spaCy/website/docs/usage/training.md

title	teaser	next	menu
Training Pipelines & Models	Train and update components on your own data and integrate custom models	/usage/layers-architectures	Introduction basics Quickstart quickstart Config System config Custom Training config-custom Custom Functions custom-functions Initialization initialization Data Utilities data Parallel Training parallel-training Internal API api

Introduction to training

import Training101 from 'usage/101/_training.md'

If you need to label a lot of data, check out Prodigy, a new, active learning-powered annotation tool we've developed. Prodigy is fast and extensible, and comes with a modern web application that helps you collect training data faster. It integrates seamlessly with spaCy, pre-selects the most relevant examples for annotation, and lets you train and evaluate ready-to-use spaCy pipelines.

Quickstart

The recommended way to train your spaCy pipelines is via the spacy train command on the command line. It only needs a single config.cfg configuration file that includes all settings and hyperparameters. You can optionally overwrite settings on the command line, and load in a Python file to register custom functions and architectures. This quickstart widget helps you generate a starter config with the recommended settings for your specific use case. It's also available in spaCy as the init config command.

Upgrade to the latest version of spaCy to use the quickstart widget. For earlier releases, follow the CLI instructions to generate a compatible config.

Instructions: widget

1. Select your requirements and settings.
2. Use the buttons at the bottom to save the result to your clipboard or a file base_config.cfg.
3. Run init fill-config to create a full config.
4. Run train with your config and data.

Instructions: CLI

1. Run the init config command and specify your requirements and settings as CLI arguments.
2. Run train with the exported config and data.

import QuickstartTraining from 'widgets/quickstart-training.js'

After you've saved the starter config to a file base_config.cfg, you can use the init fill-config command to fill in the remaining defaults. Training configs should always be complete and without hidden defaults, to keep your experiments reproducible.

$ python -m spacy init fill-config base_config.cfg config.cfg

Tip: Debug your data

The debug data command lets you analyze and validate your training and development data, get useful stats, and find problems like invalid entity annotations, cyclic dependencies, low data labels and more.

$ python -m spacy debug data config.cfg

Instead of exporting your starter config from the quickstart widget and auto-filling it, you can also use the init config command and specify your requirement and settings as CLI arguments. You can now add your data and run train with your config. See the convert command for details on how to convert your data to spaCy's binary .spacy format. You can either include the data paths in the [paths] section of your config, or pass them in via the command line.

$ python -m spacy train config.cfg --output ./output --paths.train ./train.spacy --paths.dev ./dev.spacy

Tip: Enable your GPU

Use the --gpu-id option to select the GPU:

$ python -m spacy train config.cfg --gpu-id 0

The recommended config settings generated by the quickstart widget and the init config command are based on some general best practices and things we've found to work well in our experiments. The goal is to provide you with the most useful defaults.

Under the hood, the quickstart_training.jinja template defines the different combinations – for example, which parameters to change if the pipeline should optimize for efficiency vs. accuracy. The file quickstart_training_recommendations.yml collects the recommended settings and available resources for each language including the different transformer weights. For some languages, we include different transformer recommendations, depending on whether you want the model to be more efficient or more accurate. The recommendations will be evolving as we run more experiments.

The easiest way to get started is to clone a project template and run it – for example, this end-to-end template that lets you train a part-of-speech tagger and dependency parser on a Universal Dependencies treebank.

Training config system

Training config files include all settings and hyperparameters for training your pipeline. Instead of providing lots of arguments on the command line, you only need to pass your config.cfg file to spacy train. Under the hood, the training config uses the configuration system provided by our machine learning library Thinc. This also makes it easy to integrate custom models and architectures, written in your framework of choice. Some of the main advantages and features of spaCy's training config are:

• Structured sections. The config is grouped into sections, and nested sections are defined using the . notation. For example, [components.ner] defines the settings for the pipeline's named entity recognizer. The config can be loaded as a Python dict.
• References to registered functions. Sections can refer to registered functions like model architectures, optimizers or schedules and define arguments that are passed into them. You can also register your own functions to define custom architectures or methods, reference them in your config and tweak their parameters.
• Interpolation. If you have hyperparameters or other settings used by multiple components, define them once and reference them as variables.
• Reproducibility with no hidden defaults. The config file is the "single source of truth" and includes all settings.
• Automated checks and validation. When you load a config, spaCy checks if the settings are complete and if all values have the correct types. This lets you catch potential mistakes early. In your custom architectures, you can use Python type hints to tell the config which types of data to expect.

%%GITHUB_SPACY/spacy/default_config.cfg

Under the hood, the config is parsed into a dictionary. It's divided into sections and subsections, indicated by the square brackets and dot notation. For example, [training] is a section and [training.batch_size] a subsection. Subsections can define values, just like a dictionary, or use the @ syntax to refer to registered functions. This allows the config to not just define static settings, but also construct objects like architectures, schedules, optimizers or any other custom components. The main top-level sections of a config file are:

Section	Description
nlp	Definition of the nlp object, its tokenizer and processing pipeline component names.
components	Definitions of the pipeline components and their models.
paths	Paths to data and other assets. Re-used across the config as variables, e.g. ${paths.train}, and can be overwritten on the CLI.
system	Settings related to system and hardware. Re-used across the config as variables, e.g. ${system.seed}, and can be overwritten on the CLI.
training	Settings and controls for the training and evaluation process.
pretraining	Optional settings and controls for the language model pretraining.
initialize	Data resources and arguments passed to components when nlp.initialize is called before training (but not at runtime).

For a full overview of spaCy's config format and settings, see the data format documentation and Thinc's config system docs. The settings available for the different architectures are documented with the model architectures API. See the Thinc documentation for optimizers and schedules.

Config lifecycle at runtime and training

A pipeline's config.cfg is considered the "single source of truth", both at training and runtime. Under the hood, Language.from_config takes care of constructing the nlp object using the settings defined in the config. An nlp object's config is available as nlp.config and it includes all information about the pipeline, as well as the settings used to train and initialize it.

