spaCy/website/docs/usage/training.md

title	next	menu
Training spaCy's Statistical Models	/usage/adding-languages	Basics basics NER ner Tagger & Parser tagger-parser Text Classification textcat Entity Linking entity-linker Tips and Advice tips

This guide describes how to train new statistical models for spaCy's part-of-speech tagger, named entity recognizer, dependency parser, text classifier and entity linker. Once the model is trained, you can then save and load it.

Training basics

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

Training via the command-line interface

For most purposes, the best way to train spaCy is via the command-line interface. The spacy train command takes care of many details for you, including making sure that the data is minibatched and shuffled correctly, progress is printed, and models are saved after each epoch. You can prepare your data for use in spacy train using the spacy convert command, which accepts many common NLP data formats, including .iob for named entities, and the CoNLL format for dependencies:

git clone https://github.com/UniversalDependencies/UD_Spanish-AnCora
mkdir ancora-json
python -m spacy convert UD_Spanish-AnCora/es_ancora-ud-train.conllu ancora-json
python -m spacy convert UD_Spanish-AnCora/es_ancora-ud-dev.conllu ancora-json
mkdir models
python -m spacy train es models ancora-json/es_ancora-ud-train.json ancora-json/es_ancora-ud-dev.json

If you're running spaCy v2.2 or above, you can use the debug-data command to 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 en train.json dev.json --verbose

You can also use the gold.docs_to_json helper to convert a list of Doc objects to spaCy's JSON training format.

Understanding the training output

When you train a model using the spacy train command, you'll see a table showing metrics after each pass over the data. Here's what those metrics means:

Tokenization metrics

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.

Name	Description
Dep Loss	Training loss for dependency parser. Should decrease, but usually not to 0.
NER Loss	Training loss for named entity recognizer. Should decrease, but usually not to 0.
UAS	Unlabeled attachment score for parser. The percentage of unlabeled correct arcs. Should increase.
NER P.	NER precision on development data. Should increase.
NER R.	NER recall on development data. Should increase.
NER F.	NER F-score on development data. Should increase.
Tag %	Fine-grained part-of-speech tag accuracy on development data. Should increase.
Token %	Tokenization accuracy on development data.
CPU WPS	Prediction speed on CPU in words per second, if available. Should stay stable.
GPU WPS	Prediction speed on GPU in words per second, if available. Should stay stable.

Improving accuracy with transfer learning

In most projects, you'll usually have a small amount of labelled data, and access to a much bigger sample of raw text. The raw text contains a lot of information about the language in general. Learning this general information from the raw text can help your model use the smaller labelled data more efficiently.

The two main ways to use raw text in your spaCy models are word vectors and language model pretraining. Word vectors provide information about the definitions of words. The vectors are a look-up table, so each word only has one representation, regardless of its context. Language model pretraining lets you learn contextualized word representations. Instead of initializing spaCy's convolutional neural network layers with random weights, the spacy pretrain command trains a language model to predict each word's word vector based on the surrounding words. The information used to predict this task is a good starting point for other tasks such as named entity recognition, text classification or dependency parsing.

For more details, see the documentation on vectors and similarity and the spacy pretrain command.

How do I get training data?

Collecting training data may sound incredibly painful – and it can be, if you're planning a large-scale annotation project. However, if your main goal is to update an existing model's predictions – for example, spaCy's named entity recognition – the hard part is usually not creating the actual annotations. It's finding representative examples and extracting potential candidates. The good news is, if you've been noticing bad performance on your data, you likely already have some relevant text, and you can use spaCy to bootstrap a first set of training examples. For example, after processing a few sentences, you may end up with the following entities, some correct, some incorrect.

How many examples do I need?

As a rule of thumb, you should allocate at least 10% of your project resources to creating training and evaluation data. If you're looking to improve an existing model, you might be able to start off with only a handful of examples. Keep in mind that you'll always want a lot more than that for evaluation – especially previous errors the model has made. Otherwise, you won't be able to sufficiently verify that the model has actually made the correct generalizations required for your use case.

Text	Entity	Start	End	Label
Uber blew through 1 million a week	Uber	0	4	ORG	✅
Android Pay expands to Canada	Android	0	7	PERSON	❌
Android Pay expands to Canada	Canada	23	30	GPE	✅
Spotify steps up Asia expansion	Spotify	0	8	ORG	✅
Spotify steps up Asia expansion	Asia	17	21	NORP	❌

Alternatively, the rule-based matcher can be a useful tool to extract tokens or combinations of tokens, as well as their start and end index in a document. In this case, we'll extract mentions of Google and assume they're an ORG.

