A model architecture is a function that wires up a Thinc Model instance. It describes the neural network that is run internally as part of a component in a spaCy pipeline. To define the actual architecture, you can implement your logic in Thinc directly, or you can use Thinc as a thin wrapper around frameworks such as PyTorch, TensorFlow and MXNet. Each Model can also be used as a sublayer of a larger network, allowing you to freely combine implementations from different frameworks into a single model.

spaCy’s built-in components require a Model instance to be passed to them via the config system. To change the model architecture of an existing component, you just need to update the config so that it refers to a different registered function. Once the component has been created from this config, you won’t be able to change it anymore. 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.

config.cfg (excerpt)

Type signatures

The Thinc Model class is a generic type that can specify its input and output types. Python uses a square-bracket notation for this, so the type Model[List, Dict] says that each batch of inputs to the model will be a list, and the outputs will be a dictionary. You can be even more specific and write for instanceModel[List[Doc], Dict[str, float]] to specify that the model expects a list of Doc objects as input, and returns a dictionary mapping of strings to floats. Some of the most common types you’ll see are: ​

See the Thinc type reference for details. The model type signatures help you figure out which model architectures and components can fit together. For instance, the TextCategorizer class expects a model typed Model[List[Doc],Floats2d], because the model will predict one row of category probabilities per Doc. In contrast, the Tagger class expects a model typed Model[List[Doc], List[Floats2d]], because it needs to predict one row of probabilities per token.

There’s no guarantee that two models with the same type signature can be used interchangeably. There are many other ways they could be incompatible. However, if the types don’t match, they almost surely won’t be compatible. This little bit of validation goes a long way, especially if you configure your editor or other tools to highlight these errors early. The config file is also validated at the beginning of training, to verify that all the types match correctly.

Tip: Static type checking in your editor

If you’re using a modern editor like Visual Studio Code, you can set up mypy with the custom Thinc plugin and get live feedback about mismatched types as you write code.

Swapping model architectures

If no model is specified for the TextCategorizer, the TextCatEnsemble architecture is used by default. This architecture combines a simple bag-of-words model with a neural network, usually resulting in the most accurate results, but at the cost of speed. The config file for this model would look something like this:

config.cfg (excerpt)

spaCy has two additional built-in textcat architectures, and you can easily use those by swapping out the definition of the textcat’s model. For instance, to use the simple and fast bag-of-words model TextCatBOW, you can change the config to:

config.cfg (excerpt)

For details on all pre-defined architectures shipped with spaCy and how to configure them, check out the model architectures documentation.

Defining sublayers

Model architecture functions often accept sublayers as arguments, so that you can try substituting a different layer into the network. Depending on how the architecture function is structured, you might be able to define your network structure entirely through the config system, using layers that have already been defined. ​

In most neural network models for NLP, the most important parts of the network are what we refer to as the embed and encode steps. These steps together compute dense, context-sensitive representations of the tokens, and their combination forms a typical Tok2Vec layer:

config.cfg (excerpt)

By defining these sublayers specifically, it becomes straightforward to swap out a sublayer for another one, for instance changing the first sublayer to a character embedding with the CharacterEmbed architecture:

config.cfg (excerpt)

Most of spaCy’s default architectures accept a tok2vec layer as a sublayer within the larger task-specific neural network. This makes it easy to switch between transformer, CNN, BiLSTM or other feature extraction approaches. The transformers documentation section shows an example of swapping out a model’s standard tok2vec layer with a transformer. And if you want to define your own solution, all you need to do is register a Model[List[Doc], List[Floats2d]] architecture function, and you’ll be able to try it out in any of the spaCy components. ​

Wrapping PyTorch, TensorFlow and other frameworks

Thinc allows you to wrap models written in other machine learning frameworks like PyTorch, TensorFlow and MXNet using a unified Model API. This makes it easy to use a model implemented in a different framework to power a component in your spaCy pipeline. For example, to wrap a PyTorch model as a Thinc Model, you can use Thinc’s PyTorchWrapper:

Let’s use PyTorch to define a very simple neural network consisting of two hidden Linear layers with ReLU activation and dropout, and a softmax-activated output layer:

PyTorch model

The resulting wrapped Model can be used as a custom architecture as such, or can be a subcomponent of a larger model. For instance, we can use Thinc’s chain combinator, which works like Sequential in PyTorch, to combine the wrapped model with other components in a larger network. This effectively means that you can easily wrap different components from different frameworks, and “glue” them together with Thinc:

In the above example, we have combined our custom PyTorch model with a character embedding layer defined by spaCy. CharacterEmbed returns a Model that takes a List[Doc] as input, and outputs a List[Floats2d]. To make sure that the wrapped PyTorch model receives valid inputs, we use Thinc’s with_array helper.

