spaCy projects let you manage and share end-to-end spaCy workflows for different use cases and domains, and orchestrate training, packaging and serving your custom pipelines. You can start off by cloning a pre-defined project template, adjust it to fit your needs, load in your data, train a pipeline, export it as a Python package, upload your outputs to a remote storage and share your results with your team. spaCy projects can be used via the new spacy project command and we provide templates in our projects repo.

spaCy projects make it easy to integrate with many other awesome tools in the data science and machine learning ecosystem to track and manage your data and experiments, iterate on demos and prototypes and ship your models into production.

1. Clone a project template

The spacy project clone command clones an existing project template and copies the files to a local directory. You can then run the project, e.g. to train a pipeline and edit the commands and scripts to build fully custom workflows.

By default, the project will be cloned into the current working directory. You can specify an optional second argument to define the output directory. The --repo option lets you define a custom repo to clone from if you don’t want to use the spaCy projects repo. You can also use any private repo you have access to with Git.

2. Fetch the project assets

Assets are data files your project needs – for example, the training and evaluation data or pretrained vectors and embeddings to initialize your model with. Each project template comes with a project.yml that defines the assets to download and where to put them. The spacy project assets will fetch the project assets for you:

Asset URLs can be a number of different protocols: HTTP, HTTPS, FTP, SSH, and even cloud storage such as GCS and S3. You can also fetch assets using git, by replacing the url string with a git block. spaCy will use Git’s “sparse checkout” feature to avoid downloading the whole repository.

Sometimes your project configuration may include large assets that you don’t necessarily want to download when you run spacy project assets. That’s why assets can be marked as extra - by default, these assets are not downloaded. If they should be, run spacy project assets --extra.

3. Run a command

Commands consist of one or more steps and can be run with spacy project run. The following will run the command preprocess defined in the project.yml:

Commands can define their expected dependencies and outputs using the deps (files the commands require) and outputs (files the commands create) keys. This allows your project to track changes and determine whether a command needs to be re-run. For instance, if your input data changes, you want to re-run the preprocess command. But if nothing changed, this step can be skipped. You can also set --force to force re-running a command, or --dry to perform a “dry run” and see what would happen (without actually running the script).

Since spaCy v3.4.2, spacy projects run checks your installed dependencies to verify that your environment is properly set up and aligns with the project’s requirements.txt, if there is one. If missing or conflicting dependencies are detected, a corresponding warning is displayed. If you’d like to disable the dependency check, set check_requirements: false in your project’s project.yml.

4. Run a workflow

Workflows are series of commands that are run in order and often depend on each other. For instance, to generate a pipeline package, you might start by converting your data, then run spacy train to train your pipeline on the converted data and if that’s successful, run spacy package to turn the best trained artifact into an installable Python package. The following command runs the workflow named all defined in the project.yml, and executes the commands it specifies, in order:

Using the expected dependencies and outputs defined in the commands, spaCy can determine whether to re-run a command (if its inputs or outputs have changed) or whether to skip it. If you’re looking to implement more advanced data pipelines and track your changes in Git, check out the Data Version Control (DVC) integration. The spacy project dvc command generates a DVC config file from a workflow defined in your project.yml so you can manage your spaCy project as a DVC repo.

5. Optional: Push to remote storage

After training a pipeline, you can optionally use the spacy project push command to upload your outputs to a remote storage, using protocols like S3, Google Cloud Storage or SSH. This can help you export your pipeline packages, share work with your team, or cache results to avoid repeating work.

The remotes section in your project.yml lets you assign names to the different storages. To download state from a remote storage, you can use the spacy project pull command. For more details, see the docs on remote storage.