Text	Entity	Start	End	Label
let me google this for you	google	7	13	ORG	❌
Google Maps launches location sharing	Google	0	6	ORG	❌
Google rebrands its business apps	Google	0	6	ORG	✅
look what i found on google! 😂	google	21	27	ORG	✅

Based on the few examples above, you can already create six training sentences with eight entities in total. Of course, what you consider a "correct annotation" will always depend on what you want the model to learn. While there are some entity annotations that are more or less universally correct – like Canada being a geopolitical entity – your application may have its very own definition of the NER annotation scheme.

train_data = [
    ("Uber blew through $1 million a week", [(0, 4, 'ORG')]),
    ("Android Pay expands to Canada", [(0, 11, 'PRODUCT'), (23, 30, 'GPE')]),
    ("Spotify steps up Asia expansion", [(0, 8, "ORG"), (17, 21, "LOC")]),
    ("Google Maps launches location sharing", [(0, 11, "PRODUCT")]),
    ("Google rebrands its business apps", [(0, 6, "ORG")]),
    ("look what i found on google! 😂", [(21, 27, "PRODUCT")])]

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 models.

Training with annotations

The GoldParse object collects the annotated training examples, also called the gold standard. It's initialized with the Doc object it refers to, and keyword arguments specifying the annotations, like tags or entities. Its job is to encode the annotations, keep them aligned and create the C-level data structures required for efficient access. Here's an example of a simple GoldParse for part-of-speech tags:

vocab = Vocab(tag_map={"N": {"pos": "NOUN"}, "V": {"pos": "VERB"}})
doc = Doc(vocab, words=["I", "like", "stuff"])
gold = GoldParse(doc, tags=["N", "V", "N"])

Using the Doc and its gold-standard annotations, the model can be updated to learn a sentence of three words with their assigned part-of-speech tags. The tag map is part of the vocabulary and defines the annotation scheme. If you're training a new language model, this will let you map the tags present in the treebank you train on to spaCy's tag scheme.

doc = Doc(Vocab(), words=["Facebook", "released", "React", "in", "2014"])
gold = GoldParse(doc, entities=["U-ORG", "O", "U-TECHNOLOGY", "O", "U-DATE"])

The same goes for named entities. The letters added before the labels refer to the tags of the BILUO scheme – O is a token outside an entity, U an single entity unit, B the beginning of an entity, I a token inside an entity and L the last token of an entity.

• Training data: The training examples.
• Text and label: The current example.
• Doc: A Doc object created from the example text.
• GoldParse: A GoldParse object of the Doc and label.
• nlp: The nlp object with the model.
• Optimizer: A function that holds state between updates.
• Update: Update the model's weights.

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.

• begin_training(): Start the training and return an optimizer function to update the model's weights. Can take an optional function converting the training data to spaCy's training format.
• update(): Update the model with the training example and gold data.
• to_disk(): Save the updated model to a directory.

### Example training loop
optimizer = nlp.begin_training(get_data)
for itn in range(100):
    random.shuffle(train_data)
    for raw_text, entity_offsets in train_data:
        doc = nlp.make_doc(raw_text)
        gold = GoldParse(doc, entities=entity_offsets)
        nlp.update([doc], [gold], drop=0.5, sgd=optimizer)
nlp.to_disk("/model")

The nlp.update method takes the following arguments:

Name	Description
docs	Doc objects. The update method takes a sequence of them, so you can batch up your training examples. Alternatively, you can also pass in a sequence of raw texts.
golds	GoldParse objects. The update method takes a sequence of them, so you can batch up your training examples. Alternatively, you can also pass in a dictionary containing the annotations.
drop	Dropout rate. Makes it harder for the model to just memorize the data.
sgd	An optimizer, i.e. a callable to update the model's weights. If not set, spaCy will create a new one and save it for further use.

Instead of writing your own training loop, you can also use the built-in train command, which expects data in spaCy's JSON format. On each epoch, a model will be saved out to the directory. After training, you can use the package command to generate an installable Python package from your model.

python -m spacy convert /tmp/train.conllu /tmp/data
python -m spacy train en /tmp/model /tmp/data/train.json -n 5

Simple training style

Instead of sequences of Doc and GoldParse objects, you can also use the "simple training style" and pass raw texts and dictionaries of annotations to nlp.update. The dictionaries can have the keys entities, heads, deps, tags and cats. This is generally recommended, as it removes one layer of abstraction, and avoids unnecessary imports. It also makes it easier to structure and load your training data.

