New services should be deployed to tower01 if possible since this is the only machine with GPU Support.
We use a new model: https://github.com/facebookresearch/deit/
All datasets will be stored at sg01:/var/www/picalike.corpex-kunden.de/sg01/htdocs/ml_data The folder structure is as follows $root/$service_name/$date
The content of the folder MUST contain:
A file with one URL per line for each class
A sqlite with the image data (3,224,224) as a backup.
sg01:/var/www/picalike.corpex-kunden.de/sg01/htdocs/ml_data/size_table
sg01:/var/www/picalike.corpex-kunden.de/sg01/htdocs/ml_data/person
sg01:/var/www/picalike.corpex-kunden.de/sg01/htdocs/ml_data/freigestellt
A task is to update the OS of the machine to allow Docker + cuda + Pytorch. But since this is a bigger task, new v5 services will run in CPU-only mode for now.
A service should provide an endpoint (POST). The input is a list of URLs.
In case of a 2-class service, the result should be a dictionary with URL→{True,False}.