At runtime spaCy will only use the [nlp] and [components] blocks of the config and load all data, including tokenization rules, model weights and other resources from the pipeline directory. The [training] block contains the settings for training the model and is only used during training. Similarly, the [initialize] block defines how the initial nlp object should be set up before training and whether it should be initialized with vectors or pretrained tok2vec weights, or any other data needed by the components.

The initialization settings are only loaded and used when nlp.initialize is called (typically right before training). This allows you to set up your pipeline using local data resources and custom functions, and preserve the information in your config – but without requiring it to be available at runtime. You can also use this mechanism to provide data paths to custom pipeline components and custom tokenizers – see the section on custom initialization for details.

Overwriting config settings on the command line

The config system means that you can define all settings in one place and in a consistent format. There are no command-line arguments that need to be set, and no hidden defaults. However, there can still be scenarios where you may want to override config settings when you run spacy train. This includes file paths to vectors or other resources that shouldn't be hard-code in a config file, or system-dependent settings.

For cases like this, you can set additional command-line options starting with -- that correspond to the config section and value to override. For example, --paths.train ./corpus/train.spacy sets the train value in the [paths] block.

$ python -m spacy train config.cfg --paths.train ./corpus/train.spacy --paths.dev ./corpus/dev.spacy --training.batch_size 128

Only existing sections and values in the config can be overwritten. At the end of the training, the final filled config.cfg is exported with your pipeline, so you'll always have a record of the settings that were used, including your overrides. Overrides are added before variables are resolved, by the way – so if you need to use a value in multiple places, reference it across your config and override it on the CLI once.

💡 Tip: Verbose logging

If you're using config overrides, you can set the --verbose flag on spacy train to make spaCy log more info, including which overrides were set via the CLI and environment variables.

Adding overrides via environment variables

Instead of defining the overrides as CLI arguments, you can also use the SPACY_CONFIG_OVERRIDES environment variable using the same argument syntax. This is especially useful if you're training models as part of an automated process. Environment variables take precedence over CLI overrides and values defined in the config file.

$ SPACY_CONFIG_OVERRIDES="--system.gpu_allocator pytorch --training.batch_size 128" ./your_script.sh

Reading from standard input

Setting the config path to - on the command line lets you read the config from standard input and pipe it forward from a different process, like init config or your own custom script. This is especially useful for quick experiments, as it lets you generate a config on the fly without having to save to and load from disk.

💡 Tip: Writing to stdout

When you run init config, you can set the output path to - to write to stdout. In a custom script, you can print the string config, e.g. print(nlp.config.to_str()).

$ python -m spacy init config - --lang en --pipeline ner,textcat --optimize accuracy | python -m spacy train - --paths.train ./corpus/train.spacy --paths.dev ./corpus/dev.spacy

Using variable interpolation

Another very useful feature of the config system is that it supports variable interpolation for both values and sections. This means that you only need to define a setting once and can reference it across your config using the ${section.value} syntax. In this example, the value of seed is reused within the [training] block, and the whole block of [training.optimizer] is reused in [pretraining] and will become pretraining.optimizer.

### config.cfg (excerpt) {highlight="5,18"}
[system]
seed = 0

[training]
seed = ${system.seed}

[training.optimizer]
@optimizers = "Adam.v1"
beta1 = 0.9
beta2 = 0.999
L2_is_weight_decay = true
L2 = 0.01
grad_clip = 1.0
use_averages = false
eps = 1e-8

[pretraining]
optimizer = ${training.optimizer}

You can also use variables inside strings. In that case, it works just like f-strings in Python. If the value of a variable is not a string, it's converted to a string.

[paths]
version = 5
root = "/Users/you/data"
train = "${paths.root}/train_${paths.version}.spacy"
# Result: /Users/you/data/train_5.spacy

If you need to change certain values between training runs, you can define them once, reference them as variables and then override them on the CLI. For example, --paths.root /other/root will change the value of root in the block [paths] and the change will be reflected across all other values that reference this variable.

Customizing the pipeline and training

Defining pipeline components

You typically train a pipeline of one or more components. The [components] block in the config defines the available pipeline components and how they should be created – either by a built-in or custom factory, or sourced from an existing trained pipeline. For example, [components.parser] defines the component named "parser" in the pipeline. There are different ways you might want to treat your components during training, and the most common scenarios are:

1. Train a new component from scratch on your data.
2. Update an existing trained component with more examples.
3. Include an existing trained component without updating it.
4. Include a non-trainable component, like a rule-based EntityRuler or Sentencizer, or a fully custom component.

If a component block defines a factory, spaCy will look it up in the built-in or custom components and create a new component from scratch. All settings defined in the config block will be passed to the component factory as arguments. This lets you configure the model settings and hyperparameters. If a component block defines a source, the component will be copied over from an existing trained pipeline, with its existing weights. This lets you include an already trained component in your pipeline, or update a trained component with more data specific to your use case.

### config.cfg (excerpt)
[components]

# "parser" and "ner" are sourced from a trained pipeline
[components.parser]
source = "en_core_web_sm"

[components.ner]
source = "en_core_web_sm"

# "textcat" and "custom" are created blank from a built-in / custom factory
[components.textcat]
factory = "textcat"

[components.custom]
factory = "your_custom_factory"
your_custom_setting = true

The pipeline setting in the [nlp] block defines the pipeline components added to the pipeline, in order. For example, "parser" here references [components.parser]. By default, spaCy will update all components that can be updated. Trainable components that are created from scratch are initialized with random weights. For sourced components, spaCy will keep the existing weights and resume training.

If you don't want a component to be updated, you can freeze it by adding it to the frozen_components list in the [training] block. Frozen components are not updated during training and are included in the final trained pipeline as-is. They are also excluded when calling nlp.initialize.

Note on frozen components

Even though frozen components are not updated during training, they will still run during evaluation. This is very important, because they may still impact your model's performance – for instance, a sentence boundary detector can impact what the parser or entity recognizer considers a valid parse. So the evaluation results should always reflect what your pipeline will produce at runtime.