Example Annotations

{
   "entities": [(0, 4, "ORG")],
   "heads": [1, 1, 1, 5, 5, 2, 7, 5],
   "deps": ["nsubj", "ROOT", "prt", "quantmod", "compound", "pobj", "det", "npadvmod"],
   "tags": ["PROPN", "VERB", "ADP", "SYM", "NUM", "NUM", "DET", "NOUN"],
   "cats": {"BUSINESS": 1.0},
}

### Simple training loop
TRAIN_DATA = [
        ("Uber blew through $1 million a week", {"entities": [(0, 4, "ORG")]}),
        ("Google rebrands its business apps", {"entities": [(0, 6, "ORG")]})]

nlp = spacy.blank("en")
optimizer = nlp.begin_training()
for i in range(20):
    random.shuffle(TRAIN_DATA)
    for text, annotations in TRAIN_DATA:
        nlp.update([text], [annotations], sgd=optimizer)
nlp.to_disk("/model")

The above training loop leaves out a few details that can really improve accuracy – but the principle really is that simple. Once you've got your pipeline together and you want to tune the accuracy, you usually want to process your training examples in batches, and experiment with minibatch sizes and dropout rates, set via the drop keyword argument. See the Language and Pipe API docs for available options.

Training the named entity recognizer

All spaCy models support online learning, so you can update a pretrained model with new examples. You'll usually need to provide many examples to meaningfully improve the system — a few hundred is a good start, although more is better.

You should avoid iterating over the same few examples multiple times, or the model is likely to "forget" how to annotate other examples. If you iterate over the same few examples, you're effectively changing the loss function. The optimizer will find a way to minimize the loss on your examples, without regard for the consequences on the examples it's no longer paying attention to. One way to avoid this "catastrophic forgetting" problem is to "remind" the model of other examples by augmenting your annotations with sentences annotated with entities automatically recognized by the original model. Ultimately, this is an empirical process: you'll need to experiment on your data to find a solution that works best for you.

Tip: Converting entity annotations

You can train the entity recognizer with entity offsets or annotations in the BILUO scheme. The spacy.gold module also exposes two helper functions to convert offsets to BILUO tags, and BILUO tags to entity offsets.

Updating the Named Entity Recognizer

This example shows how to update spaCy's entity recognizer with your own examples, starting off with an existing, pretrained model, or from scratch using a blank Language class. To do this, you'll need example texts and the character offsets and labels of each entity contained in the texts.

https://github.com/explosion/spaCy/tree/master/examples/training/train_ner.py

Step by step guide

1. Load the model you want to start with, or create an empty model using spacy.blank with the ID of your language. If you're using a blank model, don't forget to add the entity recognizer to the pipeline. If you're using an existing model, make sure to disable all other pipeline components during training using nlp.select_pipes. This way, you'll only be training the entity recognizer.
2. Shuffle and loop over the examples. For each example, update the model by calling nlp.update, which steps through the words of the input. At each word, it makes a prediction. It then consults the annotations to see whether it was right. If it was wrong, it adjusts its weights so that the correct action will score higher next time.
3. Save the trained model using nlp.to_disk.
4. Test the model to make sure the entities in the training data are recognized correctly.

Training an additional entity type

This script shows how to add a new entity type ANIMAL to an existing pretrained NER model, or an empty Language class. To keep the example short and simple, only a few sentences are provided as examples. In practice, you'll need many more — a few hundred would be a good start. You will also likely need to mix in examples of other entity types, which might be obtained by running the entity recognizer over unlabelled sentences, and adding their annotations to the training set.

https://github.com/explosion/spaCy/tree/master/examples/training/train_new_entity_type.py

If you're using an existing model, make sure to mix in examples of other entity types that spaCy correctly recognized before. Otherwise, your model might learn the new type, but "forget" what it previously knew. This is also referred to as the "catastrophic forgetting" problem.

Step by step guide

1. Load the model you want to start with, or create an empty model using spacy.blank with the ID of your language. If you're using a blank model, don't forget to add the entity recognizer to the pipeline. If you're using an existing model, make sure to disable all other pipeline components during training using nlp.select_pipes. This way, you'll only be training the entity recognizer.
2. Add the new entity label to the entity recognizer using the add_label method. You can access the entity recognizer in the pipeline via nlp.get_pipe('ner').
3. Loop over the examples and call nlp.update, which steps through the words of the input. At each word, it makes a prediction. It then consults the annotations, to see whether it was right. If it was wrong, it adjusts its weights so that the correct action will score higher next time.
4. Save the trained model using nlp.to_disk.
5. Test the model to make sure the new entity is recognized correctly.

