Currently the TextCategorizer defaults to a fairly complicated model, designed partly around the active learning requirements of Prodigy. The model's a bit slow, and not very GPU-friendly.
This patch implements a straightforward CNN model that still performs pretty well. The replacement model also makes it easy to use the LMAO pretraining, since most of the parameters are in the CNN.
The replacement model has a flag to specify whether labels are mutually exclusive, which defaults to True. This has been a common problem with the text classifier. We'll also now be able to support adding labels to pretrained models again.
Resolves#2934, #2756, #1798, #1748.
<!--- Provide a general summary of your changes in the title. -->
## Description
- [x] Use [`black`](https://github.com/ambv/black) to auto-format all `.py` files.
- [x] Update flake8 config to exclude very large files (lemmatization tables etc.)
- [x] Update code to be compatible with flake8 rules
- [x] Fix various small bugs, inconsistencies and messy stuff in the language data
- [x] Update docs to explain new code style (`black`, `flake8`, when to use `# fmt: off` and `# fmt: on` and what `# noqa` means)
Once #2932 is merged, which auto-formats and tidies up the CLI, we'll be able to run `flake8 spacy` actually get meaningful results.
At the moment, the code style and linting isn't applied automatically, but I'm hoping that the new [GitHub Actions](https://github.com/features/actions) will let us auto-format pull requests and post comments with relevant linting information.
### Types of change
enhancement, code style
## Checklist
<!--- Before you submit the PR, go over this checklist and make sure you can
tick off all the boxes. [] -> [x] -->
- [x] I have submitted the spaCy Contributor Agreement.
- [x] I ran the tests, and all new and existing tests passed.
- [x] My changes don't require a change to the documentation, or if they do, I've added all required information.
These experiments were completed a few weeks ago, but I didn't make the PR, pending model release.
Token vector width: 128->96
Hidden width: 128->64
Embed size: 5000->2000
Dropout: 0.2->0.1
Updated optimizer defaults (unclear how important?)
This should improve speed, model size and load time, while keeping
similar or slightly better accuracy.
The tl;dr is we prefer to prevent over-fitting by reducing model size,
rather than using more dropout.
Add a new class for the Tagger model, MultiSoftmax. This allows softmax
prediction of multiple classes on the same output layer, e.g. one
variable with 3 classes, another with 4 classes. This makes a layer with
7 output neurons, which we softmax into two distributions.
* Add spacy.errors module
* Update deprecation and user warnings
* Replace errors and asserts with new error message system
* Remove redundant asserts
* Fix whitespace
* Add messages for print/util.prints statements
* Fix typo
* Fix typos
* Move CLI messages to spacy.cli._messages
* Add decorator to display error code with message
An implementation like this is nice because it only modifies the string when it's retrieved from the containing class – so we don't have to worry about manipulating tracebacks etc.
* Remove unused link in spacy.about
* Update errors for invalid pipeline components
* Improve error for unknown factories
* Add displaCy warnings
* Update formatting consistency
* Move error message to spacy.errors
* Update errors and check if doc returned by component is None
This patch addresses #1660, which was caused by keying all pre-trained
vectors with the same ID when telling Thinc how to refer to them. This
meant that if multiple models were loaded that had pre-trained vectors,
errors or incorrect behaviour resulted.
The vectors class now includes a .name attribute, which defaults to:
{nlp.meta['lang']_nlp.meta['name']}.vectors
The vectors name is set in the cfg of the pipeline components under the
key pretrained_vectors. This replaces the previous cfg key
pretrained_dims.
In order to make existing models compatible with this change, we check
for the pretrained_dims key when loading models in from_disk and
from_bytes, and add the cfg key pretrained_vectors if we find it.
