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.
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
The helper method state.B(1) gets the index of the first token of the
buffer, or -1 if no such token exists. Normally this is safe because we
pass this to functions like state.safe_get(), which returns an empty
token. Here we used it directly as an array index, which is not okay!
This error may have been the cause of out-of-bounds access errors during
training. Similar errors may still be around, so much be hunted down.
Hunting this one down took a long time...I printed out values across
training runs and diffed, looking for points of divergence between
runs, when no randomness should be allowed.
The helper method state.B(1) gets the index of the first token of the
buffer, or -1 if no such token exists. Normally this is safe because we
pass this to functions like state.safe_get(), which returns an empty
token. Here we used it directly as an array index, which is not okay!
This error may have been the cause of out-of-bounds access errors during
training. Similar errors may still be around, so much be hunted down.
Hunting this one down took a long time...I printed out values across
training runs and diffed, looking for points of divergence between
runs, when no randomness should be allowed.
The parser.begin_training() method was rewritten in v2.1. The rewrite
introduced a regression, where if you added labels prior to
begin_training(), these labels were discarded. This patch fixes that.
* 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
* 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.
This patch does a few smallish things that tighten up the training workflow a little, and allow memory use during training to be reduced by letting the GoldCorpus stream data properly.
Previously, the parser and entity recognizer read and saved labels as lists, with extra labels noted separately. Lists were used becaue ordering is very important, to ensure that the label-to-class mapping is stable.
We now manage labels as nested dictionaries, first keyed by the action, and then keyed by the label. Values are frequencies. The trick is, how do we save new labels? We need to make sure we iterate over these in the same order they're added. Otherwise, we'll get different class IDs, and the model's predictions won't make sense.
To allow stable sorting, we map the new labels to negative values. If we have two new labels, they'll be noted as having "frequency" -1 and -2. The next new label will then have "frequency" -3. When we sort by (frequency, label), we then get a stable sort.
Storing frequencies then allows us to make the next nice improvement. Previously we had to iterate over the whole training set, to pre-process it for the deprojectivisation. This led to storing the whole training set in memory. This was most of the required memory during training.
To prevent this, we now store the frequencies as we stream in the data, and deprojectivize as we go. Once we've built the frequencies, we can then apply a frequency cut-off when we decide how many classes to make.
Finally, to allow proper data streaming, we also have to have some way of shuffling the iterator. This is awkward if the training files have multiple documents in them. To solve this, the GoldCorpus class now writes the training data to disk in msgpack files, one per document. We can then shuffle the data by shuffling the paths.
This is a squash merge, as I made a lot of very small commits. Individual commit messages below.
* Simplify label management for TransitionSystem and its subclasses
* Fix serialization for new label handling format in parser
* Simplify and improve GoldCorpus class. Reduce memory use, write to temp dir
* Set actions in transition system
* Require thinc 6.11.1.dev4
* Fix error in parser init
* Add unicode declaration
* Fix unicode declaration
* Update textcat test
* Try to get model training on less memory
* Print json loc for now
* Try rapidjson to reduce memory use
* Remove rapidjson requirement
* Try rapidjson for reduced mem usage
* Handle None heads when projectivising
* Stream json docs
* Fix train script
* Handle projectivity in GoldParse
* Fix projectivity handling
* Add minibatch_by_words util from ud_train
* Minibatch by number of words in spacy.cli.train
* Move minibatch_by_words util to spacy.util
* Fix label handling
* More hacking at label management in parser
* Fix encoding in msgpack serialization in GoldParse
* Adjust batch sizes in parser training
* Fix minibatch_by_words
* Add merge_subtokens function to pipeline.pyx
* Register merge_subtokens factory
* Restore use of msgpack tmp directory
* Use minibatch-by-words in train
* Handle retokenization in scorer
* Change back-off approach for missing labels. Use 'dep' label
* Update NER for new label management
* Set NER tags for over-segmented words
* Fix label alignment in gold
* Fix label back-off for infrequent labels
* Fix int type in labels dict key
* Fix int type in labels dict key
* Update feature definition for 8 feature set
* Update ud-train script for new label stuff
* Fix json streamer
* Print the line number if conll eval fails
* Update children and sentence boundaries after deprojectivisation
* Export set_children_from_heads from doc.pxd
* Render parses during UD training
* Remove print statement
* Require thinc 6.11.1.dev6. Try adding wheel as install_requires
* Set different dev version, to flush pip cache
* Update thinc version
* Update GoldCorpus docs
* Remove print statements
* Fix formatting and links [ci skip]
Matthew Honnibal