RSPN Prototype: a simple learn & estimate test using Instacart's orders table

[kihyuk-nam]

• use a tiny fraction of InstaCart or TPC-H, TPC-DS' store_sales
• build an spn(mspn, rspn) for some data
• count query for the spn

[kihyuk-nam]

Started from this

• DataLearningTest code
• DataLoaderTest code

Support for PyTorch seems to be incomplete. (in contrast to TensorFlow/GPU)

• Failed to reproduce this

[kihyuk-nam]

ASSUMPTION:

• Skeleton only (no verification of correctness, performance)

TODO :

• [x] setup code: args, logger, launch restapi b984074681ea93e2639381e983a9746841eeb7bb (cf. TrainDBModelRunner)
• [x] traindb CLI interface (RSPN.py d47be847872a965daf36457beb0b80da76a55cb0, cf. TrainDBBaseModel, )
• [x] estimate (COUNT) a3a1071a5607cd5a810998d2431beec382d796b5
• [x] data preparation (instacart) #28 cf. tpc-ds codes of deepdb)
• [x] train a single table model #30
• [x] implement the REST API for dev (fill the gap between interface/dev/main.py and model/types/RSPN.py)

ISSUES

• [ ] location of models: directly access the 'traindb-model' (access required) vs manage them here and PR to the 'traindb-model' repo
• [ ] dual-license #29 \
• [ ] incorrect results (e.g., the confidence interval and result of COUNT) #30
• [ ] hanging threads after training #30, ee40048f007534a2a782685581f7b330132a9e5f

[kihyuk-nam]

The skeleton implementation is finished.

Next: RSPN v0 (solve critical issues and integrate this into the main traindb