A web app to generate template code for machine learning
Generate custom template code for PyTorch & sklearn, using a simple web UI built with streamlit. Traingenerator offers multiple options for preprocessing, model setup, training, and visualization (using Tensorboard or comet.ml). It exports to .py, Jupyter Notebook, or Google Colab. The perfect tool to jumpstart your next machine learning project!
You can add your own template in 4 easy steps (see below), without changing any code in the app itself. Your new template will be automatically discovered by Traingenerator and shown in the sidebar. That's it!
Some ideas for new templates: Keras/Tensorflow, Pytorch Lightning, object detection, segmentation, text classification, ...
./templates. The folder name should be the task that your template solves (e.g.
Image classification). Optionally, you can add a framework name (e.g.
Image classification_PyTorch). Both names are automatically shown in the first two dropdowns in the sidebar (see image). Tip: Copy the example template to get started more quickly.
sidebar.pyto the folder (see example). It needs to contain a method
show(), which displays all template-specific streamlit components in the sidebar (i.e. everything below Task) and returns a dictionary of user inputs.
code-template.py.jinjato the folder (see example). This Jinja2 template is used to generate the code. You can write normal Python code in it and modify it (through Jinja) based on the user inputs in the sidebar (e.g. insert a parameter value from the sidebar or show different code parts based on the user's selection).
test-inputs.ymlto the folder (see example). This simple YAML file should define a few possible user inputs that can be used for testing. If you run pytest (see below), it will automatically pick up this file, render the code template with its values, and check that the generated code runs without errors. This file is optional but it's required if you want to contribute your template to this repo.
Note: You only need to install Traingenerator if you want to contribute or run it locally. If you just want to use it, go here.
git clone https://github.com/jrieke/traingenerator.git cd traingenerator pip install -r requirements.txt
Optional: For the "Open in Colab" button to work you need to set up a Github repo
where the notebook files can be stored (Colab can only open public files if
they are on Github). After setting up the repo, create a file
.env with content:
If you don't set this up, the app will still work but the "Open in Colab" button will only show an error message.
streamlit run app/main.py
Make sure to run always from the
traingenerator dir (not from the
otherwise the app will not be able to find the templates.
First, install heroku and login.
To create a new deployment, run inside
heroku create git push heroku main heroku open
To update the deployed app, commit your changes and run:
git push heroku main
Optional: If you set up a Github repo to enable the "Open in Colab" button (see above), you also need to run:
heroku config:set GITHUB_TOKEN=<your-github-access-token> heroku config:set REPO_NAME=<user/notebooks-repo>
First, install pytest and required plugins via:
pip install -r requirements-dev.txt
To run all tests:
Note that this only tests the code templates (i.e. it renders them with different input values and makes sure that the code executes without error). The streamlit app itself is not tested at the moment.
You can also test an individual template by passing the name of the template dir to
pytest ./tests --template "Image classification_scikit-learn"
The mage image used in Traingenerator is from Twitter's Twemoji library and released under Creative Commons Attribution 4.0 International Public License.