[nlp]
lang = "en"
pipeline = ["parser", "ner", "textcat", "custom"]

[training]
frozen_components = ["parser", "custom"]

When the components in your pipeline share an embedding layer, the performance of your frozen component will be degraded if you continue training other layers with the same underlying Tok2Vec instance. As a rule of thumb, ensure that your frozen components are truly independent in the pipeline.

To automatically replace a shared token-to-vector listener with an independent copy of the token-to-vector layer, you can use the replace_listeners setting of a sourced component, pointing to the listener layer(s) in the config. For more details on how this works under the hood, see Language.replace_listeners.

[training]
frozen_components = ["tagger"]

[components.tagger]
source = "en_core_web_sm"
replace_listeners = ["model.tok2vec"]

Using predictions from preceding components

By default, components are updated in isolation during training, which means that they don't see the predictions of any earlier components in the pipeline. A component receives Example.predicted as input and compares its predictions to Example.reference without saving its annotations in the predicted doc.

Instead, if certain components should set their annotations during training, use the setting annotating_components in the [training] block to specify a list of components. For example, the feature DEP from the parser could be used as a tagger feature by including DEP in the tok2vec attrs and including parser in annotating_components:

### config.cfg (excerpt) {highlight="7,12"}
[nlp]
pipeline = ["parser", "tagger"]

[components.tagger.model.tok2vec.embed]
@architectures = "spacy.MultiHashEmbed.v1"
width = ${components.tagger.model.tok2vec.encode.width}
attrs = ["NORM","DEP"]
rows = [5000,2500]
include_static_vectors = false

[training]
annotating_components = ["parser"]

Any component in the pipeline can be included as an annotating component, including frozen components. Frozen components can set annotations during training just as they would set annotations during evaluation or when the final pipeline is run. The config excerpt below shows how a frozen ner component and a sentencizer can provide the required doc.sents and doc.ents for the entity linker during training:

### config.cfg (excerpt)
[nlp]
pipeline = ["sentencizer", "ner", "entity_linker"]

[components.ner]
source = "en_core_web_sm"

[training]
frozen_components = ["ner"]
annotating_components = ["sentencizer", "ner"]

Be aware that non-frozen annotating components with statistical models will run twice on each batch, once to update the model and once to apply the now-updated model to the predicted docs.

Using registered functions

The training configuration defined in the config file doesn't have to only consist of static values. Some settings can also be functions. For instance, the batch_size can be a number that doesn't change, or a schedule, like a sequence of compounding values, which has shown to be an effective trick (see Smith et al., 2017).

### With static value
[training]
batch_size = 128

To refer to a function instead, you can make [training.batch_size] its own section and use the @ syntax to specify the function and its arguments – in this case compounding.v1 defined in the function registry. All other values defined in the block are passed to the function as keyword arguments when it's initialized. You can also use this mechanism to register custom implementations and architectures and reference them from your configs.

How the config is resolved

The config file is parsed into a regular dictionary and is resolved and validated bottom-up. Arguments provided for registered functions are checked against the function's signature and type annotations. The return value of a registered function can also be passed into another function – for instance, a learning rate schedule can be provided as the an argument of an optimizer.

### With registered function
[training.batch_size]
@schedules = "compounding.v1"
start = 100
stop = 1000
compound = 1.001

Model architectures

💡 Model type annotations

In the documentation and code base, you may come across type annotations and descriptions of Thinc model types, like Model[List[Doc], List[Floats2d]]. This so-called generic type describes the layer and its input and output type – in this case, it takes a list of Doc objects as the input and list of 2-dimensional arrays of floats as the output. You can read more about defining Thinc models here. Also see the type checking for how to enable linting in your editor to see live feedback if your inputs and outputs don't match.

A model architecture is a function that wires up a Thinc Model instance, which you can then use in a component or as a layer of a larger network. You can use Thinc as a thin wrapper around frameworks such as PyTorch, TensorFlow or MXNet, or you can implement your logic in Thinc directly. For more details and examples, see the usage guide on layers and architectures.

spaCy's built-in components will never construct their Model instances themselves, so you won't have to subclass the component to change its model architecture. You can just update the config so that it refers to a different registered function. Once the component has been created, its Model instance has already been assigned, so you cannot change its model architecture. The architecture is like a recipe for the network, and you can't change the recipe once the dish has already been prepared. You have to make a new one. spaCy includes a variety of built-in architectures for different tasks. For example:

Architecture	Description
HashEmbedCNN	Build spaCy’s "standard" embedding layer, which uses hash embedding with subword features and a CNN with layer-normalized maxout. Model[List[Doc], List[Floats2d]]
TransitionBasedParser	Build a transition-based parser model used in the default EntityRecognizer and DependencyParser. Model[List[Docs], List[List[Floats2d]]]
TextCatEnsemble	Stacked ensemble of a bag-of-words model and a neural network model with an internal CNN embedding layer. Used in the default TextCategorizer. Model[List[Doc], Floats2d]

Metrics, training output and weighted scores

When you train a pipeline using the spacy train command, you'll see a table showing the metrics after each pass over the data. The available metrics depend on the pipeline components. Pipeline components also define which scores are shown and how they should be weighted in the final score that decides about the best model.

The training.score_weights setting in your config.cfg lets you customize the scores shown in the table and how they should be weighted. In this example, the labeled dependency accuracy and NER F-score count towards the final score with 40% each and the tagging accuracy makes up the remaining 20%. The tokenization accuracy and speed are both shown in the table, but not counted towards the score.

Why do I need score weights?

At the end of your training process, you typically want to select the best model – but what "best" means depends on the available components and your specific use case. For instance, you may prefer a pipeline with higher NER and lower POS tagging accuracy over a pipeline with lower NER and higher POS accuracy. You can express this preference in the score weights, e.g. by assigning ents_f (NER F-score) a higher weight.

[training.score_weights]
dep_las = 0.4
dep_uas = null
ents_f = 0.4
tag_acc = 0.2
token_acc = 0.0
speed = 0.0

The score_weights don't have to sum to 1.0 – but it's recommended. When you generate a config for a given pipeline, the score weights are generated by combining and normalizing the default score weights of the pipeline components. The default score weights are defined by each pipeline component via the default_score_weights setting on the @Language.factory decorator. By default, all pipeline components are weighted equally. If a score weight is set to null, it will be excluded from the logs and the score won't be weighted.

