Ecosyste.ms: Awesome
An open API service indexing awesome lists of open source software.
https://github.com/mle-infrastructure/mle-logging
Lightweight ML Experiment Logging 📖
https://github.com/mle-infrastructure/mle-logging
Last synced: 2 months ago
JSON representation
Lightweight ML Experiment Logging 📖
- Host: GitHub
- URL: https://github.com/mle-infrastructure/mle-logging
- Owner: mle-infrastructure
- License: mit
- Created: 2021-08-03T13:27:02.000Z (almost 3 years ago)
- Default Branch: main
- Last Pushed: 2023-03-07T16:37:05.000Z (over 1 year ago)
- Last Synced: 2024-01-25T21:37:34.983Z (5 months ago)
- Language: Python
- Homepage: https://mle-infrastructure.github.io/mle_logging
- Size: 9.4 MB
- Stars: 77
- Watchers: 1
- Forks: 6
- Open Issues: 1
-
Metadata Files:
- Readme: README.md
- Changelog: CHANGELOG.md
- Contributing: CONTRIBUTING.md
- License: LICENSE
Lists
- awesome-stars - mle-infrastructure/mle-logging - A Lightweight ML Experiment Logging Tool 📖 (Python)
- awesome-stars - mle-infrastructure/mle-logging - Lightweight ML Experiment Logging 📖 (Python)
README
# A Lightweight Logger for ML Experiments 📖
[](https://pypi.python.org/pypi/mle-logging)
[](https://badge.fury.io/py/mle-logging)
[](https://github.com/psf/black)
[](https://codecov.io/gh/mle-infrastructure/mle-logging)
[](https://colab.research.google.com/github/mle-infrastructure/mle-logging/blob/main/examples/getting_started.ipynb)
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](https://github.com/mle-infrastructure/mle-logging/blob/main/examples/getting_started.ipynb) 🚀
## The API 🎮
```python
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.
```python
# 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...
```python
log.update(time_tic, stats_tic, model, fig, extra, save=True)
```
### File Structure & Re-Loading 📚
The `MLELogger` will create a nested directory, which looks as follows:
```
experiment_dir
├── extra: Stores saved .pkl object files
├── figures: Stores saved .png figures
├── logs: Stores .hdf5 log files (meta, stats, time)
├── models: Stores different model checkpoints
├── init: Stores initial checkpoint
├── final: Stores most recent checkpoint
├── every_k: Stores every k-th checkpoint provided in update
├── top_k: Stores portfolio of top-k checkpoints based on performance
├── tboards: Stores tensorboards for model checkpointing
├── .json: Copy of configuration file (if provided)
```
For visualization and post-processing load the results via
```python
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:
```python
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 ⏳
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 🚴
### Merging Multiple Logs 👫
**Merging Multiple Random Seeds** 🌱 + 🌱
```python
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** 🔖 + 🔖
```python
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 🧑🎨
```python
meta_log = load_meta_log("multi_config_dir/meta_log.hdf5")
meta_log.plot("train_loss", "num_updates")
```
### Storing Checkpoint Portfolios 📂
**Logging every k-th checkpoint update** ❗ ⏩ ... ⏩ ❗
```python
# 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** 🔱
```python
# 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)
```
### Weights&Biases Backend Integration 🧑🎨
You can also use W&B as a backend for logging. All results are stored as before but additionally we report to the W&B server:
```python
# Provide all configuration details as option
log = MLELogger(time_to_track=['num_updates', 'num_epochs'],
what_to_track=['train_loss', 'test_loss'],
use_wandb=True,
wandb_config={
"key": "sadfasd", # Only needed if not logged in
"entity": "roberttlange", # Only needed if not logged in
"project": "some-project-name",
"group": "some-group-name"
})
```
### Citing the MLE-Infrastructure ✏️
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 👷
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](CONTRIBUTING.md) 🤗.