# Welcome to HPSearch
> Experiment tracking framework.
`HPSearch` is a flexible experiment tracking framework that provides features such as a simple and powerful query mechanism for studying and visualizing the performance of past experiments meeting given criteria, the possibility to resume a previous experiment with or without modifying the original hyper-parameters (e.g., extending the number of epochs of a promising past experiment, or gradually changing the hyper-parameters to obtain an approximate curriculum learning type of approach, etc.), the capability of visualizing the learning evolution of multiple experiments based on different metrics, and many other features.
## Key features
`HPSearch` provides, among others, the following features:
- Query the performance of past experiments meeting desired criteria. This can be done either from command line with a simple command, or programmatically. Queried experiments are shown as a table sorted by performance, and visualized in graphical form, comparing the evolution of the metrics of the selected experiments during training.
- Visualize the evolution of current experiments and compare them against previous ones using multiple metrics
- Resume previous experiments with or without modifying their original hyper-parameters. This can be applied, for instance, when we start by performing a quick exploration of hyper-parameters by allocating a small time budget for each experiment. Once this is done, we may want to increase the exploration of a subset of experiments that were promising, e.g., by extending their number of epochs, or using curriculum learning to gradually changing their hyper-parameters across epochs. We may also want to change only to explore hyper-parameters affecting the final part of a pipeline, where the first steps are dedicated to pre-processing and normalization techniques that might be computationally expensive and which we want to reuse. With `HPSearch`, it is easy to do that, and annotate the fact that this was done when characterizing the new experiments.
- Good support for default values. When introducing a new hyper-parameter, all previous experiments are automatically assigned a default value for such parameter. This makes it easy to avoid repeating previous experiments in the case when the default value is one of the possible values to be explored.
- High level of decoupling between the experiment tracking code and the model code.
## Installation
`HPSearch` can be installed using pip:
```bash
pip install hpsearch
```
## Documentation
Documentation can be found [here](https://jaume-jci.github.io/hpsearch/)
Raw data
{
"_id": null,
"home_page": "https://github.com/Jaume-JCI/hpsearch/tree/master/",
"name": "hpsearch",
"maintainer": "",
"docs_url": null,
"requires_python": ">=3.6",
"maintainer_email": "",
"keywords": "some keywords",
"author": "Jaume Amores",
"author_email": "jamorej@jci.com",
"download_url": "https://files.pythonhosted.org/packages/68/5e/0b7606b7977f6cb2613b58ea7c7fe0b692a6013665a62019dcac61d0b472/hpsearch-0.0.7.tar.gz",
"platform": null,
"description": "# Welcome to HPSearch\n> Experiment tracking framework.\n\n\n`HPSearch` is a flexible experiment tracking framework that provides features such as a simple and powerful query mechanism for studying and visualizing the performance of past experiments meeting given criteria, the possibility to resume a previous experiment with or without modifying the original hyper-parameters (e.g., extending the number of epochs of a promising past experiment, or gradually changing the hyper-parameters to obtain an approximate curriculum learning type of approach, etc.), the capability of visualizing the learning evolution of multiple experiments based on different metrics, and many other features.\n\n## Key features\n\n`HPSearch` provides, among others, the following features:\n\n- Query the performance of past experiments meeting desired criteria. This can be done either from command line with a simple command, or programmatically. Queried experiments are shown as a table sorted by performance, and visualized in graphical form, comparing the evolution of the metrics of the selected experiments during training.\n- Visualize the evolution of current experiments and compare them against previous ones using multiple metrics\n- Resume previous experiments with or without modifying their original hyper-parameters. This can be applied, for instance, when we start by performing a quick exploration of hyper-parameters by allocating a small time budget for each experiment. Once this is done, we may want to increase the exploration of a subset of experiments that were promising, e.g., by extending their number of epochs, or using curriculum learning to gradually changing their hyper-parameters across epochs. We may also want to change only to explore hyper-parameters affecting the final part of a pipeline, where the first steps are dedicated to pre-processing and normalization techniques that might be computationally expensive and which we want to reuse. With `HPSearch`, it is easy to do that, and annotate the fact that this was done when characterizing the new experiments.\n- Good support for default values. When introducing a new hyper-parameter, all previous experiments are automatically assigned a default value for such parameter. This makes it easy to avoid repeating previous experiments in the case when the default value is one of the possible values to be explored.\n- High level of decoupling between the experiment tracking code and the model code.\n\n## Installation\n\n`HPSearch` can be installed using pip:\n\n```bash\npip install hpsearch\n```\n\n## Documentation \n\nDocumentation can be found [here](https://jaume-jci.github.io/hpsearch/) \n",
"bugtrack_url": null,
"license": "MIT License",
"summary": "A description of your project",
"version": "0.0.7",
"split_keywords": [
"some",
"keywords"
],
"urls": [
{
"comment_text": "",
"digests": {
"md5": "f92af6680364bf8c2662dd6fe279cb05",
"sha256": "d31683bf06c29a41bc53f2db1c4f002d28692214c67c5f3a754a13d316b7d104"
},
"downloads": -1,
"filename": "hpsearch-0.0.7-py3-none-any.whl",
"has_sig": false,
"md5_digest": "f92af6680364bf8c2662dd6fe279cb05",
"packagetype": "bdist_wheel",
"python_version": "py3",
"requires_python": ">=3.6",
"size": 99790,
"upload_time": "2022-12-02T17:24:53",
"upload_time_iso_8601": "2022-12-02T17:24:53.323344Z",
"url": "https://files.pythonhosted.org/packages/d5/cd/d18bc87eeec5928050a619c423d68f9554d204956bb93fbf325a10086ffb/hpsearch-0.0.7-py3-none-any.whl",
"yanked": false,
"yanked_reason": null
},
{
"comment_text": "",
"digests": {
"md5": "e7f5472be14ff37a98986ff402085605",
"sha256": "ec70653a63a8e52baf0ce0aacbce68c6c84952828c9ef0bc2b1472a12c7964c9"
},
"downloads": -1,
"filename": "hpsearch-0.0.7.tar.gz",
"has_sig": false,
"md5_digest": "e7f5472be14ff37a98986ff402085605",
"packagetype": "sdist",
"python_version": "source",
"requires_python": ">=3.6",
"size": 73312,
"upload_time": "2022-12-02T17:24:56",
"upload_time_iso_8601": "2022-12-02T17:24:56.232479Z",
"url": "https://files.pythonhosted.org/packages/68/5e/0b7606b7977f6cb2613b58ea7c7fe0b692a6013665a62019dcac61d0b472/hpsearch-0.0.7.tar.gz",
"yanked": false,
"yanked_reason": null
}
],
"upload_time": "2022-12-02 17:24:56",
"github": false,
"gitlab": false,
"bitbucket": false,
"lcname": "hpsearch"
}
JSON