Stef-util

My utility functions to start machine learning projects

Usage

from stefutil import *
# Change those to your project
from os.path import join as os_join

BASE_PATH = '/Users/stefanh/Documents/UMich/Research/Clarity Lab/Zeroshot Text Classification'
PROJ_DIR = 'Zeroshot-Text-Classification'
PKG_NM = 'zeroshot_encoder'
DSET_DIR = 'dataset'
MODEL_DIR = 'models'

# Setup project-level functions
# Have a config `json` file ready
sconfig = SConfig(config_file=os_join(BASE_PATH, PROJ_DIR, PKG_NM, 'util', 'config.json')).__call__
_sutil = PathUtil(base_path=BASE_PATH, project_dir=PROJ_DIR, package_name=PKG_NM, dataset_dir=DSET_DIR, model_dir=MODEL_DIR)
save_fig = _sutil.save_fig
# Now you can call `sconfig` and `save_fig`

# Set argument "enum" checks
ca.cache_options('Bar Plot Orientation', attr_name='bar_orient', options=['v', 'h', 'vertical', 'horizontal'])
# Now you can call `ca` like so:
ori = 'v'
ca(bar_orient=ori)

For the nlp module which uses spaCy, run

python -m spacy download en_core_web_sm

first.

Highlights

An incomplete list of features.

Modified IceCream Debugging

IceCream debugging print with custom output width, intended for various terminal sizes.

from stefutil import sic
sic.output_width = 256 

Custom colored logger formatting

from stefutil import get_logger
logger = get_logger('<Script-Name>')

Prettier Training Log

Updates to HuggingFace transformers training & logging, including custom tqdm progress bar and TensorBoard logging.

from stefutil import MlPrettier, LogStep, s

# Before training
prettier = MlPrettier(ref=train_meta, metric_keys=['acc', 'recall', 'auc', 'ikr'])
ls = LogStep(trainer=trainer, prettier=prettier, logger=logger, file_logger=file_logger, tb_writer=tb_writer)

"""
...
"""

# During each training/eval step
for step in training_or_eval_steps:
    logs = prettier(logs)
    logger.info(s.i(logs))
    file_logger.info(s.nc(logs))