Skip to content

Latest commit

 

History

History
71 lines (52 loc) · 3.72 KB

README.md

File metadata and controls

71 lines (52 loc) · 3.72 KB
Raw

SAM for multi-relation database generation

Getting Started

Datasets For multi-relation database, we conduct our experiments on IMDB. User can download the dataset by running the script

bash scripts/download_imdb.sh

Pretrained Models We have provided two pretrained model for IMDB dataset (Job-light-ranges schema).

./models/uaeq-mscn.pt: Trained from the full MSCN workload (./queries/mscn_full.csv).

./models/uaeq-mscn-400.pt: Trained from the first 400 queries in the MSCN workload (./queries/mscn_400.csv).

Database Generation To generate database from trained models using SAM, use the following commands.

python run_dbgen.py --run data-generation-job-light-mscn-worklod

By default, this generates the database using the model ./models/uaeq-mscn-400.pt. The generation process runs for 100 iterations. The generated data csv files are saved at ./generated_database/imdb.

Test the generated database To test the fidelity of generated database, import the files to a PostgreSQL database:

create table title (id int PRIMARY KEY, production_year int, kind_id int);
copy title from '/SAM/sam_multi/generated_database/imdb/title_100.csv' delimiter ',' header csv;

create table movie_keyword (movie_id int, keyword_id int);
copy movie_keyword from '/SAM/sam_multi/generated_database/imdb/movie_keyword_100.csv' delimiter ',' header csv;

create table movie_info_idx (movie_id int, info_type_id int);
copy movie_info_idx from 'SAM/sam_multi/generated_database/imdb/movie_info_idx_100.csv' delimiter ',' header csv;

create table movie_info (movie_id int, info_type_id int);
copy movie_info from '/SAM/sam_multi/generated_database/imdb/movie_info_100.csv' delimiter ',' header csv;

create table movie_companies (movie_id int, company_type_id int, company_id int);
copy movie_companies from '/SAM/sam_multi/generated_database/imdb/movie_companies_100.csv' delimiter ',' header csv;

create table cast_info (movie_id int, role_id int, person_id int);
copy cast_info from '/SAM/sam_multi/generated_database/imdb/cast_info_100.csv' delimiter ',' header csv;

Run the 400 training queries (./queries/mscn_400.sql) on the generated database and get the result Q-error:

python query_execute.py --queries ./queries/mscn_400.sql --cards ./queries/mscn_400_card.csv

Configuration for database generation

All configuration of SAM, including model training and database generation, can be set in experiments.py. To configure the generation process, locate the configure for ['data-generation-job-light-MSCN-worklod'], where you can set the autoregressive model to load, the number of iteration to run, as well as the schema of the generated database.

To reproduce our results and generate database using the pretrained model from the full MSCN workload, set the following in experiments.py

'checkpoint_to_load': 'models/uaeq-mscn.pt'
'total_iterations': 1000

*To speedup the generation process, use a larger 'save_frequency'.

The sample test set of 1000 queries can be found at ./queries/mscn_sample_1000.sql and ./queries/mscn_sample_100_card.csv

SAM model training

SAM uses UAE-Q to train a deep autoregressive model from query workloads,

To train the model from the full MSCN dataset

python run_uae.py --run job-light-ranges-mscn-workload

To test the model on sub-queries of JOB-light

python run_uae.py --run uae-job-light-ranges-reload