eric8607242 / SGNAS

[CVPR 2021] 'Searching by Generating: Flexible and Efficient One-Shot NAS with Architecture Generator'

[CVPR2021] Searching by Generating: Flexible and Efficient One-Shot NAS with Architecture Generator

Overview

This is the entire codebase for the paper Searching by Generating: Flexible and Efficient One-Shot NAS with Architecture Generator

In one-shot NAS, sub-networks need to be searched from the supernet to meet different hardware constraints. However, the search cost is high and N times of searches are needed for N different constraints. In this work, we propose a novel search strategy called architecture generator to search sub-networks by generating them, so that the search process can be much more efficient and flexible. With the trained architecture generator, given target hardware constraints as the input, N good architectures can be generated for N constraints by just one forward pass without researching and supernet retraining. Moreover, we propose a novel single-path supernet, called unified supernet, to further improve search efficiency and reduce GPU memory consumption of the architecture generator. With the architecture generator and the unified supernet, we pro- pose a flexible and efficient one-shot NAS framework, called Searching by Generating NAS (SGNAS). The search time of SGNAS for N different hardware constraints is only 5 GPU hours, which is 4N times faster than previous SOTA single-path methods. The top1-accuracy of SGNAS on ImageNet is 77.1%, which is comparable with the SOTAs.

Model Zoo

Model	FLOPs (M)	Param (M)	Top-1 (%)	Weights
SGNAS-A	373	6.0	77.1	Google drive
SGNAS-B	326	5.5	76.8	Google drive
SGNAS-C	281	4.7	76.2	Google drive

Requirements

pip3 install -r requirements.txt

• [Optional] Transfer Imagenet dataset into LMDB format by utils/folder2lmdb.py
  • With LMDB format, you can speed up entire training process(30 mins per epoch with 4 GeForce GTX 1080 Ti)

Getting Started

Search

• To search the architecture, we sample 20% images from the training set as the validation set, and the reset is kept as the training set.
  • For cifar10/100, set train_portion in ./config_file/config.yml to 0.8.
  • For Imagenet, users should split the dataset manually.

Training Unified Supernet

• For Imagenet training, set the config file ./config_file/imagenet_config.yml.
• For cifar100 training, set the config file ./config_file/cifar_config.yml.
• Set the hyperparameter warmup_epochs in the config file for training the unified supernet.

python3 search.py --cfg [CONFIG_FILE] --title [EXPERIMENT_TITLE]

Training Architecture Generator

• For Imagenet training, set the config file ./config_file/imagenet_config.yml for [CONFIG_FILE].

• For cifar100 training, set the config file ./config_file/cifar_config.yml for [CONFIG_FILE].

• If you have trained the supernet first, you can directly train the architecture generator with the pretrained supernet weight.

  • Set the hyperparameter warmup_epochs in the config file to 0 to skip the supernet training, and set the hyperparameter search_epochs for training the architecture generator.

python3 search.py --cfg [PATH_TO_CONFIG_FILE] --title [EXPERIMENT_TITLE]

• [EXPERIMENT_TITLE] is the tile for this experiment. (You can set different title for each experiment).

Train From Scratch

CIFAR10 or CIFAR100

• Set train_portion in ./config_file/cifar_config.yml to 1 to train the searched network from scratch with full training dataset.

python3 train_cifar.py --cfg [CONFIG_FILE] --flops [TARGET_FLOPS] --title [EXPERIMENT_TITLE]

• [EXPERIMENT_TITLE] is the tile for this experiment. (You can set different title for each experiment).
• [TARGET_FLOPS] is the target flops of the architecture generated from arhcitecture generator.

ImageNet

• Set the target flops and correspond config file path in run_example.sh

bash ./run_example.sh

Validate

ImageNet

• Download the ImageNet validation dataset.
• Download the checkpoint from the url above.

• SGNAS-A

python3 validate.py [PATH_TO_IMAGENET_VALIDATION_DIR] --checkpoint [CHECKPOINT_PATH] --config_path [CONFIG_FILE] --target_flops 365 --se True --activation hswish

• SGNAS-B

python3 validate.py [PATH_TO_IMAGENET_VALIDATION_DIR] --checkpoint [CHECKPOINT_PATH] --config_path [CONFIG_FILE] --target_flops 320 --se True --activation hswish

• SGNAS-C

python3 validate.py [PATH_TO_IMAGENET_VALIDATION_DIR] --checkpoint [CHECKPOINT_PATH] --config_path [CONFIG_FILE] --target_flops 275 --se True --activation hswish

Reference

• DNA
• Efficient-Pytorch
• Image2LMDB
• pytorch-image-models
• slimmable_networks

Citation

@InProceedings{sgnas,
author = {Sian-Yao Huang and Wei-Ta Chu},
title = {Searching by Generating: Flexible and Efficient One-Shot NAS with Architecture Generator},
booktitle = {Proceedings of IEEE Conference on Computer Vision and Pattern Recognition},
year = {2021}
}