Commit Graph

12 Commits

Author SHA1 Message Date
Sofie Van Landeghem
2998131416
Reproducibility for TextCat and Tok2Vec (#6218)
* ensure fixed seed in HashEmbed layers

* forgot about the joys of python 2
2020-10-08 00:43:46 +02:00
Matthew Honnibal
3e78e82a83
Experimental character-based pretraining (#5700)
* Use cosine loss in Cloze multitask

* Fix char_embed for gpu

* Call resume_training for base model in train CLI

* Fix bilstm_depth default in pretrain command

* Implement character-based pretraining objective

* Use chars loss in ClozeMultitask

* Add method to decode predicted characters

* Fix number characters

* Rescale gradients for mlm

* Fix char embed+vectors in ml

* Fix pipes

* Fix pretrain args

* Move get_characters_loss

* Fix import

* Fix import

* Mention characters loss option in pretrain

* Remove broken 'self attention' option in pretrain

* Revert "Remove broken 'self attention' option in pretrain"

This reverts commit 56b820f6af.

* Document 'characters' objective of pretrain
2020-07-05 15:48:39 +02:00
Ines Montani
09cec3e41b
Replace function registries with catalogue (#4584)
* Replace functions registries with catalogue

* Update __init__.py

* Fix test

* Revert unrelated flag [ci skip]
2019-11-07 11:45:22 +01:00
Ines Montani
2c107f02a4 Auto-format [ci skip] 2019-10-31 15:01:56 +01:00
Matthew Honnibal
e82306937e Put Tok2Vec refactor behind feature flag (#4563)
* Add back pre-2.2.2 tok2vec

* Add simple tok2vec tests

* Add simple tok2vec tests

* Reformat

* Fix CharacterEmbed in new tok2vec

* Fix legacy tok2vec

* Resolve circular imports

* Fix test for Python 2
2019-10-31 15:01:15 +01:00
Ines Montani
5e9849b60f Auto-format [ci skip] 2019-10-30 19:27:18 +01:00
Matthew Honnibal
9e210fa7fd
Fix tok2vec structure after model registry refactor (#4549)
The model registry refactor of the Tok2Vec function broke loading models
trained with the previous function, because the model tree was slightly
different. Specifically, the new function wrote:

    concatenate(norm, prefix, suffix, shape)

To build the embedding layer. In the previous implementation, I had used
the operator overloading shortcut:

    ( norm | prefix | suffix | shape )

This actually gets mapped to a binary association, giving something
like:

    concatenate(norm, concatenate(prefix, concatenate(suffix, shape)))

This is a different tree, so the layers iterate differently and we
loaded the weights wrongly.
2019-10-28 23:59:03 +01:00
Matthew Honnibal
d5509e0989 Support Mish activation (requires Thinc 7.3) (#4536)
* Add arch for MishWindowEncoder

* Support mish in tok2vec and conv window >=2

* Pass new tok2vec settings from parser

* Syntax error

* Fix tok2vec setting

* Fix registration of MishWindowEncoder

* