Embeddings, Transformers and Transfer Learning

Using transformer embeddings like BERT in spaCy

spaCy supports a number of transfer and multi-task learning workflows that can often help improve your pipeline’s efficiency or accuracy. Transfer learning refers to techniques such as word vector tables and language model pretraining. These techniques can be used to import knowledge from raw text into your pipeline, so that your models are able to generalize better from your annotated examples.

You can convert word vectors from popular tools like FastText and Gensim, or you can load in any pretrained transformer model if you install spacy-transformers. You can also do your own language model pretraining via the spacy pretrain command. You can even share your transformer or other contextual embedding model across multiple components, which can make long pipelines several times more efficient. To use transfer learning, you’ll need at least a few annotated examples for what you’re trying to predict. Otherwise, you could try using a “one-shot learning” approach using vectors and similarity.

Transformers are large and powerful neural networks that give you better accuracy, but are harder to deploy in production, as they require a GPU to run effectively. Word vectors are a slightly older technique that can give your models a smaller improvement in accuracy, and can also provide some additional capabilities.

The key difference between word-vectors and contextual language models such as transformers is that word vectors model lexical types, rather than tokens. If you have a list of terms with no context around them, a transformer model like BERT can’t really help you. BERT is designed to understand language in context, which isn’t what you have. A word vectors table will be a much better fit for your task. However, if you do have words in context – whole sentences or paragraphs of running text – word vectors will only provide a very rough approximation of what the text is about.

Word vectors are also very computationally efficient, as they map a word to a vector with a single indexing operation. Word vectors are therefore useful as a way to improve the accuracy of neural network models, especially models that are small or have received little or no pretraining. In spaCy, word vector tables are only used as static features. spaCy does not backpropagate gradients to the pretrained word vectors table. The static vectors table is usually used in combination with a smaller table of learned task-specific embeddings.

Word vectors are not compatible with most transformer models, but if you’re training another type of NLP network, it’s almost always worth adding word vectors to your model. As well as improving your final accuracy, word vectors often make experiments more consistent, as the accuracy you reach will be less sensitive to how the network is randomly initialized. High variance due to random chance can slow down your progress significantly, as you need to run many experiments to filter the signal from the noise.

Word vector features need to be enabled prior to training, and the same word vectors table will need to be available at runtime as well. You cannot add word vector features once the model has already been trained, and you usually cannot replace one word vectors table with another without causing a significant loss of performance.

Shared embedding layers

spaCy lets you share a single transformer or other token-to-vector (“tok2vec”) embedding layer between multiple components. You can even update the shared layer, performing multi-task learning. Reusing the tok2vec layer between components can make your pipeline run a lot faster and result in much smaller models. However, it can make the pipeline less modular and make it more difficult to swap components or retrain parts of the pipeline. Multi-task learning can affect your accuracy (either positively or negatively), and may require some retuning of your hyper-parameters.

Pipeline components using a shared embedding component vs. independent embedding layers
smaller: models only need to include a single copy of the embeddingslarger: models need to include the embeddings for each component
faster: embed the documents once for your whole pipelineslower: rerun the embedding for each component
less composable: all components require the same embedding component in the pipelinemodular: components can be moved and swapped freely

You can share a single transformer or other tok2vec model between multiple components by adding a Transformer or Tok2Vec component near the start of your pipeline. Components later in the pipeline can “connect” to it by including a listener layer like Tok2VecListener within their model.

Pipeline components listening to shared embedding component

At the beginning of training, the Tok2Vec component will grab a reference to the relevant listener layers in the rest of your pipeline. When it processes a batch of documents, it will pass forward its predictions to the listeners, allowing the listeners to reuse the predictions when they are eventually called. A similar mechanism is used to pass gradients from the listeners back to the model. The Transformer component and TransformerListener layer do the same thing for transformer models, but the Transformer component will also save the transformer outputs to the Doc._.trf_data extension attribute, giving you access to them after the pipeline has finished running.

Example: Shared vs. independent config

The config system lets you express model configuration for both shared and independent embedding layers. The shared setup uses a single Tok2Vec component with the Tok2Vec architecture. All other components, like the entity recognizer, use a Tok2VecListener layer as their model’s tok2vec argument, which connects to the tok2vec component model.


