},
  year      = {2019}
}

Others

If you want to contribute to this repo, please see CONTRIBUTING.md. Besides, please follow CODE-OF-CONDUCT.md.

License

The entire codebase is under MIT license