Project directory and assets

project.yml

The project.yml defines the assets a project depends on, like datasets and pretrained weights, as well as a series of commands that can be run separately or as a workflow – for instance, to preprocess the data, convert it to spaCy’s format, train a pipeline, evaluate it and export metrics, package it and spin up a quick web demo. It looks pretty similar to a config file used to define CI pipelines.

explosion/projects/v3/pipelines/tagger_parser_ud/project.yml

Data assets

Assets are any files that your project might need, like training and development corpora or pretrained weights for initializing your model. Assets are defined in the assets block of your project.yml and can be downloaded using the project assets command. Defining checksums lets you verify that someone else running your project will use the same files you used. Asset URLs can be a number of different protocols: HTTP, HTTPS, FTP, SSH, and even cloud storage such as GCS and S3. You can also download assets from a Git repo instead.

Downloading from a URL or cloud storage

Under the hood, spaCy uses the smart-open library so you can use any protocol it supports. Note that you may need to install extra dependencies to use certain protocols.

Downloading from a Git repo

If a git block is provided, the asset is downloaded from the given Git repository. You can download from any repo that you have access to. Under the hood, this uses Git’s “sparse checkout” feature, so you’re only downloading the files you need and not the whole repo.

Working with private assets

For many projects, the datasets and weights you’re working with might be company-internal and not available over the internet. In that case, you can specify the destination paths and a checksum, and leave out the URL. When your teammates clone and run your project, they can place the files in the respective directory themselves. The project assets command will alert you about missing files and mismatched checksums, so you can ensure that others are running your project with the same data.

Dependencies and outputs

Each command defined in the project.yml can optionally define a list of dependencies and outputs. These are the files the command requires and creates. For example, a command for training a pipeline may depend on a config.cfg and the training and evaluation data, and it will export a directory model-best, which you can then re-use in other commands.

project.yml

If you’re running a command and it depends on files that are missing, spaCy will show you an error. If a command defines dependencies and outputs that haven’t changed since the last run, the command will be skipped. This means that you’re only re-running commands if they need to be re-run. Commands can also set no_skip: true if they should never be skipped – for example commands that run tests. Commands without outputs are also never skipped. To force re-running a command or workflow, even if nothing changed, you can set the --force flag.

Note that spacy project doesn’t compile any dependency graphs based on the dependencies and outputs, and won’t re-run previous steps automatically. For instance, if you only run the command train that depends on data created by preprocess and those files are missing, spaCy will show an error – it won’t just re-run preprocess. If you’re looking for more advanced data management, check out the Data Version Control (DVC) integration. If you’re planning on integrating your spaCy project with DVC, you can also use outputs_no_cache instead of outputs to define outputs that won’t be cached or tracked.

Files and directory structure

The project.yml can define a list of directories that should be created within a project – for instance, assets, training, corpus and so on. spaCy will make sure that these directories are always available, so your commands can write to and read from them. Project directories will also include all files and directories copied from the project template with spacy project clone. Here’s an example of a project directory:

Example project directory

If you don’t want a project to create a directory, you can delete it and remove its entry from the project.yml – just make sure it’s not required by any of the commands. Custom templates can use any directories they need – the only file that’s required for a project is the project.yml.

Custom scripts and projects

The project.yml lets you define any custom commands and run them as part of your training, evaluation or deployment workflows. The script section defines a list of commands that are called in a subprocess, in order. This lets you execute other Python scripts or command-line tools. Let’s say you’ve written a few integration tests that load the best model produced by the training command and check that it works correctly. You can now define a test command that calls into pytest, runs your tests and uses pytest-html to export a test report:

project.yml

Adding training/model-best to the command’s deps lets you ensure that the file is available. If not, spaCy will show an error and the command won’t run. Setting no_skip: true means that the command will always run, even if the dependencies (the trained pipeline) haven’t changed. This makes sense here, because you typically don’t want to skip your tests.

