Skip to content

v6.0.0: Add thinc.neural for NLP-oriented deep learning
Compare
Choose a tag to compare
@honnibal honnibal released this 31 Dec 00:14
v6.0.0
bf48795

✨ Major features and improvements

  • NEW: Add thinc.neural to develop neural networks for spaCy.
  • Introduce support for Affine, Maxout, ReLu and Softmax vector-to-vector layers.
  • Introduce support for efficient static word embedding layer with projection matrix and per-word-type memoisation.
  • Introduce support for efficient word vector convolution layer, which also supports per-word-type memoisation.
  • Introduce support for MeanPooling, MaxPooling and MinPooling. Add MultiPooling layer for concatenative pooling.
  • Introduce support for annealed dropout training.
  • Introduce support for classical momentum, Adam and Eve optimisers.
  • Introduce support for averaged parameters for each optimiser.

⚠️ Backwards incompatibilities

The Example class now holds a pointer to its ExampleC struct, where previously it held the struct value. This introduces a small backwards incompatibility in spaCy.

Assets 2
Loading