Finish the MixPath

To finish the MixPath code

[email protected]

Done:

• NSGA-II (use pymoo)
• Plot the result

TODO:

• SNPE/OPENVINO's LookupTable

Cifar

run

Train

python S1/train_search.py \
    --exp_name experiment_name \
    --m 4\
    --data_dir ~/.torch/datasets \
    --seed 2020

Search

python S1/eval_search.py \
    --exp_name search_cifar\
    --m 4\
    --data_dir ~/.torch/datasets \
    --model_path ./super_train/experiment_name/super_train_states.pt.tar\
    --batch_size 500\
    --n_generations 40\
    --pop_size 40\
    --n_offsprings 10

Result

result of search, f1: Accuracy, f2: parameter amount, f3: GPU latency

result of search, f1: Accuracy, f2: GPU latency

Accuracy

According to https://github.com/kuangliu/pytorch-cifar

Model	Acc.
VGG16	92.64%
ResNet18	93.02%
ResNet50	93.62%
ResNet101	93.75%
MobileNetV2	94.43%
ResNeXt29(32x4d)	94.73%
ResNeXt29(2x64d)	94.82%
DenseNet121	95.04%
PreActResNet18	95.11%
DPN92	95.16%
MixPath_S1(my)	95.29%

MixPath: A Unified Approach for One-shot Neural Architecture Search

This repo provides the supernet of S₁ and our confirmatory experiments on NAS-Bench-101.

招聘 We Are Hiring！

Dear DL folks, we are opening several precious positions both for professionals and interns avid in AutoML/NAS, please send your resume/cv to [email protected]. 全职/实习生申请投递至前述邮箱。

Requirements

Python >= 3.6, Pytorch >= 1.0.0, torchvision >= 0.2.0

Datasets

CIFAR-10 can be automatically downloaded by torchvision. It has 50,000 images for training and 10,000 images for validation.

Usage

python S1/train_search.py \
    --exp_name experiment_name \
    --m number_of_paths[1,2,3,4]
    --data_dir /path/to/dataset \
    --seed 2020 \

python NasBench101/nas_train_search.py \
    --exp_name experiment_name \
    --m number_of_paths[1,2,3,4]
    --data_dir /path/to/dataset \
    --seed 2020 \

Citation

@article{chu2020mixpath,
  title={MixPath: A Unified Approach for One-shot Neural Architecture Search},
  author={Chu, Xiangxiang and Li, Xudong and Lu, Yi and Zhang, Bo and Li, Jixiang},
  journal={arXiv preprint arXiv:2001.05887},
  url={https://arxiv.org/abs/2001.05887},
  year={2020}
}