Important

We are no longer actively maintaining this repository. All active work by the Allen Institute for Cell Science is located under the AllenCell organization.

Cell ABM Pipeline

Installation

Installation using Poetry

This project uses Poetry to manage dependencies and virtual environments.

1. Create the virtual environment:

poetry install

2. Activate the environment:

poetry shell

Alternative installation using pip

This project also includes a requirements.txt generated from the poetry.lock file. Install dependencies directly from this file using:

pip install -r requirements.txt

Install the package (note that you need pip ≥ 21.3):

pip install -e .

Usage

The pipeline uses Prefect for workflows and Hydra for composable configuration. Workflows can be run via CLI or through the Prefect UI as deployments.

When running via CLI, configurations can be passed in three ways: inline, using a single config file, or using composable config files. Note that Hydra supports additional overriding configurations via CLI for all three options.

Run with inline configs

Configurations can be passed directly using:

abmpipe demo :: parameters.name=demo_parameters context.name=demo_context series.name=demo_series

Run using single config file

Create a config file demo.yaml with the following contents:

context:
  name: demo_context
series:
  name: demo_series
parameters:
  name: demo_parameters

Then use:

abmpipe demo /path/to/demo.yaml

Run using composable config files

Create a configs directory with the following structure:

configs
├── context
│   └── demo.yaml
├── parameters
│   └── demo.yaml
└── series
    └── demo.yaml

Each demo.yaml should contain the field name: <name>. Then use:

abmpipe demo parameters=demo context=demo series=demo

Additional flags

Use the flag --dryrun to display the composed configuration without running the workflow.

Use the flag --deploy to create a Prefect deployment.

Adding secrets

Configs can use Secret fields. In configs, any field in the form ${secret:name-of-secret} will be resolved using the Prefect Secret loader. These values must be configured as a Secret Block in Prefect via a script:

from prefect.blocks.system import Secret

Secret(value="secret-value").save(name="name-of-secret")

or in the Prefect UI under Blocks.

Development

New flows can be added to the flows module with following structure:

from dataclasses import dataclass

from prefect import flow


@dataclass
class ParametersConfig:
    # TODO: add parameter config


@dataclass
class ContextConfig:
    # TODO: add context config


@dataclass
class SeriesConfig:
    # TODO: add series config


@flow(name="name-of-flow")
def run_flow(context: ContextConfig, series: SeriesConfig, parameters: ParametersConfig) -> None:
    # TODO: add flow

The command:

abmpipe name-of-flow

will create a new flow template under the flows module with the name name_of_flow.

Using notebooks

Notebooks can be helpful for prototyping flows.

Use dataclasses to specify configuration

Create dataclasses for all relevant configuration for the flow. Specify types and default values, if relevant. For flows in this repo, three types of configs are used:

• ParametersConfig specifies all parameters for the flow
• ContextConfig specifies the infrastructure context (e.g. local working path or S3 bucket names)
• SeriesConfig specifies the simulation series the flow is applied to (e.g. simulation name, conditions, seeds)

Load configuration into dataclasses

Configurations can be loaded in multiple ways.

1. Load entire configuration directly from an existing configuration file using the make_config_from_file function. Works best for simple configurations without interpolation.

config = make_config_from_file(ConfigDataclass, f"/path/to/config.yaml")

2. Load partial configuration directly from an existing configuration file using the make_config_from_file function. Missing fields in the config can be loaded from other configuration files using OmegaConf.load or set directly. Works best for configurations that use interpolation.

config = make_config_from_file(ConfigDataclass, f"/path/to/config.yaml")
config.field = OmegaConf.load(f"/path/to/another/config.yaml").field
config.field = "value"

3. Directly instantiate the config object. Fields in object initialization can also be loaded using OmegaConf.load. Works best for custom configurations or testing configurations.

config = ConfigDataclass(
    field="value",
    field=OmegaConf.load(f"/path/to/config.yaml").field,
    ...
)

Call tasks from collections

Import tasks from collections in the undecorated form:

from collection.module.task import task

Tasks can also be imported in decorated form:

from collection.module import task

but will need to called using task.fn() because we are not in a Prefect flow environment.

Converting notebooks to flows

Make sure the main flow method has the @flow decorator and imports should be switched to their @task decorated form to take advantage of Prefect task and flow monitoring.