- Notifications
You must be signed in to change notification settings - Fork3
slowbull/DDG
Folders and files
| Name | Name | Last commit message | Last commit date | |
|---|---|---|---|---|
Repository files navigation
This repository is the PyTorch code for the paperDecoupled Parallel Backpropagation with Convergence Guarantee" by Zhouyuan Huo, Bin Gu, Qian Yang and Huang, Heng, ICML 2018.
For any question or suggestions, feel free to contactzhouyuan.huo@pitt.edu
- Python 3.6.2 (Anaconda)
- PyTorch 0.3.1
- CUDA 8.0
Clone and cd into the repo directory, run:
python main.py ./data/cifar.python --dataset cifar10 --arch resnet_ddg --epochs 300 --learning_rate 0.01 --schedule 150 225 --gammas 0.1 0.1 --batch_size 128 --manualSeed 2 --depth 110 --splits 2Experimental results for optimizing ResNet110 on CIFAR10 using DDG.
Note that to obtain the speedup in the paper, the code should be run in a machine with 4 GPUs.
@article{huo2018decoupled, title={Decoupled Parallel Backpropagation with Convergence Guarantee}, author={Huo, Zhouyuan and Gu, Bin and Yang, Qian and Huang, Heng}, journal={arXiv preprint arXiv:1804.10574}, year={2018}}