You could also implement a model that only uses PyTorch for the transformer layers, and “native” Thinc layers to do fiddly input and output transformations and add on task-specific “heads”, as efficiency is less of a consideration for those parts of the network.

Using wrapped models

To use our custom model including the PyTorch subnetwork, all we need to do is register the architecture using the architectures registry. This assigns the architecture a name so spaCy knows how to find it, and allows passing in arguments like hyperparameters via the config. The full example then becomes:

Registering the architecture

The model definition can now be used in any existing trainable spaCy component, by specifying it in the config file. In this configuration, all required parameters for the various subcomponents of the custom architecture are passed in as settings via the config.

config.cfg (excerpt)

Note that when using a PyTorch or Tensorflow model, it is recommended to set the GPU memory allocator accordingly. When gpu_allocator is set to “pytorch” or “tensorflow” in the training config, cupy will allocate memory via those respective libraries, preventing OOM errors when there’s available memory sitting in the other library’s pool.

config.cfg (excerpt)

Custom models with Thinc

Of course it’s also possible to define the Model from the previous section entirely in Thinc. The Thinc documentation provides details on the various layers and helper functions available. Combinators can be used to overload operators and a common usage pattern is to bind chain to >>. The “native” Thinc version of our simple neural network would then become:

Shape inference in Thinc

It is not strictly necessary to define all the input and output dimensions for each layer, as Thinc can perform shape inference between sequential layers by matching up the output dimensionality of one layer to the input dimensionality of the next. This means that we can simplify the layers definition:

Thinc can even go one step further and deduce the correct input dimension of the first layer, and output dimension of the last. To enable this functionality, you have to call Model.initialize with an input sample X and an output sample Y with the correct dimensions:

Shape inference with initialization

The built-in pipeline components in spaCy ensure that their internal models are always initialized with appropriate sample data. In this case, X is typically a List[Doc], while Y is typically a List[Array1d] or List[Array2d], depending on the specific task. This functionality is triggered when nlp.initialize is called.

Dropout and normalization in Thinc

Many of the available Thinc layers allow you to define a dropout argument that will result in “chaining” an additional Dropout layer. Optionally, you can often specify whether or not you want to add layer normalization, which would result in an additional LayerNorm layer. That means that the following layers definition is equivalent to the previous:

Create new trainable components

In addition to swapping out layers in existing components, you can also implement an entirely new, trainable pipeline component from scratch. This can be done by creating a new class inheriting from TrainablePipe, and linking it up to your custom model implementation.

Example: Entity relation extraction component

This section outlines an example use-case of implementing a novel relation extraction component from scratch. We’ll implement a binary relation extraction method that determines whether or not two entities in a document are related, and if so, what type of relation connects them. We allow multiple types of relations between two such entities (a multi-label setting). There are two major steps required:

1. Implement a machine learning model specific to this task. It will have to extract candidate relation instances from a Doc and predict the corresponding scores for each relation label.
2. Implement a custom pipeline component - powered by the machine learning model from step 1 - that translates the predicted scores into annotations that are stored on the Doc objects as they pass through the nlp pipeline.

Step 1: Implementing the Model

We need to implement a Model that takes a list of documents (List[Doc]) as input, and outputs a two-dimensional matrix (Floats2d) of predictions:

The model architecture

We adapt a modular approach to the definition of this relation model, and define it as chaining two layers together: the first layer that generates an instance tensor from a given set of documents, and the second layer that transforms the instance tensor into a final tensor holding the predictions:

The model architecture

The classification_layer could be something like a Linear layer followed by a logistic activation function:

The classification layer

The first layer that creates the instance tensor can be defined by implementing a custom forward function with an appropriate backpropagation callback. We also define an initialization method that ensures that the layer is properly set up for training.

