youngminpil / autolr

AutoLR: Layer-wise Pruning and Auto-tuning of Learning Rates in Fine-tuning of Deep Networks

AutoLR: Layer-wise Pruning and Auto-tuning of Learning Rates in Fine-tuning of Deep Networks

Official Pytorch implementation of paper:

AutoLR: Layer-wise Pruning and Auto-tuning of Learning Rates in Fine-tuning of Deep Networks (AAAI 2021).

Project page

Environment

Python 3.6, Pytorch 0.4.1, Torchvision, tensorboard

Train

Default setting:

• Architecture: ResNet-50
• Dataset: CUB2011 or Cars-196 retrieval
• Batch size: 40
• Image size: 224X224

prepare

The dataset path should be changed to your own path.

CUB2011-200 dataset are available on https://drive.google.com/file/d/1hbzc_P1FuxMkcabkgn9ZKinBwW683j45/view

Cars-196 dataset are available on https://ai.stanford.edu/~jkrause/cars/car_dataset.html

prepare_cub.py 

train network.

The dataset path(data_dir='/home/ro/FG/CUB_200_2011/pytorch') should be changed to your own path.

train_CUB.py --dataset CUB-200 --max_f 0.4 --min_f 2

In the case of Cars-196 retrieval dataset training,

train_CUB.py --dataset Cars-196 --max_f 0.4 --min_f 2

Citation

@inproceedings{ro2021autolr,
      title={AutoLR: Layer-wise Pruning and Auto-tuning of Learning Rates in Fine-tuning of Deep Networks}, 
      author={Youngmin Ro and Jin Young Choi},
      year={2021},
      eprint={2002.06048},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

Youngmin Ro and Jin Young Choi, "AutoLR: Layer-wise Pruning and Auto-tuning of Learning Rates in Fine-tuning of Deep Networks", CoRR, 2020.

autolr's Issues

Experiments Results

I tried the code you provided with default parameters for the CUB-200 dataset, but I didn't get results the same as those in the paper. In this paper, the Recall@1 Score using AutoLR method is 70.1%, but I only reached about 56%.