Please follow the instructions from Neurocard project to download IMDB dataset in ./datasets.
python run_uae.py --run job-light-ranges-mscn-workload
Instruction for model testing of the pretrained UAE (with job-light-ranges basic settings) on subqueries of JOB-light
python run_uae.py --run uae-job-light-ranges-reload
If you want to evaluate UAE on the original JOB-light queries, simply change the value of 'job_light_queries_csv' at line 627 in experiments.py to './queries/job-light.csv'.
train_virtual_cols: True: allow the virtual columns to be involved in query learning. False: fix the virtual columns during query learning.
run_uaeq: True: train query-driven uae-q from queries. False: train uae from both data and queries.
To change the the configuration for the run, please modify the corresponding section in experiments.py.
The current UAE model was trained using the column settings of job-light-ranges, and thus the hyper-parameters were tuned accordingly. If you want to train UAE mode using other column settings (e.g., job-light), please carefully tune the hyper-parameters (especially q_weight and warmups for UAE).
This code is based on Neurocard. Thanks to the contributors of Neurocard.