Name	Description
Loss	The training loss representing the amount of work left for the optimizer. Should decrease, but usually not to 0.
Precision (P)	Percentage of predicted annotations that were correct. Should increase.
Recall (R)	Percentage of reference annotations recovered. Should increase.
F-Score (F)	Harmonic mean of precision and recall. Should increase.
UAS / LAS	Unlabeled and labeled attachment score for the dependency parser, i.e. the percentage of correct arcs. Should increase.
Speed	Prediction speed in words per second (WPS). Should stay stable.

Note that if the development data has raw text, some of the gold-standard entities might not align to the predicted tokenization. These tokenization errors are excluded from the NER evaluation. If your tokenization makes it impossible for the model to predict 50% of your entities, your NER F-score might still look good.

Custom functions

Registered functions in the training config files can refer to built-in implementations, but you can also plug in fully custom implementations. All you need to do is register your function using the @spacy.registry decorator with the name of the respective registry, e.g. @spacy.registry.architectures, and a string name to assign to your function. Registering custom functions allows you to plug in models defined in PyTorch or TensorFlow, make custom modifications to the nlp object, create custom optimizers or schedules, or stream in data and preprocesses it on the fly while training.

Each custom function can have any number of arguments that are passed in via the config, just the built-in functions. If your function defines default argument values, spaCy is able to auto-fill your config when you run init fill-config. If you want to make sure that a given parameter is always explicitly set in the config, avoid setting a default value for it.

Training with custom code

### Training
$ python -m spacy train config.cfg --code functions.py

### Packaging
$ python -m spacy package ./model-best ./packages --code functions.py

The spacy train recipe lets you specify an optional argument --code that points to a Python file. The file is imported before training and allows you to add custom functions and architectures to the function registry that can then be referenced from your config.cfg. This lets you train spaCy pipelines with custom components, without having to re-implement the whole training workflow. When you package your trained pipeline later using spacy package, you can provide one or more Python files to be included in the package and imported in its __init__.py. This means that any custom architectures, functions or components will be shipped with your pipeline and registered when it's loaded. See the documentation on saving and loading pipelines for details.

Example: Modifying the nlp object

For many use cases, you don't necessarily want to implement the whole Language subclass and language data from scratch – it's often enough to make a few small modifications, like adjusting the tokenization rules or language defaults like stop words. The config lets you provide five optional callback functions that give you access to the language class and nlp object at different points of the lifecycle:

Callback	Description
nlp.before_creation	Called before the nlp object is created and receives the language subclass like English (not the instance). Useful for writing to the Language.Defaults aside from the tokenizer settings.
nlp.after_creation	Called right after the nlp object is created, but before the pipeline components are added to the pipeline and receives the nlp object.
nlp.after_pipeline_creation	Called right after the pipeline components are created and added and receives the nlp object. Useful for modifying pipeline components.
initialize.before_init	Called before the pipeline components are initialized and receives the nlp object for in-place modification. Useful for modifying the tokenizer settings, similar to the v2 base model option.
initialize.after_init	Called after the pipeline components are initialized and receives the nlp object for in-place modification.

The @spacy.registry.callbacks decorator lets you register your custom function in the callbacks registry under a given name. You can then reference the function in a config block using the @callbacks key. If a block contains a key starting with an @, it's interpreted as a reference to a function. Because you've registered the function, spaCy knows how to create it when you reference "customize_language_data" in your config. Here's an example of a callback that runs before the nlp object is created and adds a custom stop word to the defaults:

config.cfg

[nlp.before_creation]
@callbacks = "customize_language_data"

### functions.py {highlight="3,6"}
import spacy

@spacy.registry.callbacks("customize_language_data")
def create_callback():
    def customize_language_data(lang_cls):
        lang_cls.Defaults.stop_words.add("good")
        return lang_cls

    return customize_language_data

Remember that a registered function should always be a function that spaCy calls to create something. In this case, it creates a callback – it's not the callback itself.

Any registered function – in this case create_callback – can also take arguments that can be set by the config. This lets you implement and keep track of different configurations, without having to hack at your code. You can choose any arguments that make sense for your use case. In this example, we're adding the arguments extra_stop_words (a list of strings) and debug (boolean) for printing additional info when the function runs.

config.cfg

[nlp.before_creation]
@callbacks = "customize_language_data"
extra_stop_words = ["ooh", "aah"]
debug = true

### functions.py {highlight="5,7-9"}
from typing import List
import spacy

@spacy.registry.callbacks("customize_language_data")
def create_callback(extra_stop_words: List[str] = [], debug: bool = False):
    def customize_language_data(lang_cls):
        lang_cls.Defaults.stop_words.update(extra_stop_words)
        if debug:
            print("Updated stop words")
        return lang_cls

    return customize_language_data

spaCy's configs are powered by our machine learning library Thinc's configuration system, which supports type hints and even advanced type annotations using pydantic. If your registered function provides type hints, the values that are passed in will be checked against the expected types. For example, debug: bool in the example above will ensure that the value received as the argument debug is a boolean. If the value can't be coerced into a boolean, spaCy will raise an error. debug: pydantic.StrictBool will force the value to be a boolean and raise an error if it's not – for instance, if your config defines 1 instead of true.

With your functions.py defining additional code and the updated config.cfg, you can now run spacy train and point the argument --code to your Python file. Before loading the config, spaCy will import the functions.py module and your custom functions will be registered.

$ python -m spacy train config.cfg --output ./output --code ./functions.py

Example: Modifying tokenizer settings

Use the initialize.before_init callback to modify the tokenizer settings when training a new pipeline. Write a registered callback that modifies the tokenizer settings and specify this callback in your config:

config.cfg

[initialize]

[initialize.before_init]
@callbacks = "customize_tokenizer"

### functions.py
from spacy.util import registry, compile_suffix_regex

@registry.callbacks("customize_tokenizer")
def make_customize_tokenizer():
    def customize_tokenizer(nlp):
        # remove a suffix
        suffixes = list(nlp.Defaults.suffixes)
        suffixes.remove("\\[")
        suffix_regex = compile_suffix_regex(suffixes)
        nlp.tokenizer.suffix_search = suffix_regex.search

        # add a special case
        nlp.tokenizer.add_special_case("_SPECIAL_", [{"ORTH": "_SPECIAL_"}])
    return customize_tokenizer

When training, provide the function above with the --code option:

$ python -m spacy train config.cfg --code ./functions.py

Because this callback is only called in the one-time initialization step before training, the callback code does not need to be packaged with the final pipeline package. However, to make it easier for others to replicate your training setup, you can choose to package the initialization callbacks with the pipeline package or to publish them separately.

