(optional) Install custom dependencies: swift-loader

(1) Install dependencies in requirements.txt

(2) Install this package in development mode: pip3 install -e . from the root of this package

(3) Initialize this tool by running:

mlproject init \
    --authors ${AUTHOR1_NAME,AUTHOR1_EMAIL1,AUTHOR1_EMAIL2;AUTHOR2_name,AUTHOR2_EMAIL..} \
    --company ${YOUR_DEFAULT_COMPANY} \
    --license ${YOUR_DEFAULT_LICENSE}

where:

• --authors: specifies the default authors for future projects created by the command mlproject new-project if each author is separated by colon (;), name and email is separated by comma for example --authors "First1 Last1,[email protected],[email protected];First2 Last2,[email protected],[email protected]"
• --company: specifies the name of the default company of the authors
• --license: specifies default license for future projects created by mlproject new-project. available options: proprietary, apache, mit

To create a new project template, we can run:

mlproject new-project \
    --project-name "name of this project" \
    --path "parent directory to create project under" \
    --authors "authors in this project" \
    --company "company that owns this project" \
    --license "license to use, current support proprietary/apache/mit"
    --template "template to use [currently support only generic]"

In addition, mlproject also provides convenient datastructures and classes that can be used to quickly implement new ML ideas.

When inside a project created by mlproject, we could create an empty file by using the following command:

mlproject new-file \
    --filename ${NAME_OF_THE_FILE} \
    --path ${PATH_TO_CREATE_FILE_UNDER} \
    --desc ${SHORT_DESCRIPTION_OF_THIS_FILE}

with --path is defaulted to .

The new file when created will have the same header structure as other files in the project with proper date, author names, emails, licenses, etc.

When we want to modify the metadata of a project such as the list of authors, the license or the company, we could run the following command in the main directory of a project created by ˋmlproject new-projectˋ as follows:

mlproject modify-metadata \
    --project-name "new name for this project" \
    --authors "new author list" \
    --company "new company that owns this project" \
    --license "new license"

The above command will change content of the metadata file and all the headers of all python files. The convention for --authors is similar to the project creation command above

If you're using the project template created by mlproject new-project, you'll run a single experiment configuration via the entry.py script.

What if your machine has many GPUs or CPUs and you would like to run many experiment configurations in parallel?

mlproject launch-exp is the command to use:

mlproject launch-exp \
    --entry-script "path to entry script" \
    --config-path "path to configuration file" \
    --device "either cpu or cuda" \
    --gpu-indices "the list of GPUs to use, comma separated. Default to all GPUs if device is cuda" \
    --gpu-per-exp "the number of GPUs to use for one experiment configuration" \
    --log-prefix "the prefix to dump logs from workers" \
    --nb-parallel-exp "number of parallel experiments to run. Only needed when device is cpu"

If you're using the project template created by mlproject new-project and you've run a lot of experiments, you could create a table that summarizes

the results by running the following command:

mlproject summarize-exp \
    --entry-script "path to entry script" \
    --config-path "path to configuration file" \
    --metrics "the list of metrics you want to include in the report. Comma separated"

The last switch --metrics is especially helpful if you only want to take a look at a subset of metrics.

Overall,

• mlproject.data provides abstraction for data processing. Take a look at mlproject.data for description.
• mlproject.trainer provides trainer class in pytorch. Take a look at mlproject.trainer for description.
• mlproject.metric provides abstraction for metrics, which are used in mlproject.trainer. Take a look at mlproject.metric for description.
• mlproject.loss provides abstraction for losses, which are used in mlproject.loss. Take a look at mlproject.loss for description.

It's also a good idea to take a look at the source code to get an idea about the working mechanism of different abstractions. In addition, examples under examples also serve as a good starting point

Dat Tran ([email protected])