Guild AI  streamline your TensorFlow and Keras development

Guild AI automates TensorFlow and Keras deep learning workflow, letting you focus on optimizing your models and getting them into production as quickly as possible. The command line toolset, supports your deep learning work end-to-end — from model acquisition, through training and testing, to deployment. You can even publish your models for your collaborators to use!


Automatically save each training run for analysis and deployment.

Track experiments


Compare performance across runs including accuracy, loss, and hyperparameters.

Compare runs


Deploy your models to Cloud ML or serve them locally as a REST API.

Deploy to Cloud ML


Visualize results across multiple runs with TensorBoard and Guild View.



Package and distribute your models for others to use and learn from.

Package models


Explore state-of-the-art models in Guild AI's ever-growing ecosystem.

Platform support

Quick start

Step 1. Install Guild AI

Guild AI is installed using pip. Select one of the installation methods below.

Install Guild AI using pip by running:

            pip install guildai


If you are unable to install Guild AI due to permission errors, you may need to run the command using sudo:

              sudo pip install guildai

Install Guild AI in a virtual environment named guild by running:

              virtualenv guild
              . guild/bin/activate
              pip install guildai

To install in a different virtual env, replace guild above with the alternate location.

Install Guild AI in a Conda environment named guild by running:

              conda create -n guild python=3.6
              source activate guild
              pip install guildai

To install in a different virtual env, replace guild above with the alternate location.

For more information, see Installing Guild AI.

When Guild AI is installed, initialize the Guild environment by running:

      guild init --env

This step verifies that your environment is setup correctly and prompts you to install TensorFlow if it isn't already installed.

If you encounter errors at this stage, see Troubleshooting for more information or open an issue on GitHub to get help.

Step 2. Find and install models

Guild AI lets you find and install models in seconds. Search for mnist by running:

      guild search mnist

For this quick start, we'll work with the base mnist package. Install it by running:

      guild install mnist

List the models you just installed by running:

      guild models

Step 3. Train the models

In this step we'll train two models: mnist-softmax and mnist-cnn.

First, train the softmax version by running:

      guild train mnist-softmax

Review the default values and press ENTER. The mnist-softmax model trains quickly even on systems that don't have a GPU!

Next we'll train the CNN. Choose a method based on your system type.

Most GPU accelerated systems can train the CNN model in a minute or two. If your system system has a GPU, train mnist-cnn for the default number of epochs by running:

            guild train mnist-cnn

If you system doesn't have a GPU, the mnist-cnn model will train slowly. You can complete this step faster by training fewer epochs. Train for one epoch by running:

            guild train mnist-cnn epochs=1

When both models are trained, view the list of runs by running:

      guild runs

Step 4. Compare model performance

In this step we'll view the training results. Open a separate command line console and run:

      guild view

Guild View is a visual application that lets you explore runs, compare model performance, and view generated files. Guild View opens automatically in your browser when run the command.

Guild View

Step 4.1. Compare runs in Guild View

In your browser, click Compare runs in the left sidebar. This displays a table containing the results of your two runs.

Compare runs table

Use this view to find the run with the best performance. In this case, it's the CNN, and by quite a margin!

Step 4.2. Compare runs in TensorBoard

In your browser, click View in TensorBoard in the left sidebar. This opens another tab running TensorBoard, which lets you view detailed training data. Use the tabs at the top of TensorBoard to view different types of data.


Step 5. Serve locally

In the previous step, we saw that the CNN model performs much better than the softmax! Let's serve that model locally as a REST prediction API.

In a new command line console, serve the CNN model by running:

      guild serve -o mnist-cnn --host localhost --port 8083

This command opens a new browser window for Guild Serve, which describes the REST endpoint for the trained MNIST CNN model. You can build your application and test locally before deploying to a production environment.

Guild Serve
Guild Serve

Step 6. Deploy to Cloud ML


This step requires a Google Cloud Machine Learning Engine account. To setup your account and environment, follow the steps in Cloud ML Engine - Getting Started

When you're ready to run your prediction service in production, you can deploy to Cloud ML by running:

      guild run mnist-cnn:cloudml-deploy bucket=$BUCKET_NAME

$BUCKET_NAME must refer to a Google Cloud Storage bucket that you have write permission to.

For a deep dive into Guild AI's Cloud ML support, see Train and predict with Cloud ML.

Guild View

Next steps

Go deeper with Guild AI

Go deeper into the Quick Start material above with a step-by-step tutorial on training and deploying with Guild AI and Cloud ML.

Go deeper with Guild AI

Explore models

Guild AI supports an ever-growing ecosystem of TensorFlow and Keras models that you can install and train with a few simple commands.

Discover the models

Browse the docs

If you're interested in a complete picture of Guild AI, start by browsing its comprehensives documentation.

Browse the docs