• nlp.before_creation is the best place to modify language defaults other than the tokenizer settings.
• initialize.before_init is the best place to modify tokenizer settings when training a new pipeline.

Unlike the other language defaults, the tokenizer settings are saved with the pipeline with nlp.to_disk(), so modifications made in nlp.before_creation will be clobbered by the saved settings when the trained pipeline is loaded from disk.

Example: Custom logging function

During training, the results of each step are passed to a logger function. By default, these results are written to the console with the ConsoleLogger. There is also built-in support for writing the log files to Weights & Biases with the WandbLogger. On each step, the logger function receives a dictionary with the following keys:

Key	Value
epoch	How many passes over the data have been completed. int
step	How many steps have been completed. int
score	The main score from the last evaluation, measured on the dev set. float
other_scores	The other scores from the last evaluation, measured on the dev set. Dict[str, Any]
losses	The accumulated training losses, keyed by component name. Dict[str, float]
checkpoints	A list of previous results, where each result is a (score, step) tuple. List[Tuple[float, int]]

You can easily implement and plug in your own logger that records the training results in a custom way, or sends them to an experiment management tracker of your choice. In this example, the function my_custom_logger.v1 writes the tabular results to a file:

### config.cfg (excerpt)
[training.logger]
@loggers = "my_custom_logger.v1"
log_path = "my_file.tab"

### functions.py
import sys
from typing import IO, Tuple, Callable, Dict, Any, Optional
import spacy
from spacy import Language
from pathlib import Path

@spacy.registry.loggers("my_custom_logger.v1")
def custom_logger(log_path):
    def setup_logger(
        nlp: Language,
        stdout: IO=sys.stdout,
        stderr: IO=sys.stderr
    ) -> Tuple[Callable, Callable]:
        stdout.write(f"Logging to {log_path}\\n")
        log_file = Path(log_path).open("w", encoding="utf8")
        log_file.write("step\\t")
        log_file.write("score\\t")
        for pipe in nlp.pipe_names:
            log_file.write(f"loss_{pipe}\\t")
        log_file.write("\\n")

        def log_step(info: Optional[Dict[str, Any]]):
            if info:
                log_file.write(f"{info['step']}\\t")
                log_file.write(f"{info['score']}\\t")
                for pipe in nlp.pipe_names:
                    log_file.write(f"{info['losses'][pipe]}\\t")
                log_file.write("\\n")

        def finalize():
            log_file.close()

        return log_step, finalize

    return setup_logger

Example: Custom batch size schedule

You can also implement your own batch size schedule to use during training. The @spacy.registry.schedules decorator lets you register that function in the schedules registry and assign it a string name:

Why the version in the name?

A big benefit of the config system is that it makes your experiments reproducible. We recommend versioning the functions you register, especially if you expect them to change (like a new model architecture). This way, you know that a config referencing v1 means a different function than a config referencing v2.

### functions.py
import spacy

@spacy.registry.schedules("my_custom_schedule.v1")
def my_custom_schedule(start: int = 1, factor: float = 1.001):
   while True:
      yield start
      start = start * factor

In your config, you can now reference the schedule in the [training.batch_size] block via @schedules. If a block contains a key starting with an @, it's interpreted as a reference to a function. All other settings in the block will be passed to the function as keyword arguments. Keep in mind that the config shouldn't have any hidden defaults and all arguments on the functions need to be represented in the config.

### config.cfg (excerpt)
[training.batch_size]
@schedules = "my_custom_schedule.v1"
start = 2
factor = 1.005

Defining custom architectures

Built-in pipeline components such as the tagger or named entity recognizer are constructed with default neural network models. You can change the model architecture entirely by implementing your own custom models and providing those in the config when creating the pipeline component. See the documentation on layers and model architectures for more details.

### config.cfg
[components.tagger]
factory = "tagger"

[components.tagger.model]
@architectures = "custom_neural_network.v1"
output_width = 512

### functions.py
from typing import List
from thinc.types import Floats2d
from thinc.api import Model
import spacy
from spacy.tokens import Doc

@spacy.registry.architectures("custom_neural_network.v1")
def custom_neural_network(output_width: int) -> Model[List[Doc], List[Floats2d]]:
    return create_model(output_width)

Customizing the initialization

When you start training a new model from scratch, spacy train will call nlp.initialize to initialize the pipeline and load the required data. All settings for this are defined in the [initialize] block of the config, so you can keep track of how the initial nlp object was created. The initialization process typically includes the following:

config.cfg (excerpt)

[initialize]
vectors = ${paths.vectors}
init_tok2vec = ${paths.init_tok2vec}

[initialize.components]
# Settings for components

1. Load in data resources defined in the [initialize] config, including word vectors and pretrained tok2vec weights.
2. Call the initialize methods of the tokenizer (if implemented, e.g. for Chinese) and pipeline components with a callback to access the training data, the current nlp object and any custom arguments defined in the [initialize] config.
3. In pipeline components: if needed, use the data to infer missing shapes and set up the label scheme if no labels are provided. Components may also load other data like lookup tables or dictionaries.

The initialization step allows the config to define all settings required for the pipeline, while keeping a separation between settings and functions that should only be used before training to set up the initial pipeline, and logic and configuration that needs to be available at runtime. Without that separation, it would be very difficult to use the same, reproducible config file because the component settings required for training (load data from an external file) wouldn't match the component settings required at runtime (load what's included with the saved nlp object and don't depend on external file).

