Swin-Unet

In this work swin-net transformer is used for sementic segmentation of Polyp Dataset for Computer Aided Gastrointestinal Disease Detection. The data set can be downloaded from the link given below

https://datasets.simula.no/kvasir-seg/

The rest of the repo. is divided as follows

1. Requirements
2. Data Prepration
3. Traning and Inference
4. Results

Requirements

The main library requirements for this projects are as follows

• Tensorflow
• Matplotlib
• Opencv

The complete requirements can be installed using the following set of commands

cd Swin-unet/
conda create  -n <environment -name> python==3.7.4
conda activate <environment-name>
pip install -r requirements.txt

Data Prepration

Download the data from * "https://datasets.simula.no/kvasir-seg/" and extract the data folder. Make a folder name "data" inside Swin-unet folder and copy images and mask from kvasir-seg folder inside data folder then run the following commands.

python preprocess.py

This code will preprocess data and convert it in to png mask folder with the name " masks_png".

Training and Inferene

This work genrates a swin-unet model from scratch. However, it can save checkpoint, resume traning , test on data and perform infrence on a dataset. Please follow the following set of commands

Train

python Swin_UNET_128.py  --data=<data set path>--class=< number of classes>  --inps= < Train, test, Resume>  --b_s=< batch size> --e=< Epoch> --p= < patience>  --model_dir= < Check_Point>
Example Command
python Swin_UNET_128.py  --data=./data --class=2 --inps=train --b_s=16 --e=10 --p=10 --model_dir='./checkpoint/'

Test

python Swin_UNET_128.py  --data=<data> --class=< number of classes>  --inps= < train, test, resume> --model_dir='./checkpoint/'
Example Command
python Swin_UNET_128.py  --data=./data --class=2 --inps=test --model_dir='./checkpoint/'

Resume

python Swin_UNET_128.py  --data=<data set path>--class=< number of classes>  --inps= < Train, test, Resume>  --b_s=< batch size> --e=< Epoch> --p= < patience>  --model_dir= < Check_Point>

Example Command

python Swin_UNET_128.py  --data=./data --class=2 --inps=resume --b_s=16 --e=10 --p=10 --model_dir='./checkpoint/'

Infer

python Swin_UNET_128.py  --data=< data for infrence > --class=2 --inps=< test, train, resume, infer> --model_dir=< checkp point>

python Swin_UNET_128.py  --data=./data --class=2 --inps=infer --model_dir='./checkpoint/'

Results

The prediction accuracy, latency , Flops, size on disk and the number of parameter are shown in the following table

Model	Parameters	Accuracy	Latency(sec)	Size on Disk (MB)	Flops
Swin-unet	3,783,510	0.836	0.0004	38.89	0.681 G

The visual results of this work are shown in the following figure where first column shows the ground truth sample images. The second coulmn shows their respective predictions . The third coulmn shows the error. between the predicted and the ground truth.