Refactor to use Pytorch for training models

[percevalw]

Description

This PR refactors EDS-NLP to allow training models and performing inference using PyTorch as the deep-learning backend. Rather than a mere wrapper of Pytorch using spaCy, this is a new framework to build hybrid multi-task models.

To achieve this, instead of patching spaCy's pipeline, a new pipeline was implemented in a similar fashion to aphp/edspdf#12. The new pipeline tries to preserve the existing API, especially for non-machine learning uses such as rule-based components. This means that users can continue to use the library in the same way as before (spacy.blank('xx'), nlp.add_pipe(...)), while also having the option to train models using PyTorch. We still use spaCy data structures such as Doc and Span to represent the texts and their annotations.

It should be noted that this is a work-in-progress and will require further testing before it can be released. We should maybe release it under alpha version number ? Once testing is complete, the new version will be released as a stable version.

Core changes / new features:

• Use the confit package to instantiate components (soon to be published)
• Language.factory -> edsnlp.registry.factory.register (confit registry)
• Lazy loading components from their entry point (had to patch spacy.Language.__init__) to avoid having to wrap every import torch statement for pure rule-based use cases. Hence, torch is not a required dependency
• Training script with Pytorch only (tests/training/)
• Re-implemented the trainable NER component with the new system under eds.ner
• New efficient implementation for eds.transformer (to be used in place of spacy-transformer)
• New eds.text_cnn embedding contextualizer

Checklist

• Publish confit
• Add Span sourcing options to eds.ner (from_ents, from_span_groups)
• Add a training recipe ?
• Re-implement the span qualifier from SpanQualifier trainable component #193
• Update the documentation for NER
• Add documentation for embedding components (eds.transformer, eds.text_cnn)
• Add documentation for the new pipeline system
• Add unit tests for the new pipeline
• Update changelog

[codecov]

Codecov Report

Attention: 36 lines in your changes are missing coverage. Please review.

Comparison is base (1b62d35) 94.76% compared to head (df2bf0a) 96.58%.

❗ Current head df2bf0a differs from pull request most recent head 3ec32ab. Consider uploading reports for the commit 3ec32ab to get more accurate results

Files	Patch %	Lines
edsnlp/optimization.py	91.89%	6 Missing ⚠️
edsnlp/core/pipeline.py	98.45%	5 Missing ⚠️
edsnlp/data/base.py	87.17%	5 Missing ⚠️
edsnlp/data/brat.py	0.00%	5 Missing ⚠️
edsnlp/core/torch_component.py	97.84%	4 Missing ⚠️
edsnlp/data/standoff.py	98.25%	3 Missing ⚠️
edsnlp/core/registry.py	98.34%	2 Missing ⚠️
edsnlp/data/json.py	97.77%	2 Missing ⚠️
edsnlp/pipes/ner/adicap/models.py	85.71%	2 Missing ⚠️
edsnlp/data/converters.py	99.47%	1 Missing ⚠️
... and 1 more

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #202      +/-   ##
==========================================
+ Coverage   94.76%   96.58%   +1.81%     
==========================================
  Files         233      254      +21     
  Lines        6099     8356    +2257     
==========================================
+ Hits         5780     8071    +2291     
+ Misses        319      285      -34     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.