spaCy/website/docs/usage/embeddings-transformers.md

title	teaser	menu	next
Embeddings, Transformers and Transfer Learning	Using transformer embeddings like BERT in spaCy	Embedding Layers embedding-layers Transformers transformers Static Vectors static-vectors Pretraining pretraining	/usage/training

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.

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.

Shared	Independent
✅ smaller: models only need to include a single copy of the embeddings	❌ larger: models need to include the embeddings for each component
✅ faster: embed the documents once for your whole pipeline	❌ slower: rerun the embedding for each component
❌ less composable: all components require the same embedding component in the pipeline	✅ modular: 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.

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.

### Shared {highlight="1-2,4-5,19-20"}
[components.tok2vec]
factory = "tok2vec"

[components.tok2vec.model]
@architectures = "spacy.Tok2Vec.v1"

[components.tok2vec.model.embed]
@architectures = "spacy.MultiHashEmbed.v1"

[components.tok2vec.model.encode]
@architectures = "spacy.MaxoutWindowEncoder.v1"

[components.ner]
factory = "ner"

[components.ner.model]
@architectures = "spacy.TransitionBasedParser.v1"

[components.ner.model.tok2vec]
@architectures = "spacy.Tok2VecListener.v1"

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.

### Independent {highlight="7-8"}
[components.ner]
factory = "ner"

[components.ner.model]
@architectures = "spacy.TransitionBasedParser.v1"

[components.ner.model.tok2vec]
@architectures = "spacy.Tok2Vec.v1"

[components.ner.model.tok2vec.embed]
@architectures = "spacy.MultiHashEmbed.v1"

[components.ner.model.tok2vec.encode]
@architectures = "spacy.MaxoutWindowEncoder.v1"

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

System requirements

We recommend an NVIDIA GPU with at least 10GB of memory in order to work with transformer models. Make sure your GPU drivers are up to date and you have CUDA v9+ installed.

The exact requirements will depend on the transformer model. Training a transformer-based model without a GPU will be too slow for most practical purposes.

Provisioning a new machine will require about 5GB of data to be downloaded: 3GB CUDA runtime, 800MB PyTorch, 400MB CuPy, 500MB weights, 200MB spaCy and dependencies.

Once you have CUDA installed, you'll need to install two pip packages, cupy and spacy-transformers. cupy is just like numpy, but for GPU. The best way to install it is to choose a wheel that matches the version of CUDA you're using. 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 10.2 in /opt/nvidia/cuda, you would run:

### Installation with CUDA
$ export CUDA_PATH="/opt/nvidia/cuda"
$ pip install cupy-cuda102
$ pip install spacy-transformers

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.

By default, the Transformer component sets the Doc._.trf_data extension attribute, which lets you access the transformers outputs at runtime.

$ python -m spacy download en_core_trf_lg

### Example
import spacy
from thinc.api import use_pytorch_for_gpu_memory, require_gpu

# Use the GPU, with memory allocations directed via PyTorch.
# This prevents out-of-memory errors that would otherwise occur from competing
# memory pools.
use_pytorch_for_gpu_memory()
require_gpu(0)

nlp = spacy.load("en_core_trf_lg")
for doc in nlp.pipe(["some text", "some other text"]):
    tokvecs = doc._.trf_data.tensors[-1]

You can customize how the Transformer component sets annotations onto the Doc, by changing the annotation_setter. 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.

def custom_annotation_setter(docs, trf_data):
    doc_data = list(trf_data.doc_data)
    for doc, data in zip(docs, doc_data):
        doc._.custom_attr = data

nlp = spacy.load("en_core_trf_lg")
nlp.get_pipe("transformer").annotation_setter = custom_annotation_setter
doc = nlp("This is a text")
assert isinstance(doc._.custom_attr, TransformerData)
print(doc._.custom_attr.tensors)

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:

Python equivalent

from spacy_transformers import Transformer, TransformerModel
from spacy_transformers.annotation_setters import configure_trfdata_setter
from spacy_transformers.span_getters import get_doc_spans

trf = Transformer(
    nlp.vocab,
    TransformerModel(
        "bert-base-cased",
        get_spans=get_doc_spans,
        tokenizer_config={"use_fast": True},
    ),
    annotation_setter=configure_trfdata_setter(),
    max_batch_items=4096,
)

### config.cfg (excerpt)
[components.transformer]
factory = "transformer"
max_batch_items = 4096

[components.transformer.model]
@architectures = "spacy-transformers.TransformerModel.v1"
name = "bert-base-cased"
tokenizer_config = {"use_fast": true}

[components.transformer.model.get_spans]
@span_getters = "doc_spans.v1"

[components.transformer.annotation_setter]
@annotation_setters = "spacy-transformers.trfdata_setter.v1"

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.v1 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.

