Part of Final Project for 2021 Intro to Deep Learning System. DINO algorithm implementation, from DINO official.
Author: Yuanzhe Liu yl9539
For the official implementation, follow this link: https://github.com/facebookresearch/dino
For the official README, README_DINO.md is the one file.
In my Intro to Deep Learning System Final Project, I used MoBY to train Swin Transformer as the backbone.
In this repo, I use DINO to train ResNet-50 as a backbone.
For Swin Transformer with MoBY, check out this repo below:
https://github.com/ggflow123/DLS_Final_Project
For Mask R-CNN with FPN using Swin Transformer or ResNet-50 as backbone, please check:
https://github.com/ggflow123/DLS_FINAL_mmdetection
For the program, first make a director called resnet-50, then run:
python -m torch.distributed.launch --nproc_per_node=1 main_dino.py --arch resnet50 --data_path /scratch/yl9539/Transformer-SSL/imagenet/train --output_dir ./resnet-50/MODEL_NAME
On NYU HPC Greene or similar Linux Environment cloud computing platform, for 4GPU, 24 hours training, do:
sbatch run_resnet.slurm
For testing whether the code can run on the platform, for 4GPU, 1 hours testing, do:
sbatch run_test.slurm