In the independent setup, the entity recognizer component defines its own Tok2Vec instance. Other components will do the same. This makes them fully independent and doesn’t require an upstream Tok2Vec component to be present in the pipeline.


Using transformer models

Transformers are a family of neural network architectures that compute dense, context-sensitive representations for the tokens in your documents. Downstream models in your pipeline can then use these representations as input features to improve their predictions. You can connect multiple components to a single transformer model, with any or all of those components giving feedback to the transformer to fine-tune it to your tasks. spaCy’s transformer support interoperates with PyTorch and the HuggingFace transformers library, giving you access to thousands of pretrained models for your pipelines. There are many great guides to transformer models, but for practical purposes, you can simply think of them as drop-in replacements that let you achieve higher accuracy in exchange for higher training and runtime costs.

Setup and installation

Once you have CUDA installed, we recommend installing PyTorch following the PyTorch installation guidelines for your package manager and CUDA version. If you skip this step, pip will install PyTorch as a dependency below, but it may not find the best version for your setup.

Example: Install PyTorch 1.11.0 for CUDA 11.3 with pip

Next, install spaCy with the extras for your CUDA version and transformers. The CUDA extra (e.g., cuda102, cuda113) installs the correct version of cupy, which is just like numpy, but for GPU. You may also need to set the CUDA_PATH environment variable if your CUDA runtime is installed in a non-standard location. Putting it all together, if you had installed CUDA 11.3 in /opt/nvidia/cuda, you would run:

Installation with CUDA

For transformers v4.0.0+ and models that require SentencePiece (e.g., ALBERT, CamemBERT, XLNet, Marian, and T5), install the additional dependencies with:

Install sentencepiece

Runtime usage

Transformer models can be used as drop-in replacements for other types of neural networks, so your spaCy pipeline can include them in a way that’s completely invisible to the user. Users will download, load and use the model in the standard way, like any other spaCy pipeline. Instead of using the transformers as subnetworks directly, you can also use them via the Transformer pipeline component.

The processing pipeline with the transformer component

The Transformer component sets the Doc._.trf_data extension attribute, which lets you access the transformers outputs at runtime. The trained transformer-based pipelines provided by spaCy end on _trf, e.g. en_core_web_trf.


You can also customize how the Transformer component sets annotations onto the Doc by specifying a custom set_extra_annotations function. This callback will be called with the raw input and output data for the whole batch, along with the batch of Doc objects, allowing you to implement whatever you need. The annotation setter is called with a batch of Doc objects and a FullTransformerBatch containing the transformers data for the batch.

Training usage

The recommended workflow for training is to use spaCy’s config system, usually via the spacy train command. The training config defines all component settings and hyperparameters in one place and lets you describe a tree of objects by referring to creation functions, including functions you register yourself. For details on how to get started with training your own model, check out the training quickstart.

The [components] section in the config.cfg describes the pipeline components and the settings used to construct them, including their model implementation. Here’s a config snippet for the Transformer component, along with matching Python code. In this case, the [components.transformer] block describes the transformer component:


The [components.transformer.model] block describes the model argument passed to the transformer component. It’s a Thinc Model object that will be passed into the component. Here, it references the function spacy-transformers.TransformerModel.v3 registered in the architectures registry. If a key in a block starts with @, it’s resolved to a function and all other settings are passed to the function as arguments. In this case, name, tokenizer_config and get_spans.

get_spans is a function that takes a batch of Doc objects and returns lists of potentially overlapping Span objects to process by the transformer. Several built-in functions are available – for example, to process the whole document or individual sentences. When the config is resolved, the function is created and passed into the model as an argument.

The name value is the name of any HuggingFace model, which will be downloaded automatically the first time it’s used. You can also use a local file path. For full details, see the TransformerModel docs.

A wide variety of PyTorch models are supported, but some might not work. If a model doesn’t seem to work feel free to open an issue. Additionally note that Transformers loaded in spaCy can only be used for tensors, and pretrained task-specific heads or text generation features cannot be used as part of the transformer pipeline component.

