classification

Learning in the Frequency Domain

Classification on ImageNet with PyTorch.

Prerequisites

• PyTorch compatible GPU
• Python 3.7
• PyTorch >= 1.2.0
• opencv-python 4.1.1
• libjpeg-turbo 2.0.3
• jpeg2dct

Install

• Install PyTorch

• Clone this repo recursively

  git clone --recursive https://github.com/calmevtime1990/supp
  

• Install required packages

  pip install -r requirements.txt
  

• Install libjpeg-turbo

  bash install_libjpegturbo.sh
  

• Download pretrained models and extract to pretrained. The folder structure should look like this:

pretrained
├── resnet50dct_upscaled_static_24
│   ├── log.txt
│   └── model_best.pth.tar
└── resnet50dct_upscaled_static_64
    ├── log.txt
    └── model_best.pth.tar

• Prepare datasets It is recommended to symlink the dataset root to data. The folder structure should look like this:

data
├── train
├── val
└── README.md

Evaluation

Run resnet_upscaled_static.sh to start testing. Change the --data $imagenet_dir to the location of the ImageNet dataset.

Testing the proposed model with 24 channels

bash scripts/resnet_upscaled_static.sh 24

Testing the proposed model with 64 channels

bash scripts/resnet_upscaled_static.sh 64

Results

Performance of the proposed model - ResNet-50

ResNet-50	#Channels	Size Per Channel	Top-1	Top-5	Normalized Input Size
RGB	3	224x224	75.780	92.650	1.0
DCT-24 (ours)	24	56x56	76.792	93.254	0.5
DCT-64 (ours)	64	56x56	77.160	93.474	1.3

Performance of the proposed model - MobileNetV2

MobileNetV2	#Channels	Size Per Channel	Top-1	Top-5
RGB	3	224x224	71.702	90.415
DCT-24 (ours)	24	112x112	72.364	90.606
DCT-32 (ours)	32	112x112	72.282	90.592