Training the tagger and parser

Updating the Dependency Parser

This example shows how to train spaCy's dependency parser, starting off with an existing model or a blank model. You'll need a set of training examples and the respective heads and dependency label for each token of the example texts.

https://github.com/explosion/spaCy/tree/master/examples/training/train_parser.py

Step by step guide

1. Load the model you want to start with, or create an empty model using spacy.blank with the ID of your language. If you're using a blank model, don't forget to add the parser to the pipeline. If you're using an existing model, make sure to disable all other pipeline components during training using nlp.select_pipes. This way, you'll only be training the parser.
2. Add the dependency labels to the parser using the add_label method. If you're starting off with a pretrained spaCy model, this is usually not necessary – but it doesn't hurt either, just to be safe.
3. Shuffle and loop over the examples. For each example, update the model by calling nlp.update, which steps through the words of the input. At each word, it makes a prediction. It then consults the annotations to see whether it was right. If it was wrong, it adjusts its weights so that the correct action will score higher next time.
4. Save the trained model using nlp.to_disk.
5. Test the model to make sure the parser works as expected.

Updating the Part-of-speech Tagger

In this example, we're training spaCy's part-of-speech tagger with a custom tag map. We start off with a blank Language class, update its defaults with our custom tags and then train the tagger. You'll need a set of training examples and the respective custom tags, as well as a dictionary mapping those tags to the Universal Dependencies scheme.

https://github.com/explosion/spaCy/tree/master/examples/training/train_tagger.py

Step by step guide

1. Load the model you want to start with, or create an empty model using spacy.blank with the ID of your language. If you're using a blank model, don't forget to add the tagger to the pipeline. If you're using an existing model, make sure to disable all other pipeline components during training using nlp.select_pipes. This way, you'll only be training the tagger.
2. Add the tag map to the tagger using the add_label method. The first argument is the new tag name, the second the mapping to spaCy's coarse-grained tags, e.g. {'pos': 'NOUN'}.
3. Shuffle and loop over the examples. For each example, update the model by calling nlp.update, which steps through the words of the input. At each word, it makes a prediction. It then consults the annotations to see whether it was right. If it was wrong, it adjusts its weights so that the correct action will score higher next time.
4. Save the trained model using nlp.to_disk.
5. Test the model to make sure the parser works as expected.

Training a parser for custom semantics

spaCy's parser component can be used to be trained to predict any type of tree structure over your input text – including semantic relations that are not syntactic dependencies. This can be useful to for conversational applications, which need to predict trees over whole documents or chat logs, with connections between the sentence roots used to annotate discourse structure. For example, you can train spaCy's parser to label intents and their targets, like attributes, quality, time and locations. The result could look like this:

doc = nlp("find a hotel with good wifi")
print([(t.text, t.dep_, t.head.text) for t in doc if t.dep_ != '-'])
# [('find', 'ROOT', 'find'), ('hotel', 'PLACE', 'find'),
#  ('good', 'QUALITY', 'wifi'), ('wifi', 'ATTRIBUTE', 'hotel')]

The above tree attaches "wifi" to "hotel" and assigns the dependency label ATTRIBUTE. This may not be a correct syntactic dependency – but in this case, it expresses exactly what we need: the user is looking for a hotel with the attribute "wifi" of the quality "good". This query can then be processed by your application and used to trigger the respective action – e.g. search the database for hotels with high ratings for their wifi offerings.

Tip: merge phrases and entities

To achieve even better accuracy, try merging multi-word tokens and entities specific to your domain into one token before parsing your text. You can do this by running the entity recognizer or rule-based matcher to find relevant spans, and merging them using Doc.retokenize. You could even add your own custom pipeline component to do this automatically – just make sure to add it before='parser'.