Remember that the config.cfg used for training should contain no missing values and requires all settings to be defined. You don't want any hidden defaults creeping in and changing your results! spaCy will tell you if settings are missing, and you can run spacy init fill-config to automatically fill in all defaults.

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 = "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.

config.cfg

[components.transformer.model.get_spans]
@span_getters = "custom_sent_spans"
max_length = 25

### code.py
import spacy_transformers

@spacy_transformers.registry.span_getters("custom_sent_spans")
def configure_custom_sent_spans(max_length: int):
    def get_custom_sent_spans(docs):
        spans = []
        for doc in docs:
            spans.append([])
            for sent in doc.sents:
                start = 0
                end = max_length
                while end <= len(sent):
                    spans[-1].append(sent[start:end])
                    start += max_length
                    end += max_length
                if start < len(sent):
                    spans[-1].append(sent[start:len(sent)])
        return spans

    return get_custom_sent_spans

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.

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

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) {highlight="12"}
[components.ner]
factory = "ner"

[nlp.pipeline.ner.model]
@architectures = "spacy.TransitionBasedParser.v1"
nr_feature_tokens = 3
hidden_width = 128
maxout_pieces = 3
use_upper = false

[nlp.pipeline.ner.model.tok2vec]
@architectures = "spacy-transformers.TransformerListener.v1"
grad_factor = 1.0

[nlp.pipeline.ner.model.tok2vec.pooling]
@layers = "reduce_mean.v1"

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

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 also_use_static_vectors flag. This setting is also available on the MultiHashEmbedCNN layer, which builds the default token-to-vector encoding architecture.

[tagger.model.tok2vec.embed]
@architectures = "spacy.MultiHashEmbed.v1"
width = 128
rows = 7000
also_embed_subwords = true
also_use_static_vectors = true

The configuration system will look up the string "spacy.MultiHashEmbed.v1" in the architectures registry, and call the returned object with the rest of the arguments from the block. This will result in a call to the MultiHashEmbed function, which will return a Thinc model object with the type signature Model[List[Doc], List[Floats2d]]. Because the embedding layer takes a list of Doc objects as input, it does not need to store a copy of the vectors table. The vectors will be retrieved from the Doc objects that are passed in, via the doc.vocab.vectors attribute. This part of the process is handled by the StaticVectors layer.

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:

from spacy.ml.staticvectors import StaticVectors
from spacy.util import registry

print("I was imported!")

@registry.architectures("my_example.MyEmbedding.v1")
def MyEmbedding(output_width: int) -> Model[List[Doc], List[Floats2d]]:
    print("I was called!")
    return StaticVectors(nO=output_width)

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:

[tagger.model.tok2vec.embed]
@architectures = "my_example.MyEmbedding.v1"
output_width = 128

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.

from thinc.api import add, chain, remap_ids, Embed, FeatureExtractor
from spacy.ml.staticvectors import StaticVectors
from spacy.util import registry

@registry.architectures("my_example.MyEmbedding.v1")
def MyCustomVectors(
    output_width: int,
    vector_width: int,
    embed_rows: int,
    key2row: Dict[int, int]
) -> Model[List[Doc], List[Floats2d]]:
    return add(
        StaticVectors(nO=output_width),
        chain(
           FeatureExtractor(["ORTH"]),
           remap_ids(key2row),
           Embed(nO=output_width, nV=embed_rows)
        )
    )

Pretraining

Raw text format

The raw text can be provided as JSONL (newline-delimited JSON) with a key "text" per entry. This allows the data to be read in line by line, while also allowing you to include newlines in the texts.

{"text": "Can I ask where you work now and what you do, and if you enjoy it?"}
{"text": "They may just pull out of the Seattle market completely, at least until they have autonomous vehicles."}

$ python -m spacy init fill-config config.cfg config_pretrain.cfg --pretraining

$ python -m spacy pretrain raw_text.jsonl /output config_pretrain.cfg