Table2Charts is compared with DeepEye, Data2Vis, VizML and DracoLearn. Note that the scripts mentioned below are from their respective repositories unless otherwise noted.
DeepEye only provides inference APIs. We clone DeepEye repository, and use its test.py to evaluate each table in our corpora.
We re-train and evaluate Data2Vis with Data2Vis repository. See details in Baselines/Data2Vis directory in this repository.
We re-train and evaluate VizML with their code at VizML repository.
- Extract features with
feature_extraction/extract.py. - Preprocess features with
preprocessing/preprocess.py. - With the extracted features, we train and evaluate task 8 and task 11 (encoding level Mark Type task and Is on X-axis or Y-axis task) with
neural_network/paper_tasks.py. Note that we change its Mark Type task from 3 classification to 4 classification (including line, scatter, bar, pie chart).
We clone Draco repository and evaluate Draco with Baselines/test_draco.py in this repository (which is similar to tests/test_run.py in Draco repository). In Baselines/test_draco.py, you can know how we constrain the Draco model to generate charts.