Yonggan Fu, Zhongzhi Yu, Yongan Zhang, Yifan Jiang, Chaojian Li, Yongyuan Liang, Mingchao Jiang, Zhangyang Wang, Yingyan Lin
Accepted at DAC 2021 [Paper Link].
- The promise of Deep Neural Network (DNN) powered Internet of Thing (IoT) devices has motivated a tremendous demand for automated solutions to enable fast development and deployment of efficient (1) DNNs equipped with instantaneous accuracy-efficiency trade-off capability to accommodate the timevarying resources at IoT devices and (2) dataflows to optimize DNNs’ execution efficiency on different devices. Therefore, we propose InstantNet to automatically generate and deploy instantaneously switchable-precision networks (SP-Nets) which operate at variable bit-widths.
- To generate SP-Nets of which the accuracy under all bit-widths are the same or even higher than that of DNNs customized for individual bit-widths, the proposed Cascade Distillation Training (CDT) adopts all higher precisions to guide the learning of the current precision, motivated by the fact that the quantization noises of SP-Nets under adjacent or closer bit-widths are smaller.
- The proposed SP-NAS updates the weights based on our CDT method which explicitly incorporates switchable-precision property into the training process; and for updating the architecture parameters of SPNet, it activates only the weights under the lowest bit-width, for generating networks forced to inherently tackle SP-Nets’ bottleneck of high accuracy loss under the lowest bit-width.
- The proposed AutoMapper aims to accept (1) DNNs (e.g., SP-Nets generated by our SP-NAS), (2) the target device, and (3) target hardware efficiency, and then generate mapping methods that maximize both the task accuracy and hardware efficiency of the given SP-Nets under all bit-widths when being executed on the target device.
- InstantNet generated systems consistently outperforms the SOTA baselines in terms of the trade-off between accuracy and EDP (a commonly-used hardware metric for ASIC) by achieving a higher or comparable accuracy and better EDP under lower bit-widths over the baselines.
- A two-stage process is needed to generate the final system, i.e., the search and training stage. The corresponding settings are specified in config_search.py and config_train.py, respectively. In particular, plz specify the search settings in config_search.py and search the optimal SP-Nets through train_search.py, and then specify training settings in config_train.py and train the searched arch from scratch through train.py (The best searched arch will be saved at ckpt/search/arch.pt which is the default path for train.py to read.)
- See env.yml for the complete conda environment. Create a new conda environment:
conda env create -f env.yml
conda activate pytorch
- Specify the search setting in config_search.py:
C.dataset_path = "path-to-dataset"
C.flops_weight = 1e-9
- Run train_search.py:
python train_search.py
- Specify the training setting in config_train.py:
C.dataset_path = "path-to-dataset"
C.load_path = "path-to-searched-arch"
- Run train.py on each of your nodes:
python train.py
Search/training logs: the searched arch ckpt/search/arch.pt the search logs ckpt/search/logs.txt and the training logs ckpt/finetune/logs.txt.
@article{fu2021instantnet,
title={InstantNet: Automated Generation and Deployment of Instantaneously Switchable-Precision Networks},
author={Fu, Yonggan and Yu, Zhongzhi and Zhang, Yongan and Jiang, Yifan and Li, Chaojian and Liang, Yongyuan and Jiang, Mingchao and Wang, Zhangyang and Lin, Yingyan},
journal={arXiv preprint arXiv:2104.10853},
year={2021}
}