Customizing the settings

To change any of the settings, you can edit the config.cfg and re-run the training. To change any of the functions, like the span getter, you can replace the name of the referenced function – e.g. @span_getters = "spacy-transformers.sent_spans.v1" to process sentences. You can also register your own functions using the span_getters registry. For instance, the following custom function returns Span objects following sentence boundaries, unless a sentence succeeds a certain amount of tokens, in which case subsentences of at most max_length tokens are returned.

To resolve the config during training, spaCy needs to know about your custom function. You can make it available via the --code argument that can point to a Python file. For more details on training with custom code, see the training documentation.

Customizing the model implementations

The Transformer component expects a Thinc Model object to be passed in as its model argument. You’re not limited to the implementation provided by spacy-transformers – the only requirement is that your registered function must return an object of type Model[List[Doc],FullTransformerBatch]: that is, a Thinc model that takes a list of Doc objects, and returns a FullTransformerBatch object with the transformer data.

The same idea applies to task models that power the downstream components. Most of spaCy’s built-in model creation functions support a tok2vec argument, which should be a Thinc layer of type Model[List[Doc], List[Floats2d]]. This is where we’ll plug in our transformer model, using the TransformerListener layer, which sneakily delegates to the Transformer pipeline component.

config.cfg (excerpt)

The TransformerListener layer expects a pooling layer as the argument pooling, which needs to be of type Model[Ragged,Floats2d]. This layer determines how the vector for each spaCy token will be computed from the zero or more source rows the token is aligned against. Here we use the reduce_mean layer, which averages the wordpiece rows. We could instead use reduce_max, or a custom function you write yourself.

You can have multiple components all listening to the same transformer model, and all passing gradients back to it. By default, all of the gradients will be equally weighted. You can control this with the grad_factor setting, which lets you reweight the gradients from the different listeners. For instance, setting grad_factor = 0 would disable gradients from one of the listeners, while grad_factor = 2.0 would multiply them by 2. This is similar to having a custom learning rate for each component. Instead of a constant, you can also provide a schedule, allowing you to freeze the shared parameters at the start of training.

Static vectors

If your pipeline includes a word vectors table, you’ll be able to use the .similarity() method on the Doc, Span, Token and Lexeme objects. You’ll also be able to access the vectors using the .vector attribute, or you can look up one or more vectors directly using the Vocab object. Pipelines with word vectors can also use the vectors as features for the statistical models, which can improve the accuracy of your components.

Word vectors in spaCy are “static” in the sense that they are not learned parameters of the statistical models, and spaCy itself does not feature any algorithms for learning word vector tables. You can train a word vectors table using tools such as floret, Gensim, FastText or GloVe, or download existing pretrained vectors. The init vectors command lets you convert vectors for use with spaCy and will give you a directory you can load or refer to in your training configs.

Using word vectors in your models

Many neural network models are able to use word vector tables as additional features, which sometimes results in significant improvements in accuracy. spaCy’s built-in embedding layer, MultiHashEmbed, can be configured to use word vector tables using the include_static_vectors flag.

Creating a custom embedding layer

The MultiHashEmbed layer is spaCy’s recommended strategy for constructing initial word representations for your neural network models, but you can also implement your own. You can register any function to a string name, and then reference that function within your config (see the training docs for more details). To try this out, you can save the following little example to a new Python file:

If you pass the path to your file to the spacy train command using the --code argument, your file will be imported, which means the decorator registering the function will be run. Your function is now on equal footing with any of spaCy’s built-ins, so you can drop it in instead of any other model with the same input and output signature. For instance, you could use it in the tagger model as follows:

Now that you have a custom function wired into the network, you can start implementing the logic you’re interested in. For example, let’s say you want to try a relatively simple embedding strategy that makes use of static word vectors, but combines them via summation with a smaller table of learned embeddings.

Creating a custom vectors implementation v3.7

You can specify a custom registered vectors class under [nlp.vectors] in order to use static vectors in formats other than the ones supported by Vectors. Extend the abstract BaseVectors class to implement your custom vectors.

As an example, the following BPEmbVectors class implements support for BPEmb subword embeddings:

To use this in your pipeline, specify this registered function under [nlp.vectors] in your config:

Or specify it when creating a blank pipeline:

Remember to include this code with --code when using spacy train and spacy package.


