Yonggan Fu, Han Guo, Meng Li, Xin Yang, Yining Ding, Vikas Chandra, Yingyan Lin
Accepted at ICLR 2021 (Spotlight) [Paper Link].
Low-precision deep neural network (DNN) training has gained tremendous attention as reducing precision is one of the most effective knobs for boosting DNNs’ training time/energy efficiency. In this paper, we attempt to explore low-precision training from a new perspective as inspired by recent findings in understanding DNN training: we conjecture that DNNs’ precision might have a similar effect as the learning rate during DNN training, and advocate dynamic precision along the training trajectory for further boosting the time/energy efficiency of DNN training. Specifically, we propose Cyclic Precision Training (CPT) to cyclically vary the precision between two boundary values to balance the coarse-grained exploration of low precision and fine-grained optimization of high precision. Through experiments and visualization we show that CPT helps to (1) converge to a wider minima with a lower generalization error and (2) reduce training variance, which opens up a new design knob for simultaneously improving the optimization and efficiency of DNN training.
We evaluate CPT on eleven models & five datasets (i.e., ResNet-38/74/110/152/164/MobileNetV2 on CIFAR-10/100, ResNet-18/34/50 on ImageNet, Transformer on WikiText-103, LSTM on PTB). Please refer to our paper for more results.
- Test accuracy vs. training computational cost
- Loss landscape visualization
- Accuracy - training efficiency trade-off
- Boosting optimality
cpt_cifar and cpt_imagenet are the codes customized for CIFAR-10/100 and ImageNet, respectively, with a similar code structure.
See env.yml for the complete conda environment. Create a new conda environment:
conda env create -f env.yml
conda activate pytorch
In addition to the commonly considered args, e.g., the target network, dataset, and data path via --arch, --dataset, and --datadir, respectively, you also need to: (1) enable cyclic precision training via --is_cyclic_precision; (2) specify the precision bounds for both forward (weights and activations) and backward (gradients and errors) with --cyclic_num_bits_schedule and --cyclic_num_grad_bits_schedule, respectively (note that in CPT, we adopt a constant precision during backward for more stable training process as analyzed in our appendix); (3) specify the number of cyclic periods via --num_cyclic_period which can be set as 32 in all experiments and more ablation studies can be found in Sec. 4.3 of our paper.
- Example: Training ResNet-74 on CIFAR-100 with CPT (3~8-bit forward, 8-bit backward, and a cyclic periods of 32).
cd cpt_cifar
python train.py --save_folder ./logs --arch cifar100_resnet_74 --workers 4 --dataset cifar100 --datadir path-to-cifar100 --is_cyclic_precision --cyclic_num_bits_schedule 3 8 --cyclic_num_grad_bits_schedule 8 8 --num_cyclic_period 32
We also integrate SWA in our code although it is not used in the reported results of our paper.
The args for ImageNet experiments are similar with the ones on CIFAR-10/100.
- Example: Training ResNet-34 on ImageNet with CPT (3~8-bit forward, 8-bit backward, and a cyclic periods of 32).
cd cpt_imagenet
python train.py --save_folder ./logs --arch resnet34 --warm_up --datadir PATH_TO_IMAGENET --is_cyclic_precision --cyclic_num_bits_schedule 3 8 --cyclic_num_grad_bits_schedule 8 8 --num_cyclic_period 32 --automatic_resume
@article{fu2021cpt,
title={CPT: Efficient Deep Neural Network Training via Cyclic Precision},
author={Fu, Yonggan and Guo, Han and Li, Meng and Yang, Xin and Ding, Yining and Chandra, Vikas and Lin, Yingyan},
journal={arXiv preprint arXiv:2101.09868},
year={2021}
}
Please also check our work on how to fractionally squeeze out more training cost savings from the most redundant bit level, progressively along the training trajectory and dynamically per input:
Yonggan Fu, Haoran You, Yang Zhao, Yue Wang, Chaojian Li, Kailash Gopalakrishnan, Zhangyang Wang, Yingyan Lin. "FracTrain: Fractionally Squeezing Bit Savings Both Temporally and Spatially for Efficient DNN Training". NeurIPS, 2020. [Paper Link] [Code]