The output weights often return negative scores for classes, especially
via the bias terms. This means that when we add a new class, we can't
rely on just zeroing the weights, or we'll end up with positive
predictions for those labels.
To solve this, we use nan values as the initial weights for new labels.
This prevents them from ever coming out on top. During backprop, we
replace the nan values with the minimum assigned score, so that we're
still able to learn these classes.
The new spacy pretrain command implemented BERT/ULMFit/etc-like transfer learning, using our Language Modelling with Approximate Outputs version of BERT's cloze task. Pretraining is convenient, but in some ways it's a bit of a strange solution. All we're doing is initialising the weights. At the same time, we're putting a lot of work into our optimisation so that it's less sensitive to initial conditions, and more likely to find good optima. I discuss this a bit in the pseudo-rehearsal blog post: https://explosion.ai/blog/pseudo-rehearsal-catastrophic-forgetting
Support semi-supervised learning in spacy train
One obvious way to improve these pretraining methods is to do multi-task learning, instead of just transfer learning. This has been shown to work very well: https://arxiv.org/pdf/1809.08370.pdf . This patch makes it easy to do this sort of thing.
Add a new argument to spacy train, --raw-text. This takes a jsonl file with unlabelled data that can be used in arbitrary ways to do semi-supervised learning.
Add a new method to the Language class and to pipeline components, .rehearse(). This is like .update(), but doesn't expect GoldParse objects. It takes a batch of Doc objects, and performs an update on some semi-supervised objective.
Move the BERT-LMAO objective out from spacy/cli/pretrain.py into spacy/_ml.py, so we can create a new pipeline component, ClozeMultitask. This can be specified as a parser or NER multitask in the spacy train command. Example usage:
python -m spacy train en ./tmp ~/data/en-core-web/train/nw.json ~/data/en-core-web/dev/nw.json --pipeline parser --raw-textt ~/data/unlabelled/reddit-100k.jsonl --vectors en_vectors_web_lg --parser-multitasks cloze
Implement rehearsal methods for pipeline components
The new --raw-text argument and nlp.rehearse() method also gives us a good place to implement the the idea in the pseudo-rehearsal blog post in the parser. This works as follows:
Add a new nlp.resume_training() method. This allocates copies of pre-trained models in the pipeline, setting things up for the rehearsal updates. It also returns an optimizer object. This also greatly reduces confusion around the nlp.begin_training() method, which randomises the weights, making it not suitable for adding new labels or otherwise fine-tuning a pre-trained model.
Implement rehearsal updates on the Parser class, making it available for the dependency parser and NER. During rehearsal, the initial model is used to supervise the model being trained. The current model is asked to match the predictions of the initial model on some data. This minimises catastrophic forgetting, by keeping the model's predictions close to the original. See the blog post for details.
Implement rehearsal updates for tagger
Implement rehearsal updates for text categoriz
Remove hacks and wrappers, keep code in sync across our libraries and move spaCy a few steps closer to only depending on packages with binary wheels 🎉
See here: https://github.com/explosion/srsly
Serialization is hard, especially across Python versions and multiple platforms. After dealing with many subtle bugs over the years (encodings, locales, large files) our libraries like spaCy and Prodigy have steadily grown a number of utility functions to wrap the multiple serialization formats we need to support (especially json, msgpack and pickle). These wrapping functions ended up duplicated across our codebases, so we wanted to put them in one place.
At the same time, we noticed that having a lot of small dependencies was making maintainence harder, and making installation slower. To solve this, we've made srsly standalone, by including the component packages directly within it. This way we can provide all the serialization utilities we need in a single binary wheel.
srsly currently includes forks of the following packages:
ujson
msgpack
msgpack-numpy
cloudpickle
* WIP: replace json/ujson with srsly
* Replace ujson in examples
Use regular json instead of srsly to make code easier to read and follow
* Update requirements
* Fix imports
* Fix typos
* Replace msgpack with srsly
* Fix warning
Our epic matrix multiplication odyssey is drawing to a close...
I've now finally got the Blis linear algebra routines in a self-contained Python package, with wheels for Windows, Linux and OSX. The only missing platform at the moment is Windows Python 2.7. The result is at https://github.com/explosion/cython-blis
Thinc v7.0.0 will make the change to Blis. I've put a Thinc v7.0.0.dev0 up on PyPi so that we can test these changes with the CI, and even get them out to spacy-nightly, before Thinc v7.0.0 is released. This PR also updates the other dependencies to be in line with the current versions master is using. I've also resolved the msgpack deprecation problems, and gotten spaCy and Thinc up to date with the latest Cython.
The point of switching to Blis is to have control of how our matrix multiplications are executed across platforms. When we were using numpy for this, a different library would be used on pip and conda, OSX would use Accelerate, etc. This would open up different bugs and performance problems, especially when multi-threading was introduced.
With the change to Blis, we now strictly single-thread the matrix multiplications. This will make it much easier to use multiprocessing to parallelise the runtime, since we won't have nested parallelism problems to deal with.
* Use blis
* Use -2 arg to Cython
* Update dependencies
* Fix requirements
* Update setup dependencies
* Fix requirement typo
* Fix msgpack errors
* Remove Python27 test from Appveyor, until Blis works there
* Auto-format setup.py
* Fix murmurhash version
* Work on refactoring greedy parser
* Compile updated parser
* Fix refactored parser
* Update test
* Fix refactored parser
* Fix refactored parser
* Readd beam search after refactor
* Fix beam search after refactor
* Fix parser
* Fix beam parsing
* Support oracle segmentation in ud-train CLI command
* Avoid relying on final gold check in beam search
* Add a keyword argument sink to GoldParse
* Bug fixes to beam search after refactor
* Avoid importing fused token symbol in ud-run-test, untl that's added
* Avoid importing fused token symbol in ud-run-test, untl that's added
* Don't modify Token in global scope
* Fix error in beam gradient calculation
* Default to beam_update_prob 1
* Set a more aggressive threshold on the max violn update
* Disable some tests to figure out why CI fails
* Disable some tests to figure out why CI fails
* Add some diagnostics to travis.yml to try to figure out why build fails
* Tell Thinc to link against system blas on Travis
* Point thinc to libblas on Travis
* Try running sudo=true for travis
* Unhack travis.sh
* Restore beam_density argument for parser beam
* Require thinc 6.11.1.dev16
* Revert hacks to tests
* Revert hacks to travis.yml
* Update thinc requirement
* Fix parser model loading
* Fix size limits in training data
* Add missing name attribute for parser
* Fix appveyor for Windows
Matthew Honnibal