Timm Model Finetuning

Contributor: Pinyuan Feng (Tony)

Descriptiion

This repo is used to finetune the Timm models on ImageNet. Accelerate library is adopted to achieve efficient, easy-to-deployed parallel computing.

Environment Setup

conda create -n hmn python=3.9 -y
conda activate timm
conda install pytorch==1.13.1 torchvision==0.14.1 pytorch-cuda=11.7 -c pytorch -c nvidia
pip install timm==0.9.0 
pip install wandb accelerate pathlib numpy tqdm scipy torchmetrics pandas matplotlib

Commands

Initialization

accelerate config

After that you will see a list of questions. Just set up the configuration based on your situation

Dummy Training and Evaluation

Before starting your large-scale training, you can use following code to test your scrip:

accelerate launch --main_process_port 29501 main_accelerate_timm.py -dm True -mn 'resnet50.tv_in1k' -ep 10 -bs 16 -pt True

accelerate launch main_accelerate_timm.py dm True -mn 'resnet50.tv_in1k' -bs 64 -ev True -rs True -pt True

Training and Evaluation

accelerate launch --main_process_port 29501 main_accelerate_timm.py -mn 'resnet50.tv_in1k' -ep 20 -bs 256 -pt True -wb True

accelerate launch main_accelerate_timm.py dm True -mn 'resnet50.tv_in1k' -bs 256 -ev True -rs True -bt "best_acc"

Notice

• Enter your token and project name to enable W&B for logging
• Modify your data path and checkpoint path
• More data augmentation techniques could be added to enhance your training
• --main_process_port 29501 could be removed if your first GPU (0) is not used by other programs.
• You can visit this github page to find the models you want.