The following example shows a full implementation of a training loop for a custom message parser for a common "chat intent": finding local businesses. Our message semantics will have the following types of relations: ROOT, PLACE, QUALITY, ATTRIBUTE, TIME and LOCATION.

https://github.com/explosion/spaCy/tree/master/examples/training/train_intent_parser.py

Step by step guide

1. Create the training data consisting of words, their heads and their dependency labels in order. A token's head is the index of the token it is attached to. The heads don't need to be syntactically correct – they should express the semantic relations you want the parser to learn. For words that shouldn't receive a label, you can choose an arbitrary placeholder, for example -.
2. Load the model you want to start with, or create an empty model using spacy.blank with the ID of your language. If you're using a blank model, don't forget to add the custom parser to the pipeline. If you're using an existing model, make sure to remove the old parser from the pipeline, and disable all other pipeline components during training using nlp.select_pipes. This way, you'll only be training the parser.
3. Add the dependency labels to the parser using the add_label method.
4. Shuffle and loop over the examples. For each example, update the model by calling nlp.update, which steps through the words of the input. At each word, it makes a prediction. It then consults the annotations to see whether it was right. If it was wrong, it adjusts its weights so that the correct action will score higher next time.
5. Save the trained model using nlp.to_disk.
6. Test the model to make sure the parser works as expected.

Training a text classification model

Adding a text classifier to a spaCy model

This example shows how to train a convolutional neural network text classifier on IMDB movie reviews, using spaCy's new TextCategorizer component. The dataset will be loaded automatically via Thinc's built-in dataset loader. Predictions are available via Doc.cats.

https://github.com/explosion/spaCy/tree/master/examples/training/train_textcat.py

Step by step guide

1. Load the model you want to start with, or create an empty model using spacy.blank with the ID of your language. If you're using an existing model, make sure to disable all other pipeline components during training using nlp.select_pipes. This way, you'll only be training the text classifier.
2. Add the text classifier to the pipeline, and add the labels you want to train – for example, POSITIVE.
3. Load and pre-process the dataset, shuffle the data and split off a part of it to hold back for evaluation. This way, you'll be able to see results on each training iteration.
4. Loop over the training examples and partition them into batches using spaCy's minibatch and compounding helpers.
5. Update the model by calling nlp.update, which steps through the examples and makes a prediction. It then consults the annotations to see whether it was right. If it was wrong, it adjusts its weights so that the correct prediction will score higher next time.
6. Optionally, you can also evaluate the text classifier on each iteration, by checking how it performs on the development data held back from the dataset. This lets you print the precision, recall and F-score.
7. Save the trained model using nlp.to_disk.
8. Test the model to make sure the text classifier works as expected.

Entity linking

To train an entity linking model, you first need to define a knowledge base (KB).

Creating a knowledge base

A KB consists of a list of entities with unique identifiers. Each such entity has an entity vector that will be used to measure similarity with the context in which an entity is used. These vectors have a fixed length and are stored in the KB.

The following example shows how to build a knowledge base from scratch, given a list of entities and potential aliases. The script requires an nlp model with pretrained word vectors to obtain an encoding of an entity's description as its vector.

https://github.com/explosion/spaCy/tree/master/examples/training/create_kb.py

Step by step guide

1. Load the model you want to start with. It should contain pretrained word vectors.
2. Obtain the entity embeddings by running the descriptions of the entities through the nlp model and taking the average of all words with nlp(desc).vector. At this point, a custom encoding step can also be used.
3. Construct the KB by defining all entities with their embeddings, and all aliases with their prior probabilities.
4. Save the KB using kb.dump.
5. Print the contents of the KB to make sure the entities were added correctly.

Training an entity linking model

This example shows how to create an entity linker pipe using a previously created knowledge base. The entity linker is then trained with a set of custom examples. To do so, you need to provide example texts, and the character offsets and knowledge base identifiers of each entity contained in the texts.

https://github.com/explosion/spaCy/tree/master/examples/training/train_entity_linker.py

Step by step guide

1. Load the KB you want to start with, and specify the path to the Vocab object that was used to create this KB. Then, create an empty model using spacy.blank with the ID of your language. Add a component for recognizing sentences en one for identifying relevant entities. In practical applications, you will want a more advanced pipeline including also a component for named entity recognition. Then, create a new entity linker component, add the KB to it, and then add the entity linker to the pipeline. If you're using a model with additional components, make sure to disable all other pipeline components during training using nlp.select_pipes. This way, you'll only be training the entity linker.
2. Shuffle and loop over the examples. For each example, update the model by calling nlp.update, which steps through the annotated examples of the input. For each combination of a mention in text and a potential KB identifier, the model makes a prediction whether or not this is the correct match. It then consults the annotations to see whether it was right. If it was wrong, it adjusts its weights so that the correct combination will score higher next time.
3. Save the trained model using nlp.to_disk.
4. Test the model to make sure the entities in the training data are recognized correctly.