Writing custom scripts

Your project commands can include any custom scripts – essentially, anything you can run from the command line. Here’s an example of a custom script that uses typer for quick and easy command-line arguments that you can define via your project.yml:

scripts/custom_evaluation.py

In your project.yml, you can then run the script by calling python scripts/custom_evaluation.py with the function arguments. You can also use the vars section to define reusable variables that will be substituted in commands, paths and URLs. In this example, the batch size is defined as a variable will be added in place of ${vars.batch_size} in the script. Just like in the training config, you can also override settings on the command line – for example using --vars.batch_size.

project.yml

You can also use the env section to reference environment variables and make their values available to the commands. This can be useful for overriding settings on the command line and passing through system-level settings.

project.yml

Documenting your project

When your custom project is ready and you want to share it with others, you can use the spacy project document command to auto-generate a pretty, Markdown-formatted README file based on your project’s project.yml. It will list all commands, workflows and assets defined in the project and include details on how to run the project, as well as links to the relevant spaCy documentation to make it easy for others to get started using your project.

Under the hood, hidden markers are added to identify where the auto-generated content starts and ends. This means that you can add your own custom content before or after it and re-running the project document command will only update the auto-generated part. This makes it easy to keep your documentation up to date.

Cloning from your own repo

The spacy project clone command lets you customize the repo to clone from using the --repo option. It calls into git, so you’ll be able to clone from any repo that you have access to, including private repos.

At a minimum, a valid project template needs to contain a project.yml. It can also include other files, like custom scripts, a requirements.txt listing additional dependencies, training configs and model meta templates, or Jupyter notebooks with usage examples.

Remote Storage

You can persist your project outputs to a remote storage using the project push command. This can help you export your pipeline packages, share work with your team, or cache results to avoid repeating work. The project pull command will download any outputs that are in the remote storage and aren’t available locally.

You can list one or more remotes in the remotes section of your project.yml by mapping a string name to the URL of the storage. Under the hood, spaCy uses cloudpathlib to communicate with the remote storages, so you can use any protocol that cloudpathlib supports, including S3, Google Cloud Storage, and the local filesystem, although you may need to install extra dependencies to use certain protocols.

project.yml

For instance, let’s say you had the following command in your project.yml:

project.yml

After you finish training, you run project push to make sure the training/model-best output is saved to remote storage. spaCy will then construct a hash from your command script and the listed dependencies, corpus/train, corpus/dev and config.cfg, in order to identify the execution context of your output. It would then compute an MD5 hash of the training/model-best directory, and use those three pieces of information to construct the storage URL.

If you change the command or one of its dependencies (for instance, by editing the config.cfg file to tune the hyperparameters, a different creation hash will be calculated, so when you use project push you won’t be overwriting your previous file. The system even supports multiple outputs for the same file and the same context, which can happen if your training process is not deterministic, or if you have dependencies that aren’t represented in the command.

In summary, the spacy project remote storages are designed to make a particular set of trade-offs. Priority is placed on convenience, correctness and avoiding data loss. You can use project push freely, as you’ll never overwrite remote state, and you don’t have to come up with names or version numbers. However, it’s up to you to manage the size of your remote storage, and to remove files that are no longer relevant to you.

Integrations

Data Version Control (DVC)

Data assets like training corpora or pretrained weights are at the core of any NLP project, but they’re often difficult to manage: you can’t just check them into your Git repo to version and keep track of them. And if you have multiple steps that depend on each other, like a preprocessing step that generates your training data, you need to make sure the data is always up-to-date, and re-run all steps of your process every time, just to be safe.

Data Version Control (DVC) is a standalone open-source tool that integrates into your workflow like Git, builds a dependency graph for your data pipelines and tracks and caches your data files. If you’re downloading data from an external source, like a storage bucket, DVC can tell whether the resource has changed. It can also determine whether to re-run a step, depending on whether its input have changed or not. All metadata can be checked into a Git repo, so you’ll always be able to reproduce your experiments.

To set up DVC, install the package and initialize your spaCy project as a Git and DVC repo. You can also customize your DVC installation to include support for remote storage like Google Cloud Storage, S3, Azure, SSH and more.

