In this application docker is being using purely to provide a controlled and sandboxed runtime environment separate from the host. None of the classifier code, config or data are inside the container.

All the software runs from an external disk (/mnt/newarchive) which gets mounted to the container at runtime. Similarly we mount the incoming directory onto the container so that it can access the new data as they arrive on lt-qc.

Ideally run on lt-qc because it has access to /data/incoming, but the old OS causes problems building the docker image.

The Classifier Model

The classifer is based on the vgg16 CNN image classifier. It was trained on about 750 each good and bad image from late 2024 - early 2025, eyebaalled by JMM. The training was done using GPUs in the free online Kaggle environment. Only the weights from the model were exported and are reimported here to use with the stock vgg16 model that is included in python torchvision.

The state_dict from the trained classifier is > 500Mb which seemed too large to include in the github repository. The copy we are actually using is installed on lt-qc in /mnt/newarchive1/VMshare/shutterclassifier/Models. This needs to be backed up and maintained separatey. It is not in github.

Installation

On lt-qc

cd /mnt/newarchive1/Dockershare/
git clone https://github.com/LivTel/Docker_MirrorWobbleClassifier
cd Docker_MirrorWobbleClassifier
docker build -t mirrorwobbleclassifier .
docker run -id --name running_mirrorwobbleclassifier -v /mnt/newarchive1/Dockershare/Docker_MirrorWobbleClassifier/:/classifier/ -v /data/incoming:/data/incoming  mirrorwobbleclassifier

The container is now running and available for use.

Note that when the container was started, two directories from the host were mounted onto the container using -v

• The repository directory itself (/mnt/newarchive1/Dockershare/Docker_MirrorWobbleClassifier) is available in the container as /classifier
• The /data/incoming directory is available in the container as /data/incoming

Data Preparation - NEEDS UPDATE

The classifier was trained on 224x224 pixel FITS images created by 9x9 binning normal 2x2 binned IO:O images. Classification will be best if you feed it with similar data. It can work on larger FITS images but will not be as accurate and, for example, is likely to be hopeless if you feed it full size 2048x2048 images.

Best to shrink the FITS images down first in some external application.

Included here is trivial python script that will boxcar median bin a FITS. Typically IO:O images are preprocessed by

python3 imbin.py -b 9 h_e_99999999_1_1_1_1.fits

Running Classification

docker exec -it running_mirrorwobbleclassifier python3 /classifier/app_ml.py /data/incoming/h_e_99999999_1_1_1_1_bin9.fits

Running From Crontab

Nothing to do with this classifier per se, but careful calling dockers from within cron scripts. As with many other cron script you can have problems with the runtime environment. Do not use the -t pseudo-tty option.

docker exec -i running_mirrorwobbleclassifier python3 /classifier/app_ml.py /data/incoming/h_e_99999999_1_1_1_1_bin9.fits