Skip to content

LearnCV/mle-logging

ย 
ย 

Folders and files

NameName
Last commit message
Last commit date

Latest commit

ย 

History

55 Commits
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 

Repository files navigation

A Lightweight Logger for ML Experiments ๐Ÿ“–

Pyversions PyPI version Code style: black codecov Colab

Simple logging of statistics, model checkpoints, plots and other objects for your Machine Learning Experiments (MLE). Furthermore, the MLELogger comes with smooth multi-seed result aggregation and combination of multi-configuration runs. For a quickstart check out the notebook blog ๐Ÿš€

The API ๐ŸŽฎ

from mle_logging import MLELogger

# Instantiate logging to experiment_dir
log = MLELogger(time_to_track=['num_updates', 'num_epochs'],
                what_to_track=['train_loss', 'test_loss'],
                experiment_dir="experiment_dir/",
                model_type='torch')

time_tic = {'num_updates': 10, 'num_epochs': 1}
stats_tic = {'train_loss': 0.1234, 'test_loss': 0.1235}

# Update the log with collected data & save it to .hdf5
log.update(time_tic, stats_tic)
log.save()

You can also log model checkpoints, matplotlib figures and other .pkl compatible objects.

# Save a model (torch, tensorflow, sklearn, jax, numpy)
import torchvision.models as models
model = models.resnet18()
log.save_model(model)

# Save a matplotlib figure as .png
fig, ax = plt.subplots()
log.save_plot(fig)

# You can also save (somewhat) arbitrary objects .pkl
some_dict = {"hi" : "there"}
log.save_extra(some_dict)

Or do everything in a single line...

log.update(time_tic, stats_tic, model, fig, extra, save=True)

File Structure & Re-Loading ๐Ÿ“š

The MLELogger will create a nested directory, which looks as follows:

experiment_dir
โ”œโ”€โ”€ extra: Stores saved .pkl object files
โ”œโ”€โ”€ figures: Stores saved .png figures
โ”œโ”€โ”€ logs: Stores .hdf5 log files (meta, stats, time)
โ”œโ”€โ”€ models: Stores different model checkpoints
    โ”œโ”€โ”€ init: Stores initial checkpoint
    โ”œโ”€โ”€ final: Stores most recent checkpoint
    โ”œโ”€โ”€ every_k: Stores every k-th checkpoint provided in update
    โ”œโ”€โ”€ top_k: Stores portfolio of top-k checkpoints based on performance
โ”œโ”€โ”€ tboards: Stores tensorboards for model checkpointing
โ”œโ”€โ”€ <config_name>.json: Copy of configuration file (if provided)

For visualization and post-processing load the results via

from mle_logging import load_log
log_out = load_log("experiment_dir/")

# The results can be accessed via meta, stats and time keys
# >>> log_out.meta.keys()
# odict_keys(['experiment_dir', 'extra_storage_paths', 'fig_storage_paths', 'log_paths', 'model_ckpt', 'model_type'])
# >>> log_out.stats.keys()
# odict_keys(['test_loss', 'train_loss'])
# >>> log_out.time.keys()
# odict_keys(['time', 'num_epochs', 'num_updates', 'time_elapsed'])

If an experiment was aborted, you can reload and continue the previous run via the reload=True option:

log = MLELogger(time_to_track=['num_updates', 'num_epochs'],
                what_to_track=['train_loss', 'test_loss'],
                experiment_dir="experiment_dir/",
                model_type='torch',
                reload=True)

Installation โณ

A PyPI installation is available via:

pip install mle-logging

If you want to get the most recent commit, please install directly from the repository:

pip install git+https://github.com/mle-infrastructure/mle-logging.git@main

Advanced Options ๐Ÿšด

Merging Multiple Logs ๐Ÿ‘ซ

Merging Multiple Random Seeds ๐ŸŒฑ + ๐ŸŒฑ

from mle_logging import merge_seed_logs
merge_seed_logs("multi_seed.hdf", "experiment_dir/")
log_out = load_log("experiment_dir/")
# >>> log.eval_ids
# ['seed_1', 'seed_2']

Merging Multiple Configurations ๐Ÿ”– + ๐Ÿ”–

from mle_logging import merge_config_logs, load_meta_log
merge_config_logs(experiment_dir="experiment_dir/",
                  all_run_ids=["config_1", "config_2"])
meta_log = load_meta_log("multi_config_dir/meta_log.hdf5")
# >>> log.eval_ids
# ['config_2', 'config_1']
# >>> meta_log.config_1.stats.test_loss.keys()
# odict_keys(['mean', 'std', 'p50', 'p10', 'p25', 'p75', 'p90']))

Plotting of Logs ๐Ÿง‘โ€๐ŸŽจ

meta_log = load_meta_log("multi_config_dir/meta_log.hdf5")
meta_log.plot("train_loss", "num_updates")

Storing Checkpoint Portfolios ๐Ÿ“‚

Logging every k-th checkpoint update โ— โฉ ... โฉ โ—

# Save every second checkpoint provided in log.update (stored in models/every_k)
log = MLELogger(time_to_track=['num_updates', 'num_epochs'],
                what_to_track=['train_loss', 'test_loss'],
                experiment_dir='every_k_dir/',
                model_type='torch',
                ckpt_time_to_track='num_updates',
                save_every_k_ckpt=2)

Logging top-k checkpoints based on metric ๐Ÿ”ฑ

# Save top-3 checkpoints provided in log.update (stored in models/top_k)
# Based on minimizing the test_loss metric
log = MLELogger(time_to_track=['num_updates', 'num_epochs'],
                what_to_track=['train_loss', 'test_loss'],
                experiment_dir="top_k_dir/",
                model_type='torch',
                ckpt_time_to_track='num_updates',
                save_top_k_ckpt=3,
                top_k_metric_name="test_loss",
                top_k_minimize_metric=True)

Citing the MLE-Infrastructure โœ๏ธ

If you use mle-logging in your research, please cite it as follows:

@software{mle_infrastructure2021github,
  author = {Robert Tjarko Lange},
  title = {{MLE-Infrastructure}: A Set of Lightweight Tools for Distributed Machine Learning Experimentation},
  url = {http://github.com/mle-infrastructure},
  year = {2021},
}

Development ๐Ÿ‘ท

You can run the test suite via python -m pytest -vv tests/. If you find a bug or are missing your favourite feature, feel free to create an issue and/or start contributing ๐Ÿค—.

Packages

No packages published

Languages

  • Python 100.0%