Guild Files


Guild files are named guild.yml and are located in project directories. They provide information about your project.

  • Scripts used to generate experiments
  • User input parameters
  • Generated metrics
  • Script source code
  • Requires input files

While Guild can run scripts directly without explicit configuration, in such cases Guild makes assumptions about how to run each script. For all but simple cases, we recommend using Guild files to formally define your project operations.

More about Guild files:


Guild files are plain text files in YAML format. See Guild File Reference for details on file format.


An operation defines what Guild executes to for a run.

Consider this example, which defines a single operation named train:

  description: Train a model using a Python script
  main: train
    learning-rate: 0.1
    batch-size: 100

The operation is named train and can be run using guild run train. It runs the train Python module, which is specified by the main attribute. The operataion defines two flags: learning-rate and batch-size.

You can run the operation from a command terminal by changing to the directory containing guild.yml (the project directory) and running:

guild run train
You are about to run train
  batch-size: 100
  learning-rate: 0.1
Continue? (Y/n)

Guild shows a preview of the flags used for the operation and asks you to confirm the operation by pressing Enter. When you confirm the operation, Guild executes the train module with the specified flag values. Guild generates a run, which is a record of the operation inputs and outputs.

Guild passes flag values to Python modules by setting global variables or by passing arguments on the command line. You can configure this interface or Guild can detect it. For more information, see Flags Interface below.

Guild supports operations in Python as well as other languages. Here’s an operation that runs a shell script:

  description: Train a model using a shell script

For more information on running operations with difference languages, see Other Language Operations below.

Python Operation

Guild provides special support for Python-based operations. To define a Python based operation, use the main operation attribute to specify the Python main module. This is a Python module that runs a task when loaded by the Python interpreter as __main__.

Consider a script named

from models import cnn

def train():
    model = cnn.CNN()

if __name__ == "__main__":

To run the script using Python, you use:


In this case, the main module name is train_classifier and is specified in a Guild file operation as follows:

  main: train_classifier

Note Do not include the file name extension when specifying a main module for an operation. The attribute value specifies a Python module and not a file name.

Other Language Operations

To run a non-Python based operation, use the exec operation attribute. The value for exec is a command available on the PATH environment variable or a path to an executable program.

The following example runs a Java program, provided as a JAR file:

  exec: java -jar train.jar
    - file: train.jar

Any files needed by the operation — e.g. programs, etc. — must be specified as dependencies using the requires attribute. Refer to Dependencies below for information on specifying required files for an operation.


Flags are user inputs to an operation. Flags define model and training hyperparameters as well other script inputs, such as data set information, user defined input paths, deployment endpoints, etc.

Flags are defined for each operation using the flags attribute.

    learning-rate: 0.1
    batch-size: 100

Use flags to define operation inputs such as learning rate and batch size

When running an operation, a user sets flag values using FLAG_NAME=VALUE arguments to guild run.

guild run train learning-rate=0.01 batch-size=1000

Specify flag values as FLAG_NAME=VALUE arguments

See Flags Interface below for information on how Guild conveys flag values to a script.

Guild records flag values used for each run. Flag values are displayed in several contexts:

Flags Interface

Guild conveys flag values to a script using various methods:

  • Command line arguments
  • Environment variables
  • Global variables (Python only)

For Python based operations, Guild detects the flags interface by inspecting the main module. If the module uses Python’s argparse package, Guild assumes that the script uses command line arguments to read flag values. Otherwise, Guild assumes the script uses global variables for flags.

Specify the interface using the flags-dest operation attribute (short for flags destination).

When flags-dest is set, Guild does not inspect the file to detect the flags interface.

Flags as Command Line Arguments

To indicate that flags should be passed as command line arguments use args:

    learning-rate: 0.1
    batch-size: 100
  flags-dest: args

Flags conveyed to a script using command line arguments

In this case, Guild runs the command python -m train --learning-rate 0.1 --batch-size 100. The script must parse these command lines to read the specified flag values.

By default Guild uses the flag name as the argument name. To use a different value, specify the arg-name flag attribute.

Flags as Global Variables (Python only)

When flags-dest is globals, Guild sets flag values as script global variables.

Automatically Import Flags (Python only)

Guild can import flags from Python scripts to avoid duplicating information in a Guild file. By default, Guild does not attempt to import flags from Python scripts.

To import flags from a Python script, use the flags-import operation attribute.

Flag Definitions

See Flags for details on defining flags for an operation.