The spacy project dvc command creates a dvc.yaml config file based on a workflow defined in your project.yml. Whenever you update your project, you can re-run the command to update your DVC config. You can then manage your spaCy project like any other DVC project, run dvc add to add and track assets and dvc repro to reproduce the workflow or individual commands.

Prodigy

Prodigy is a modern annotation tool for creating training data for machine learning models, developed by us. It integrates with spaCy out-of-the-box and provides many different annotation recipes for a variety of NLP tasks, with and without a model in the loop. If Prodigy is installed in your project, you can start the annotation server from your project.yml for a tight feedback loop between data development and training.

The following example shows a workflow for merging and exporting NER annotations collected with Prodigy and training a spaCy pipeline:

project.yml

The train-curve recipe is another cool workflow you can include in your project. It will run the training with different portions of the data, e.g. 25%, 50%, 75% and 100%. As a rule of thumb, if accuracy increases in the last segment, this could indicate that collecting more annotations of the same type might improve the model further.

project.yml (excerpt)

You can use the same approach for various types of projects and annotation workflows, including named entity recognition, span categorization, text classification, dependency parsing, part-of-speech tagging or fully custom recipes. You can also use spaCy project templates to quickly start the annotation server to collect more annotations and add them to your Prodigy dataset.

Streamlit

Streamlit is a Python framework for building interactive data apps. The spacy-streamlit package helps you integrate spaCy visualizations into your Streamlit apps and quickly spin up demos to explore your pipelines interactively. It includes a full embedded visualizer, as well as individual components.

Using spacy-streamlit, your projects can easily define their own scripts that spin up an interactive visualizer, using the latest pipeline you trained, or a selection of pipelines so you can compare their results.

project.yml

The following script is called from the project.yml and takes two positional command-line argument: a comma-separated list of paths or packages to load the pipelines from and an example text to use as the default text.

explosion/projects/v3/integrations/streamlit/scripts/visualize.py

FastAPI

FastAPI is a modern high-performance framework for building REST APIs with Python, based on Python type hints. It’s become a popular library for serving machine learning models and you can use it in your spaCy projects to quickly serve up a trained pipeline and make it available behind a REST API.

project.yml

The script included in the template shows a simple REST API with a POST endpoint that accepts batches of texts and returns batches of predictions, e.g. named entities found in the documents. Type hints and pydantic are used to define the expected data types.

explosion/projects/v3/integrations/fastapi/scripts/main.py

Weights & Biases

Weights & Biases is a popular platform for experiment tracking. spaCy integrates with it out-of-the-box via the WandbLogger, which you can add as the [training.logger] block of your training config. The results of each step are then logged in your project, together with the full training config. This means that every hyperparameter, registered function name and argument will be tracked and you’ll be able to see the impact it has on your results.

Hugging Face Hub

The Hugging Face Hub lets you upload models and share them with others. It hosts models as Git-based repositories which are storage spaces that can contain all your files. It support versioning, branches and custom metadata out-of-the-box, and provides browser-based visualizers for exploring your models interactively, as well as an API for production use. The spacy-huggingface-hub package automatically adds the huggingface-hub command to your spacy CLI if it’s installed.

You can then upload any pipeline packaged with spacy package. Make sure to set --build wheel to output a binary .whl file. The uploader will read all metadata from the pipeline package, including the auto-generated pretty README.md and the model details available in the meta.json. For examples, check out the spaCy pipelines we’ve uploaded.

After uploading, you will see the live URL of your pipeline packages, as well as the direct URL to the model wheel you can install via pip install. You’ll also be able to test your pipeline interactively from your browser:

In your project.yml, you can add a command that uploads your trained and packaged pipeline to the hub. You can either run this as a manual step, or automatically as part of a workflow. Make sure to set --build wheel when running spacy package to build a wheel file for your pipeline package.

project.yml