spaCy/website/docs/usage/v3-3.mdx

250 lines
12 KiB
Plaintext

---
title: What's New in v3.3
teaser: New features and how to upgrade
menu:
- ['New Features', 'features']
- ['Upgrading Notes', 'upgrading']
---
## New features {id="features",hidden="true"}
spaCy v3.3 improves the speed of core pipeline components, adds a new trainable
lemmatizer, and introduces trained pipelines for Finnish, Korean and Swedish.
### Speed improvements {id="speed"}
v3.3 includes a slew of speed improvements:
- Speed up parser and NER by using constant-time head lookups.
- Support unnormalized softmax probabilities in `spacy.Tagger.v2` to speed up
inference for tagger, morphologizer, senter and trainable lemmatizer.
- Speed up parser projectivization functions.
- Replace `Ragged` with faster `AlignmentArray` in `Example` for training.
- Improve `Matcher` speed.
- Improve serialization speed for empty `Doc.spans`.
For longer texts, the trained pipeline speeds improve **15%** or more in
prediction. We benchmarked `en_core_web_md` (same components as in v3.2) and
`de_core_news_md` (with the new trainable lemmatizer) across a range of text
sizes on Linux (Intel Xeon W-2265) and OS X (M1) to compare spaCy v3.2 vs. v3.3:
**Intel Xeon W-2265**
| Model | Avg. Words/Doc | v3.2 Words/Sec | v3.3 Words/Sec | Diff |
| :----------------------------------------------- | -------------: | -------------: | -------------: | -----: |
| [`en_core_web_md`](/models/en#en_core_web_md) | 100 | 17292 | 17441 | 0.86% |
| (=same components) | 1000 | 15408 | 16024 | 4.00% |
| | 10000 | 12798 | 15346 | 19.91% |
| [`de_core_news_md`](/models/de/#de_core_news_md) | 100 | 20221 | 19321 | -4.45% |
| (+v3.3 trainable lemmatizer) | 1000 | 17480 | 17345 | -0.77% |
| | 10000 | 14513 | 17036 | 17.38% |
**Apple M1**
| Model | Avg. Words/Doc | v3.2 Words/Sec | v3.3 Words/Sec | Diff |
| ------------------------------------------------ | -------------: | -------------: | -------------: | -----: |
| [`en_core_web_md`](/models/en#en_core_web_md) | 100 | 18272 | 18408 | 0.74% |
| (=same components) | 1000 | 18794 | 19248 | 2.42% |
| | 10000 | 15144 | 17513 | 15.64% |
| [`de_core_news_md`](/models/de/#de_core_news_md) | 100 | 19227 | 19591 | 1.89% |
| (+v3.3 trainable lemmatizer) | 1000 | 20047 | 20628 | 2.90% |
| | 10000 | 15921 | 18546 | 16.49% |
### Trainable lemmatizer {id="trainable-lemmatizer"}
The new [trainable lemmatizer](/api/edittreelemmatizer) component uses
[edit trees](https://explosion.ai/blog/edit-tree-lemmatizer) to transform tokens
into lemmas. Try out the trainable lemmatizer with the
[training quickstart](/usage/training#quickstart)!
### displaCy support for overlapping spans and arcs {id="displacy"}
displaCy now supports overlapping spans with a new
[`span`](/usage/visualizers#span) style and multiple arcs with different labels
between the same tokens for [`dep`](/usage/visualizers#dep) visualizations.
Overlapping spans can be visualized for any spans key in `doc.spans`:
```python
import spacy
from spacy import displacy
from spacy.tokens import Span
nlp = spacy.blank("en")
text = "Welcome to the Bank of China."
doc = nlp(text)
doc.spans["custom"] = [Span(doc, 3, 6, "ORG"), Span(doc, 5, 6, "GPE")]
displacy.serve(doc, style="span", options={"spans_key": "custom"})
```
<Iframe
title="displaCy visualizer for overlapping spans"
src="/images/displacy-span.html"
height={180}
/>
## Additional features and improvements
- Config comparisons with [`spacy debug diff-config`](/api/cli#debug-diff).