For details and examples of how pipeline components can save and load data assets like model weights or lookup tables, and how the component initialization is implemented under the hood, see the usage guide on serializing and initializing component data.

Initializing labels

Built-in pipeline components like the EntityRecognizer or DependencyParser need to know their available labels and associated internal meta information to initialize their model weights. Using the get_examples callback provided on initialization, they're able to read the labels off the training data automatically, which is very convenient – but it can also slow down the training process to compute this information on every run.

The init labels command lets you auto-generate JSON files containing the label data for all supported components. You can then pass in the labels in the [initialize] settings for the respective components to allow them to initialize faster.

config.cfg

[initialize.components.ner]

[initialize.components.ner.labels]
@readers = "spacy.read_labels.v1"
path = "corpus/labels/ner.json

$ python -m spacy init labels config.cfg ./corpus --paths.train ./corpus/train.spacy

Under the hood, the command delegates to the label_data property of the pipeline components, for instance EntityRecognizer.label_data.

The JSON format differs for each component and some components need additional meta information about their labels. The format exported by init labels matches what the components need, so you should always let spaCy auto-generate the labels for you.

Data utilities

spaCy includes various features and utilities to make it easy to train models using your own data, manage training and evaluation corpora, convert existing annotations and configure data augmentation strategies for more robust models.

Converting existing corpora and annotations

If you have training data in a standard format like .conll or .conllu, the easiest way to convert it for use with spaCy is to run spacy convert and pass it a file and an output directory. By default, the command will pick the converter based on the file extension.

$ python -m spacy convert ./train.gold.conll ./corpus

💡 Tip: Converting from Prodigy

If you're using the Prodigy annotation tool to create training data, you can run the data-to-spacy command to merge and export multiple datasets for use with spacy train. Different types of annotations on the same text will be combined, giving you one corpus to train multiple components.

Training workflows often consist of multiple steps, from preprocessing the data all the way to packaging and deploying the trained model. spaCy projects let you define all steps in one file, manage data assets, track changes and share your end-to-end processes with your team.

The binary .spacy format is a serialized DocBin containing one or more Doc objects. It's extremely efficient in storage, especially when packing multiple documents together. You can also create Doc objects manually, so you can write your own custom logic to convert and store existing annotations for use in spaCy.

### Training data from Doc objects {highlight="6-9"}
import spacy
from spacy.tokens import Doc, DocBin

nlp = spacy.blank("en")
docbin = DocBin()
words = ["Apple", "is", "looking", "at", "buying", "U.K.", "startup", "."]
spaces = [True, True, True, True, True, True, True, False]
ents = ["B-ORG", "O", "O", "O", "O", "B-GPE", "O", "O"]
doc = Doc(nlp.vocab, words=words, spaces=spaces, ents=ents)
docbin.add(doc)
docbin.to_disk("./train.spacy")

Working with corpora

Example

[corpora]

[corpora.train]
@readers = "spacy.Corpus.v1"
path = ${paths.train}
gold_preproc = false
max_length = 0
limit = 0
augmenter = null

[training]
train_corpus = "corpora.train"

The [corpora] block in your config lets you define data resources to use for training, evaluation, pretraining or any other custom workflows. corpora.train and corpora.dev are used as conventions within spaCy's default configs, but you can also define any other custom blocks. Each section in the corpora config should resolve to a Corpus – for example, using spaCy's built-in corpus reader that takes a path to a binary .spacy file. The train_corpus and dev_corpus fields in the [training] block specify where to find the corpus in your config. This makes it easy to swap out different corpora by only changing a single config setting.

Instead of making [corpora] a block with multiple subsections for each portion of the data, you can also use a single function that returns a dictionary of corpora, keyed by corpus name, e.g. "train" and "dev". This can be especially useful if you need to split a single file into corpora for training and evaluation, without loading the same file twice.

By default, the training data is loaded into memory and shuffled before each epoch. If the corpus is too large to fit into memory during training, stream the corpus using a custom reader as described in the next section.

Custom data reading and batching

Some use-cases require streaming in data or manipulating datasets on the fly, rather than generating all data beforehand and storing it to disk. Instead of using the built-in Corpus reader, which uses static file paths, you can create and register a custom function that generates Example objects.

In the following example we assume a custom function read_custom_data which loads or generates texts with relevant text classification annotations. Then, small lexical variations of the input text are created before generating the final Example objects. The @spacy.registry.readers decorator lets you register the function creating the custom reader in the readers registry and assign it a string name, so it can be used in your config. All arguments on the registered function become available as config settings – in this case, source.

config.cfg

[corpora.train]
@readers = "corpus_variants.v1"
source = "s3://your_bucket/path/data.csv"

### functions.py {highlight="7-8"}
from typing import Callable, Iterator, List
import spacy
from spacy.training import Example
from spacy.language import Language
import random

@spacy.registry.readers("corpus_variants.v1")
def stream_data(source: str) -> Callable[[Language], Iterator[Example]]:
    def generate_stream(nlp):
        for text, cats in read_custom_data(source):
            # Create a random variant of the example text
            i = random.randint(0, len(text) - 1)
            variant = text[:i] + text[i].upper() + text[i + 1:]
            doc = nlp.make_doc(variant)
            example = Example.from_dict(doc, {"cats": cats})
            yield example

    return generate_stream

Remember that a registered function should always be a function that spaCy calls to create something. In this case, it creates the reader function – it's not the reader itself.