We omit some of the implementation details here, and refer to the spaCy project that has the full implementation.

The layer that creates the instance tensor

This custom layer uses an embedding layer such as a Tok2Vec component or a Transformer. This layer is assumed to be of type Model[List[Doc], List[Floats2d]] as it transforms each document into a list of tokens, with each token being represented by its embedding in the vector space.

The pooling layer will be applied to summarize the token vectors into entity vectors, as named entities (represented by Span objects) can consist of one or multiple tokens. For instance, the pooling layer could resort to calculating the average of all token vectors in an entity. Thinc provides several built-in pooling operators for this purpose.

Finally, we need a get_instances method that generates pairs of entities that we want to classify as being related or not. As these candidate pairs are typically formed within one document, this function takes a Doc as input and outputs a List of Span tuples. For instance, the following implementation takes any two entities from the same document, as long as they are within a maximum distance (in number of tokens) of each other:

Candidate generation

This function is added to the @misc registry so we can refer to it from the config, and easily swap it out for any other candidate generation function.

Intermezzo: define how to store the relations data

For our new relation extraction component, we will use a custom extension attribute doc._.rel in which we store relation data. The attribute refers to a dictionary, keyed by the start offsets of each entity involved in the candidate relation. The values in the dictionary refer to another dictionary where relation labels are mapped to values between 0 and 1. We assume anything above 0.5 to be a True relation. The Example instances that we’ll use as training data, will include their gold-standard relation annotations in example.reference._.rel.

Registering the extension attribute

Step 2: Implementing the pipeline component

To use our new relation extraction model as part of a custom trainable component, we create a subclass of TrainablePipe that holds the model.

Pipeline component skeleton

Typically, the constructor defines the vocab, the Machine Learning model, and the name of this component. Additionally, this component, just like the textcat and the tagger, stores an internal list of labels. The ML model will predict scores for each label. We add convenience methods to easily retrieve and add to them.

The constructor (continued)

After creation, the component needs to be initialized. This method can define the relevant labels in two ways: explicitly by setting the labels argument in the initialize block of the config, or implicately by deducing them from the get_examples callback that generates the full training data set, or a representative sample.

The final number of labels defines the output dimensionality of the network, and will be used to do shape inference throughout the layers of the neural network. This is triggered by calling Model.initialize.

The initialize method

The initialize method is triggered whenever this component is part of an nlp pipeline, and nlp.initialize is invoked. Typically, this happens when the pipeline is set up before training in spacy train. After initialization, the pipeline component and its internal model can be trained and used to make predictions.

During training, the method update is invoked which delegates to Model.begin_update and a get_loss function that calculates the loss for a batch of examples, as well as the gradient of loss that will be used to update the weights of the model layers. Thinc provides several loss functions that can be used for the implementation of the get_loss function.

The update method

After training the model, the component can be used to make novel predictions. The predict method needs to be implemented for each subclass of TrainablePipe. In our case, we can simply delegate to the internal model’s predict function that takes a batch of Doc objects and returns a Floats2d array:

The predict method

The final method that needs to be implemented, is set_annotations. This function takes the predictions, and modifies the given Doc object in place to store them. For our relation extraction component, we store the data in the custom attributedoc._.rel.

To interpret the scores predicted by the relation extraction model correctly, we need to refer to the model’s get_instances function that defined which pairs of entities were relevant candidates, so that the predictions can be linked to those exact entities:

The set_annotations method

Under the hood, when the pipe is applied to a document, it delegates to the predict and set_annotations methods:

The __call__ method

There is one more optional method to implement: score calculates the performance of your component on a set of examples, and returns the results as a dictionary:

The score method

This is particularly useful for calculating relevant scores on the development corpus when training the component with spacy train.

Once our TrainablePipe subclass is fully implemented, we can register the component with the @Language.factory decorator. This assigns it a name and lets you create the component with nlp.add_pipe and via the config.

Registering the pipeline component

You can extend the decorator to include information such as the type of annotations that are required for this component to run, the type of annotations it produces, and the scores that can be calculated:

Factory annotations