- Span suggester debugging with
[`SpanCategorizer.set_candidates`](/api/spancategorizer#set_candidates).
- Big endian support with
[`thinc-bigendian-ops`](https://github.com/andrewsi-z/thinc-bigendian-ops) and
updates to make `floret`, `murmurhash`, Thinc and spaCy endian neutral.
- Initial support for Lower Sorbian and Upper Sorbian.
- Language updates for English, French, Italian, Japanese, Korean, Norwegian,
Russian, Slovenian, Spanish, Turkish, Ukrainian and Vietnamese.
- New noun chunks for Finnish.
## Trained pipelines {id="pipelines"}
### New trained pipelines {id="new-pipelines"}
v3.3 introduces new CPU/CNN pipelines for Finnish, Korean and Swedish, which use
the new trainable lemmatizer and
[floret vectors](https://github.com/explosion/floret). Due to the use
[Bloom embeddings](https://explosion.ai/blog/bloom-embeddings) and subwords, the
pipelines have compact vectors with no out-of-vocabulary words.
| Package | Language | UPOS | Parser LAS | NER F |
| ----------------------------------------------- | -------- | ---: | ---------: | ----: |
| [`fi_core_news_sm`](/models/fi#fi_core_news_sm) | Finnish | 92.5 | 71.9 | 75.9 |
| [`fi_core_news_md`](/models/fi#fi_core_news_md) | Finnish | 95.9 | 78.6 | 80.6 |
| [`fi_core_news_lg`](/models/fi#fi_core_news_lg) | Finnish | 96.2 | 79.4 | 82.4 |
| [`ko_core_news_sm`](/models/ko#ko_core_news_sm) | Korean | 86.1 | 65.6 | 71.3 |
| [`ko_core_news_md`](/models/ko#ko_core_news_md) | Korean | 94.7 | 80.9 | 83.1 |
| [`ko_core_news_lg`](/models/ko#ko_core_news_lg) | Korean | 94.7 | 81.3 | 85.3 |
| [`sv_core_news_sm`](/models/sv#sv_core_news_sm) | Swedish | 95.0 | 75.9 | 74.7 |
| [`sv_core_news_md`](/models/sv#sv_core_news_md) | Swedish | 96.3 | 78.5 | 79.3 |
| [`sv_core_news_lg`](/models/sv#sv_core_news_lg) | Swedish | 96.3 | 79.1 | 81.1 |
### Pipeline updates {id="pipeline-updates"}
The following languages switch from lookup or rule-based lemmatizers to the new
trainable lemmatizer: Danish, Dutch, German, Greek, Italian, Lithuanian,
Norwegian, Polish, Portuguese and Romanian. The overall lemmatizer accuracy
improves for all of these pipelines, but be aware that the types of errors may
look quite different from the lookup-based lemmatizers. If you'd prefer to
continue using the previous lemmatizer, you can
[switch from the trainable lemmatizer to a non-trainable lemmatizer](/models#design-modify).
<figure>
| Model | v3.2 Lemma Acc | v3.3 Lemma Acc |
| ----------------------------------------------- | -------------: | -------------: |
| [`da_core_news_md`](/models/da#da_core_news_md) | 84.9 | 94.8 |
| [`de_core_news_md`](/models/de#de_core_news_md) | 73.4 | 97.7 |
| [`el_core_news_md`](/models/el#el_core_news_md) | 56.5 | 88.9 |
| [`fi_core_news_md`](/models/fi#fi_core_news_md) | - | 86.2 |
| [`it_core_news_md`](/models/it#it_core_news_md) | 86.6 | 97.2 |
| [`ko_core_news_md`](/models/ko#ko_core_news_md) | - | 90.0 |
| [`lt_core_news_md`](/models/lt#lt_core_news_md) | 71.1 | 84.8 |
| [`nb_core_news_md`](/models/nb#nb_core_news_md) | 76.7 | 97.1 |
| [`nl_core_news_md`](/models/nl#nl_core_news_md) | 81.5 | 94.0 |
| [`pl_core_news_md`](/models/pl#pl_core_news_md) | 87.1 | 93.7 |
| [`pt_core_news_md`](/models/pt#pt_core_news_md) | 76.7 | 96.9 |
| [`ro_core_news_md`](/models/ro#ro_core_news_md) | 81.8 | 95.5 |
| [`sv_core_news_md`](/models/sv#sv_core_news_md) | - | 95.5 |
</figure>
In addition, the vectors in the English pipelines are deduplicated to improve
the pruned vectors in the `md` models and reduce the `lg` model size.