Optimization tips and advice

There are lots of conflicting "recipes" for training deep neural networks at the moment. The cutting-edge models take a very long time to train, so most researchers can't run enough experiments to figure out what's really going on. For what it's worth, here's a recipe that seems to work well on a lot of NLP problems:

1. Initialize with batch size 1, and compound to a maximum determined by your data size and problem type.
2. Use Adam solver with fixed learning rate.
3. Use averaged parameters
4. Use L2 regularization.
5. Clip gradients by L2 norm to 1.
6. On small data sizes, start at a high dropout rate, with linear decay.

This recipe has been cobbled together experimentally. Here's why the various elements of the recipe made enough sense to try initially, and what you might try changing, depending on your problem.

Compounding batch size

The trick of increasing the batch size is starting to become quite popular (see Smith et al., 2017). Their recipe is quite different from how spaCy's models are being trained, but there are some similarities. In training the various spaCy models, we haven't found much advantage from decaying the learning rate – but starting with a low batch size has definitely helped. You should try it out on your data, and see how you go. Here's our current strategy:

### Batch heuristic
def get_batches(train_data, model_type):
    max_batch_sizes = {"tagger": 32, "parser": 16, "ner": 16, "textcat": 64}
    max_batch_size = max_batch_sizes[model_type]
    if len(train_data) < 1000:
        max_batch_size /= 2
    if len(train_data) < 500:
        max_batch_size /= 2
    batch_size = compounding(1, max_batch_size, 1.001)
    batches = minibatch(train_data, size=batch_size)
    return batches

This will set the batch size to start at 1, and increase each batch until it reaches a maximum size. The tagger, parser and entity recognizer all take whole sentences as input, so they're learning a lot of labels in a single example. You therefore need smaller batches for them. The batch size for the text categorizer should be somewhat larger, especially if your documents are long.

Learning rate, regularization and gradient clipping

By default spaCy uses the Adam solver, with default settings (learn_rate=0.001, beta1=0.9, beta2=0.999). Some researchers have said they found these settings terrible on their problems – but they've always performed very well in training spaCy's models, in combination with the rest of our recipe. You can change these settings directly, by modifying the corresponding attributes on the optimizer object. You can also set environment variables, to adjust the defaults.

There are two other key hyper-parameters of the solver: L2 regularization, and gradient clipping (max_grad_norm). Gradient clipping is a hack that's not discussed often, but everybody seems to be using. It's quite important in helping to ensure the network doesn't diverge, which is a fancy way of saying "fall over during training". The effect is sort of similar to setting the learning rate low. It can also compensate for a large batch size (this is a good example of how the choices of all these hyper-parameters intersect).

Dropout rate

For small datasets, it's useful to set a high dropout rate at first, and decay it down towards a more reasonable value. This helps avoid the network immediately overfitting, while still encouraging it to learn some of the more interesting things in your data. spaCy comes with a decaying utility function to facilitate this. You might try setting:

from spacy.util import decaying
dropout = decaying(0.6, 0.2, 1e-4)

You can then draw values from the iterator with next(dropout), which you would pass to the drop keyword argument of nlp.update. It's pretty much always a good idea to use at least some dropout. All of the models currently use Bernoulli dropout, for no particularly principled reason – we just haven't experimented with another scheme like Gaussian dropout yet.

Parameter averaging

The last part of our optimization recipe is parameter averaging, an old trick introduced by Freund and Schapire (1999), popularized in the NLP community by Collins (2002), and explained in more detail by Leon Bottou. Just about the only other people who seem to be using this for neural network training are the SyntaxNet team (one of whom is Michael Collins) – but it really seems to work great on every problem.

The trick is to store the moving average of the weights during training. We don't optimize this average – we just track it. Then when we want to actually use the model, we use the averages, not the most recent value. In spaCy (and Thinc) this is done by using a context manager, use_params, to temporarily replace the weights:

with nlp.use_params(optimizer.averages):
    nlp.to_disk("/model")

The context manager is handy because you naturally want to evaluate and save the model at various points during training (e.g. after each epoch). After evaluating and saving, the context manager will exit and the weights will be restored, so you resume training from the most recent value, rather than the average. By evaluating the model after each epoch, you can remove one hyper-parameter from consideration (the number of epochs). Having one less magic number to guess is extremely nice – so having the averaging under a context manager is very convenient.