gigit0000/uspto-4th-place-solution

This is the 4th Place Solution for the USPTO - Explainable AI for Patent Professionals (Kaggle competition).

• detail document: https://www.kaggle.com/competitions/uspto-explainable-ai/discussion/522200
• submission code (with magic): https://www.kaggle.com/code/ryotayoshinobu/uspto-4th-place-solution-w-magic-lb0-98
• submission code (without magic): https://www.kaggle.com/code/ryotayoshinobu/uspto-4th-place-solution-w-o-magic-lb0-91

How to Reproduce (for Competition Organizers)

Hardware

• CPU: Intel Core i9 13900KF (24 cores, 32 threads)
• GPU: NVIDIA GeForce RTX 4090
• RAM: 64GB

OS/platform

• WSL2 (version 2.0.9.0, Ubuntu 22.04.2 LTS)

3rd-party software

Please check the dockerfile in /kaggle/.devcontainer

Preprocess

Solution with magic

1. put the competition data in input/uspto-boolean-search-optimization
   • ex. input/uspto-boolean-search-optimization/test.csv
2. put the whoosh wheel in input/whoosh-wheel-2-7-4
   • (input/whoosh-wheel-2-7-4/Whoosh-2.7.4-py2.py3-none-any.whl)
3. save tokenized text data
   • input/all-index/gen.ipynb
   • input/all-index-per-patent/split.ipynb
     • Too many files may cause errors along the way. In that case, run it again (the rest will be saved in a separate directory).
   • input/all-index-per-patent/split/gen_db.ipynb
4. count cpc frequency
   • input/cpc-counts/cpc_freq.ipynb
5. count token frequency
   • input/token-counts/word_freq.ipynb
6. generate DB for mapping cpc to patents
   • input/cpc2patents/save_db.ipynb
7. generate DB for mapping rare_token to patents
   • input/rare-tokens/gen.ipynb
   • input/rare-tokens/save_db.ipynb
8. generate DB for mapping patent to cpcs
   • input/patent2cpc/save_db.ipynb

Only Solution without magic

1. complete all steps above (1.~6.)
2. generate .json for mapping cpc to patents
   • input/cpc-mapping/cpc_mapping.ipynb
3. generate DB for (cpc, token) to patents
   • input/preprocess-complete/save_cpc_token_lists.ipynb
   • input/preprocess-complete/save_patent_wise.ipynb
   • input/preprocess-complete/split/gen_leveldb.ipynb
   • input/preprocess-complete-v2/save_cpc_token_lists.ipynb
   • input/preprocess-complete-v2/save_patent_wise.ipynb
   • input/preprocess-complete-v2/split/gen_leveldb.ipynb
   • this step will take at least 2 weeks in my environment.
   • save_cpc_token_lists may stop due to OOM or other reasons. In that case, please execute again. (It will resume from where it left off.)
4. generate DB for publication_number (row id) to token
   • input/preprocess-all-token-single/save.ipynb
   • input/preprocess-all-token-single/gen_leveldb.ipynb

NOTE:

• To avoid OOM, free memory when each notebook is finished executing.
• The submission notebooks will delete all data in /kaggle/working when they are executed.
• Sufficient disk space is required.
  • At least 2TB is required.
  • If you do not have enough space, please delete unnecessary intermediate files as appropriate.

Supplemental Information for Competition Organizers

• Dockerfile is used instead of B4.requirements.txt.
• src/config.yaml is used instead of B6. SETTINGS.json.
• There are no B7. Serialized copy of the trained model.
• B8. entry_points.md is not included because my all codes are .ipynb format.