## Notes about upgrading from v3.2 {id="upgrading"}
### Span comparisons
Span comparisons involving ordering (`<`, `<=`, `>`, `>=`) now take all span
attributes into account (start, end, label, and KB ID) so spans may be sorted in
a slightly different order.
### Whitespace annotation
During training, annotation on whitespace tokens is handled in the same way as
annotation on non-whitespace tokens in order to allow custom whitespace
annotation.
### Doc.from_docs
[`Doc.from_docs`](/api/doc#from_docs) now includes `Doc.tensor` by default and
supports excludes with an `exclude` argument in the same format as
`Doc.to_bytes`. The supported exclude fields are `spans`, `tensor` and
`user_data`.
Docs including `Doc.tensor` may be quite a bit larger in RAM, so to exclude
`Doc.tensor` as in v3.2:
```diff
-merged_doc = Doc.from_docs(docs)
+merged_doc = Doc.from_docs(docs, exclude=["tensor"])
```
### Using trained pipelines with floret vectors
If you're running a new trained pipeline for Finnish, Korean or Swedish on new
texts and working with `Doc` objects, you shouldn't notice any difference with
floret vectors vs. default vectors.
If you use vectors for similarity comparisons, there are a few differences,
mainly because a floret pipeline doesn't include any kind of frequency-based
word list similar to the list of in-vocabulary vector keys with default vectors.
- If your workflow iterates over the vector keys, you should use an external
word list instead:
```diff
- lexemes = [nlp.vocab[orth] for orth in nlp.vocab.vectors]
+ lexemes = [nlp.vocab[word] for word in external_word_list]
```
- `Vectors.most_similar` is not supported because there's no fixed list of
vectors to compare your vectors to.
### Pipeline package version compatibility {id="version-compat"}
> #### Using legacy implementations
>
> In spaCy v3, you'll still be able to load and reference legacy implementations
> via [`spacy-legacy`](https://github.com/explosion/spacy-legacy), even if the
> components or architectures change and newer versions are available in the
> core library.
When you're loading a pipeline package trained with an earlier version of spaCy
v3, you will see a warning telling you that the pipeline may be incompatible.
This doesn't necessarily have to be true, but we recommend running your
pipelines against your test suite or evaluation data to make sure there are no
unexpected results.
If you're using one of the [trained pipelines](/models) we provide, you should
run [`spacy download`](/api/cli#download) to update to the latest version. To
see an overview of all installed packages and their compatibility, you can run
[`spacy validate`](/api/cli#validate).
If you've trained your own custom pipeline and you've confirmed that it's still
working as expected, you can update the spaCy version requirements in the
[`meta.json`](/api/data-formats#meta):
```diff
- "spacy_version": ">=3.2.0,<3.3.0",
+ "spacy_version": ">=3.2.0,<3.4.0",
```
### Updating v3.2 configs
To update a config from spaCy v3.2 with the new v3.3 settings, run
[`init fill-config`](/api/cli#init-fill-config):
```bash
$ python -m spacy init fill-config config-v3.2.cfg config-v3.3.cfg
```
In many cases ([`spacy train`](/api/cli#train),
[`spacy.load`](/api/top-level#spacy.load)), the new defaults will be filled in
automatically, but you'll need to fill in the new settings to run
[`debug config`](/api/cli#debug) and [`debug data`](/api/cli#debug-data).
To see the speed improvements for the
[`Tagger` architecture](/api/architectures#Tagger), edit your config to switch
from `spacy.Tagger.v1` to `spacy.Tagger.v2` and then run `init fill-config`.