This project is used for experiment to train and test models on various datasets. The core function has been packaged as "differnet-zerobox". Pleaser refer to the readme for how to use the package.
If you need more information about DifferNet, please reference to the official repository.
Differnet Officical repository The official repository to the WACV 2021 paper "Same Same But DifferNet: Semi-Supervised Defect Detection with Normalizing Flows" by Marco Rudolph, Bastian Wandt and Bodo Rosenhahn.
The project has been upgraded to python 3.9. Please setup python 3.9 virtual environment then do the following.
In order to have proper torch and torch vision to use either GPU or CPU please follow pytorch.org to install torch and torch vision
$ pip install -r requirements.txt
All configuration has default values from package differnet(from package differnet-zerobox). The project can have dict based configuration to overwrite any default value.
Common settings
#conf/settings.conf
"differnet_work_dir": "./work", #work folder
"device": "cuda", # cuda or cpu
"device_id": 0, # the device you want to use. depends on how many GPU or CPU you have.
"verbose": True, # Set to true, when you do experiments.
"meta_epochs": 10, # traing loop
"sub_epochs": 8, # sub-loop of traing
"test_anormaly_target": 10, # threshold when run testing model to identify if a given image is good or bad
Traing
python training.py
Run test cases
python -m pytest -s
The data structure under work folder looks like this. The model folder will save trained model. For experiment purpose, you would like to give test and validate folder with proper labled data. While, for zerobox it only requires train folder and data. The minimum images is 16 based on the differnet paper.
pink1/
├── model
├── test
│ ├── defect
│ └── good
├─── validate
│ ├── defect
│ └── good
└── train
└── good
├── 01.jpg
├── 02.jpg
├── 03.jpg
├── 04.jpg
├── 05.jpg
├── 06.jpg
├── 07.jpg
├── 08.jpg
├── 09.jpg
├── 10.jpg
├── 11.jpg
├── 12.jpg
├── 13.jpg
├── 14.jpg
├── 15.jpg
└── 16.jpg