FDSR

This code rebuilds the searched architecture from FGNAS based on EDSR original code.

1. Code

Clone this repository into any place you want.

git clone https://github.com/Cheeun/FDSR.git
cd FDSR

2. Download Data

save DIV2K and Benchmark dataset from benchmark datasets (250MB) for evaluation DIV2K dataset (7.1GB) for training in directory dataset/

FDSR
|-- README.md
|-- environment.yml
`-- dataset
    |-- benchmark
    |   |-- Urban100
    |   |   |-- HR
    |   |   |-- LR_bicubic
    |   |   |-- bin
    |   |-- Set5
    |   |-- Set14
    |   |-- B100
    |   |-- bin
    |-- DIV2K

2. Conda Environment setting

conda env create -f environment.yml --name FDSR
conda activate FDSR

3. Quickstart (Demo)

Train and test the searched architecture

# baseline: Full EDSR model

# for scale 2 
cd FDSR_f_x2/src       # You are now in */FDSR/FDSR_f_x2/src
python main.py --scale 2 --searched_model fdsr_full_x2_3% # training
python main.py --scale 2 --searched_model fdsr_full_x2_3% --test_only # testing pretrained model

# for scale 4
cd FDSR_f_x4/src       # You are now in */FDSR/FDSR_f_x4/src
python main.py --scale 4 --searched_model fdsr_full_x4_3% # training
python main.py --scale 4 --searched_model fdsr_full_x4_3% --test_only # testing pretrained model

4. Settings

Place the dataset as in #2

5. Results

Further compressed architectures to be done.

Name	Baseline	Training FLOPs	Pruned-ratio	Parameters[K]	Set5	Set14	B100	Urban100	Inference time*
baseline FDSR	full EDSR x4	180G	100%	38,473	32.14	28.57	27.56	25.99	35.0
3% FDSR	full EDSR x4	6G	3.3%	1,245	32.07	28.53	27.53	25.91	0.07
3% FDSR	full EDSR x2	23G	3.3%	1,206	37.27	32.87	31.64	30.32	0.23

Inference time(sec)* is calculated for a single full HD image (1920x1080)