If the corpus is too large to load into memory or the corpus reader is an infinite generator, use the setting max_epochs = -1 to indicate that the train corpus should be streamed. With this setting the train corpus is merely streamed and batched, not shuffled, so any shuffling needs to be implemented in the corpus reader itself. In the example below, a corpus reader that generates sentences containing even or odd numbers is used with an unlimited number of examples for the train corpus and a limited number of examples for the dev corpus. The dev corpus should always be finite and fit in memory during the evaluation step. max_steps and/or patience are used to determine when the training should stop.

config.cfg

[corpora.dev]
@readers = "even_odd.v1"
limit = 100

[corpora.train]
@readers = "even_odd.v1"
limit = -1

[training]
max_epochs = -1
patience = 500
max_steps = 2000

### functions.py
from typing import Callable, Iterable, Iterator
from spacy import util
import random
from spacy.training import Example
from spacy import Language


@util.registry.readers("even_odd.v1")
def create_even_odd_corpus(limit: int = -1) -> Callable[[Language], Iterable[Example]]:
    return EvenOddCorpus(limit)


class EvenOddCorpus:
    def __init__(self, limit):
        self.limit = limit

    def __call__(self, nlp: Language) -> Iterator[Example]:
        i = 0
        while i < self.limit or self.limit < 0:
            r = random.randint(0, 1000)
            cat = r % 2 == 0
            text = "This is sentence " + str(r)
            yield Example.from_dict(
                nlp.make_doc(text), {"cats": {"EVEN": cat, "ODD": not cat}}
            )
            i += 1

config.cfg

[initialize.components.textcat.labels]
@readers = "spacy.read_labels.v1"
path = "labels/textcat.json"
require = true

If the train corpus is streamed, the initialize step peeks at the first 100 examples in the corpus to find the labels for each component. If this isn't sufficient, you'll need to provide the labels for each component in the [initialize] block. init labels can be used to generate JSON files in the correct format, which you can extend with the full label set.

We can also customize the batching strategy by registering a new batcher function in the batchers registry. A batcher turns a stream of items into a stream of batches. spaCy has several useful built-in batching strategies with customizable sizes, but it's also easy to implement your own. For instance, the following function takes the stream of generated Example objects, and removes those which have the same underlying raw text, to avoid duplicates within each batch. Note that in a more realistic implementation, you'd also want to check whether the annotations are the same.

config.cfg

[training.batcher]
@batchers = "filtering_batch.v1"
size = 150

### functions.py
from typing import Callable, Iterable, Iterator, List
import spacy
from spacy.training import Example

@spacy.registry.batchers("filtering_batch.v1")
def filter_batch(size: int) -> Callable[[Iterable[Example]], Iterator[List[Example]]]:
    def create_filtered_batches(examples):
        batch = []
        for eg in examples:
            # Remove duplicate examples with the same text from batch
            if eg.text not in [x.text for x in batch]:
                batch.append(eg)
            if len(batch) == size:
                yield batch
                batch = []

    return create_filtered_batches

Data augmentation

Data augmentation is the process of applying small modifications to the training data. It can be especially useful for punctuation and case replacement – for example, if your corpus only uses smart quotes and you want to include variations using regular quotes, or to make the model less sensitive to capitalization by including a mix of capitalized and lowercase examples.

The easiest way to use data augmentation during training is to provide an augmenter to the training corpus, e.g. in the [corpora.train] section of your config. The built-in orth_variants augmenter creates a data augmentation callback that uses orth-variant replacement.

### config.cfg (excerpt) {highlight="8,14"}
[corpora.train]
@readers = "spacy.Corpus.v1"
path = ${paths.train}
gold_preproc = false
max_length = 0
limit = 0

[corpora.train.augmenter]
@augmenters = "spacy.orth_variants.v1"
# Percentage of texts that will be augmented / lowercased
level = 0.1
lower = 0.5

[corpora.train.augmenter.orth_variants]
@readers = "srsly.read_json.v1"
path = "corpus/orth_variants.json"

The orth_variants argument lets you pass in a dictionary of replacement rules, typically loaded from a JSON file. There are two types of orth variant rules: "single" for single tokens that should be replaced (e.g. hyphens) and "paired" for pairs of tokens (e.g. quotes).

### orth_variants.json
{
  "single": [{ "tags": ["NFP"], "variants": ["…", "..."] }],
  "paired": [{ "tags": ["``", "''"], "variants": [["'", "'"], ["‘", "’"]] }]
}

https://github.com/explosion/spacy-lookups-data/blob/master/spacy_lookups_data/data/en_orth_variants.json

https://github.com/explosion/spacy-lookups-data/blob/master/spacy_lookups_data/data/de_orth_variants.json

When adding data augmentation, keep in mind that it typically only makes sense to apply it to the training corpus, not the development data.

Writing custom data augmenters

Using the @spacy.augmenters registry, you can also register your own data augmentation callbacks. The callback should be a function that takes the current nlp object and a training Example and yields Example objects. Keep in mind that the augmenter should yield all examples you want to use in your corpus, not only the augmented examples (unless you want to augment all examples).

Here'a an example of a custom augmentation callback that produces text variants in "SpOnGeBoB cAsE". The registered function takes one argument randomize that can be set via the config and decides whether the uppercase/lowercase transformation is applied randomly or not. The augmenter yields two Example objects: the original example and the augmented example.

config.cfg

[corpora.train.augmenter]
@augmenters = "spongebob_augmenter.v1"
randomize = false

import spacy
import random

@spacy.registry.augmenters("spongebob_augmenter.v1")
def create_augmenter(randomize: bool = False):
    def augment(nlp, example):
        text = example.text
        if randomize:
            # Randomly uppercase/lowercase characters
            chars = [c.lower() if random.random() < 0.5 else c.upper() for c in text]
        else:
            # Uppercase followed by lowercase
            chars = [c.lower() if i % 2 else c.upper() for i, c in enumerate(text)]
        # Create augmented training example
        example_dict = example.to_dict()
        doc = nlp.make_doc("".join(chars))
        example_dict["token_annotation"]["ORTH"] = [t.text for t in doc]
        # Original example followed by augmented example
        yield example
        yield example.from_dict(doc, example_dict)

    return augment

An easy way to create modified Example objects is to use the Example.from_dict method with a new reference Doc created from the modified text. In this case, only the capitalization changes, so only the ORTH values of the tokens will be different between the original and augmented examples.

Note that if your data augmentation strategy involves changing the tokenization (for instance, removing or adding tokens) and your training examples include token-based annotations like the dependency parse or entity labels, you'll need to take care to adjust the Example object so its annotations match and remain valid.

Parallel & distributed training with Ray

Installation

$ pip install -U %%SPACY_PKG_NAME[ray]%%SPACY_PKG_FLAGS
# Check that the CLI is registered
$ python -m spacy ray --help

Ray is a fast and simple framework for building and running distributed applications. You can use Ray to train spaCy on one or more remote machines, potentially speeding up your training process. Parallel training won't always be faster though – it depends on your batch size, models, and hardware.

