- Write predict_one for CLI use
- Test different postprocessing methods
- Data augmentation
- Try non-templated commands
- Beam search with custom eval?
- Different metrics: https://huggingface.co/docs/datasets/metrics
-
notebooks
:data_preprocess.ipynb
: adds additional data sourcesgraphs.ipynb
: visualize results from experimentsold
: directory of other ipynb that we're ignoring...
src
:config.py
: configuration used bydata_utils.py
: reading/saving/context/encode/decode from competitiondataset.py
: subclassing torch datasets; LBL vs Blocked dataset?diverse_beam_search.py
: expanding hugging face beam searchgenerate.py
: prediction and scoring utilitiesmodified_beam_search.py
: another modified hugging face beam searchonnx.py
: conversion to/from onnx ML formatpreprocess.py
: preprocesses datarun.py
: main loop for trainingtrainer.py
: overrides hugging face trainer to override some optionstune.py
: contains a list of experiments by modifying config
webapp
demo_app.py
: starts a flask servereval.py
: used for predictions using the model on hugging facerequirements.txt
: python requirementstrain.py
: runsexperiments()
fromsrc/tune.py