Source Code

Guild copies operation source code to a run directory for each run. Guild uses the run copy of the source code rather than the project source code. This services two purposes:

  • The source code copy for a run is definitive — it’s the source code that is run
  • Changes to the project do not effect an in-progress run

It’s important to copy the required source code files. By default, Guild copies text files with safeguards to prevent copying too many files or files that are too big. Change this behavior by defining a sourcecode attribute for operation or the operation model.

See Guild File Reference for more information.

Output Scalars

In some cases, Guild applies additional rules to capture scalars logged by known frameworks. Refer to Framework Scalars below for more information.

The sections that follow describe how you can configure Guild’s output scalar behavior.

Custom Output Scalars

Configure output scalars for an operation by defining a output-scalars attribute. Guild supports two schemes:

  • Pattern mapping
  • Pattern list

A pattern mapping associates patterns with scalar keys. Pattern mappings work well when you have a fixed set of scalars that you want to capture, and you want to ignore everything.

The following configuration captures scalars using a pattern mapping.

    loss: 'Loss: (\value)'
    accuracy: 'Accuracy: (\value)'

Disable Output Scalars

If you want to log scalars explicitly (e.g. using a TensorFlow summary writer) you can disable Guild’s output summary support by setting output-scalars to off.

  output-scalars: off

Keras Scalars

By default, Guild applies the following patterns when running Keras operations:

Epoch (?P<step>[0-9]+) Sets the current step used for subsequently logged scalar values
- ([a-z_]+): (\value) Captures scalar values staring with lower case (skips ETA, which would otherwise be logged as a scalar)


When an operation needs a file or other resource, it defines a dependency on a resource. Guild starts each run with an empty directory. If an operation needs a file, it must define it as a dependency.

Refer to Dependencies for details on defining and using dependencies in Guild.


Pipelines are multi-step runs defined using the steps attribute.

Refer to Pipelines for details on defining and using pipelines in Guild.


A model defines a set of related operations. Generally models correspond to the structures that you train, evaluate, and deploy. However, Guild models may define any operations or even be used for non-modeling functions.

Models must be defined using full format Guild files. Models are top-level objects with a model attribute.

- model: mnist
    train: mnist_train
    validate: mnist_val

Define a model when you want to:


A resource is a set of sources required by an operation. A source typically defines one or more source files. An operation indicates it requires a resource by defining it in the requires attribute.

Resources may be defined inline or as named resources. See Dependencies for more information.

Refer to Guild File Reference for resource attributes.


Guild supports installation and use of models and operations through packages. See Packages for more information.

Reusable Config

Guild supports reusable configuration through top-level config objects.

Configuration must be defined using full format Guild files.

Configuration objects may contain any attributes. Attributes are applied based on how the object is used.

Guild supports two uses of config objects:

Below is a sample config object.

- config: model-base
    train: '{{ name }}_train'
    validate: '{{ name }}_val'

Top-level config object named base-model that defines an operations attribute

This configuration can be referenced using the extends attribute of another top-level object to inherit the configuration attributes.

- model: mnist
  extends: model-base
    name: mnist

Top-level model object that extends base-model — it defines a name param, which resolves references in the inherited attributes


Guild files support inheritance where attributes of one object (parent) are applied by default to another object (child). A child may redefine attributes as needed.

Here’s an example of using inheritance (copied from above):

- config: model-base
    train: '{{ name }}_train'
    validate: '{{ name }}_val'

- model: mnist
  extends: model-base
    name: mnist

A common use of inheritance is to reuse resource definitions.

- config: data-support
      - operation: prepare-data

- operations:
      main: prepare_data

- model: mlp
  extends: data-support   # inherit the resources defined above
      main: train_mlp
      requires: prepared-data

Attribute Includes

You can reuse config settings through a special $include attribute. This attribute is used for flags and operations.

Here’s an example:

- config: train-flags
    learning-rate: 0.1
    batch-size: 100

- operations:
        $include: train-flags
        $include: train-flags

Including Files

Guild files can include other YAML files by using a top-level include object. The include type attribute specifies the path of the file to include. Paths are considered relative to the including Guild file.

Here is a sample guild.yml file that includes two files.

- include: guild-mnist.yml
- include: guild-cifar.yml

guild.yml — includes two files

The included files must be valid full format Guild files. Their contents are included in the Guild including file at the location each is defined.

- model: mnist
    train: mnist_train
    validate: mnist_validate

guild-mnist.yml — included by guild.yml above

- model: cifar
    train: cifar_train
    validate: cifar_valuate

guild-cifar.yml — also included by guild.yml above