To use Ray with spaCy, you need the spacy-ray package installed. Installing the package will automatically add the ray command to the spaCy CLI.

The spacy ray train command follows the same API as spacy train, with a few extra options to configure the Ray setup. You can optionally set the --address option to point to your Ray cluster. If it's not set, Ray will run locally.

python -m spacy ray train config.cfg --n-workers 2

Get started with parallel training using our project template. It trains a simple model on a Universal Dependencies Treebank and lets you parallelize the training with Ray.

How parallel training works

Each worker receives a shard of the data and builds a copy of the model and optimizer from the config.cfg. It also has a communication channel to pass gradients and parameters to the other workers. Additionally, each worker is given ownership of a subset of the parameter arrays. Every parameter array is owned by exactly one worker, and the workers are given a mapping so they know which worker owns which parameter.

As training proceeds, every worker will be computing gradients for all of the model parameters. When they compute gradients for parameters they don't own, they'll send them to the worker that does own that parameter, along with a version identifier so that the owner can decide whether to discard the gradient. Workers use the gradients they receive and the ones they compute locally to update the parameters they own, and then broadcast the updated array and a new version ID to the other workers.

This training procedure is asynchronous and non-blocking. Workers always push their gradient increments and parameter updates, they do not have to pull them and block on the result, so the transfers can happen in the background, overlapped with the actual training work. The workers also do not have to stop and wait for each other ("synchronize") at the start of each batch. This is very useful for spaCy, because spaCy is often trained on long documents, which means batches can vary in size significantly. Uneven workloads make synchronous gradient descent inefficient, because if one batch is slow, all of the other workers are stuck waiting for it to complete before they can continue.

Internal training API

spaCy gives you full control over the training loop. However, for most use cases, it's recommended to train your pipelines via the spacy train command with a config.cfg to keep track of your settings and hyperparameters, instead of writing your own training scripts from scratch. Custom registered functions should typically give you everything you need to train fully custom pipelines with spacy train.

The Example object contains annotated training data, also called the gold standard. It's initialized with a Doc object that will hold the predictions, and another Doc object that holds the gold-standard annotations. It also includes the alignment between those two documents if they differ in tokenization. The Example class ensures that spaCy can rely on one standardized format that's passed through the pipeline. For instance, let's say we want to define gold-standard part-of-speech tags:

words = ["I", "like", "stuff"]
predicted = Doc(vocab, words=words)
# create the reference Doc with gold-standard TAG annotations
tags = ["NOUN", "VERB", "NOUN"]
tag_ids = [vocab.strings.add(tag) for tag in tags]
reference = Doc(vocab, words=words).from_array("TAG", numpy.array(tag_ids, dtype="uint64"))
example = Example(predicted, reference)

As this is quite verbose, there's an alternative way to create the reference Doc with the gold-standard annotations. The function Example.from_dict takes a dictionary with keyword arguments specifying the annotations, like tags or entities. Using the resulting Example object and its gold-standard annotations, the model can be updated to learn a sentence of three words with their assigned part-of-speech tags.

words = ["I", "like", "stuff"]
tags = ["NOUN", "VERB", "NOUN"]
predicted = Doc(nlp.vocab, words=words)
example = Example.from_dict(predicted, {"tags": tags})

Here's another example that shows how to define gold-standard named entities. The letters added before the labels refer to the tags of the BILUO scheme – O is a token outside an entity, U a single entity unit, B the beginning of an entity, I a token inside an entity and L the last token of an entity.

doc = Doc(nlp.vocab, words=["Facebook", "released", "React", "in", "2014"])
example = Example.from_dict(doc, {"entities": ["U-ORG", "O", "U-TECHNOLOGY", "O", "U-DATE"]})

As of v3.0, the Example object replaces the GoldParse class. It can be constructed in a very similar way – from a Doc and a dictionary of annotations. For more details, see the migration guide.

- gold = GoldParse(doc, entities=entities)
+ example = Example.from_dict(doc, {"entities": entities})

Of course, it's not enough to only show a model a single example once. Especially if you only have few examples, you'll want to train for a number of iterations. At each iteration, the training data is shuffled to ensure the model doesn't make any generalizations based on the order of examples. Another technique to improve the learning results is to set a dropout rate, a rate at which to randomly "drop" individual features and representations. This makes it harder for the model to memorize the training data. For example, a 0.25 dropout means that each feature or internal representation has a 1/4 likelihood of being dropped.

• nlp: The nlp object with the pipeline components and their models.
• nlp.initialize: Initialize the pipeline and return an optimizer to update the component model weights.
• Optimizer: Function that holds state between updates.
• nlp.update: Update component models with examples.
• Example: object holding predictions and gold-standard annotations.
• nlp.to_disk: Save the updated pipeline to a directory.

### Example training loop
optimizer = nlp.initialize()
for itn in range(100):
    random.shuffle(train_data)
    for raw_text, entity_offsets in train_data:
        doc = nlp.make_doc(raw_text)
        example = Example.from_dict(doc, {"entities": entity_offsets})
        nlp.update([example], sgd=optimizer)
nlp.to_disk("/output")

The nlp.update method takes the following arguments:

Name	Description
examples	Example objects. The update method takes a sequence of them, so you can batch up your training examples.
drop	Dropout rate. Makes it harder for the model to just memorize the data.
sgd	An Optimizer object, which updates the model's weights. If not set, spaCy will create a new one and save it for further use.

As of v3.0, the Example object replaces the GoldParse class and the "simple training style" of calling nlp.update with a text and a dictionary of annotations. Updating your code to use the Example object should be very straightforward: you can call Example.from_dict with a Doc and the dictionary of annotations:

text = "Facebook released React in 2014"
annotations = {"entities": ["U-ORG", "O", "U-TECHNOLOGY", "O", "U-DATE"]}
+ example = Example.from_dict(nlp.make_doc(text), annotations)
- nlp.update([text], [annotations])
+ nlp.update([example])