The spacy pretrain command lets you initialize your models with information from raw text. Without pretraining, the models for your components will usually be initialized randomly. The idea behind pretraining is simple: random probably isn’t optimal, so if we have some text to learn from, we can probably find a way to get the model off to a better start.

Pretraining uses the same config.cfg file as the regular training, which helps keep the settings and hyperparameters consistent. The additional [pretraining] section has several configuration subsections that are familiar from the training block: the [pretraining.batcher], [pretraining.optimizer] and [pretraining.corpus] all work the same way and expect the same types of objects, although for pretraining your corpus does not need to have any annotations, so you will often use a different reader, such as the JsonlCorpus.

You can add a [pretraining] block to your config by setting the --pretraining flag on init config or init fill-config:

You can then run spacy pretrain with the updated config and pass in optional config overrides, like the path to the raw text file:

The following defaults are used for the [pretraining] block and merged into your existing config when you run init config or init fill-config with --pretraining. If needed, you can configure the settings and hyperparameters or change the objective.


How pretraining works

The impact of spacy pretrain varies, but it will usually be worth trying if you’re not using a transformer model and you have relatively little training data (for instance, fewer than 5,000 sentences). A good rule of thumb is that pretraining will generally give you a similar accuracy improvement to using word vectors in your model. If word vectors have given you a 10% error reduction, pretraining with spaCy might give you another 10%, for a 20% error reduction in total.

The spacy pretrain command will take a specific subnetwork within one of your components, and add additional layers to build a network for a temporary task that forces the model to learn something about sentence structure and word cooccurrence statistics.

Pretraining produces a binary weights file that can be loaded back in at the start of training, using the configuration option initialize.init_tok2vec. The weights file specifies an initial set of weights. Training then proceeds as normal.

You can only pretrain one subnetwork from your pipeline at a time, and the subnetwork must be typed Model[List[Doc], List[Floats2d]] (i.e. it has to be a “tok2vec” layer). The most common workflow is to use the Tok2Vec component to create a shared token-to-vector layer for several components of your pipeline, and apply pretraining to its whole model.

Configuring the pretraining

The spacy pretrain command is configured using the [pretraining] section of your config file. The component and layer settings tell spaCy how to find the subnetwork to pretrain. The layer setting should be either the empty string (to use the whole model), or a node reference. Most of spaCy’s built-in model architectures have a reference named "tok2vec" that will refer to the right layer.


Connecting pretraining to training

To benefit from pretraining, your training step needs to know to initialize its tok2vec component with the weights learned from the pretraining step. You do this by setting initialize.init_tok2vec to the filename of the .bin file that you want to use from pretraining.

A pretraining step that runs for 5 epochs with an output path of pretrain/, as an example, produces pretrain/model0.bin through pretrain/model4.bin plus a copy of the last iteration as pretrain/model-last.bin. Additionally, you can configure n_save_epoch to tell pretraining in which epoch interval it should save the current training progress. To use the final output to initialize your tok2vec layer, you could fill in this value in your config file:


Pretraining objectives

Two pretraining objectives are available, both of which are variants of the cloze task Devlin et al. (2018) introduced for BERT. The objective can be defined and configured via the [pretraining.objective] config block.

  • PretrainCharacters: The "characters" objective asks the model to predict some number of leading and trailing UTF-8 bytes for the words. For instance, setting n_characters = 2, the model will try to predict the first two and last two characters of the word.

  • PretrainVectors: The "vectors" objective asks the model to predict the word’s vector, from a static embeddings table. This requires a word vectors model to be trained and loaded. The vectors objective can optimize either a cosine or an L2 loss. We’ve generally found cosine loss to perform better.

These pretraining objectives use a trick that we term language modelling with approximate outputs (LMAO). The motivation for the trick is that predicting an exact word ID introduces a lot of incidental complexity. You need a large output layer, and even then, the vocabulary is too large, which motivates tokenization schemes that do not align to actual word boundaries. At the end of training, the output layer will be thrown away regardless: we just want a task that forces the network to model something about word cooccurrence statistics. Predicting leading and trailing characters does that more than adequately, as the exact word sequence could be recovered with high accuracy if the initial and trailing characters are predicted accurately. With the vectors objective, the pretraining uses the embedding space learned by an algorithm such as GloVe or Word2vec, allowing the model to focus on the contextual modelling we actual care about.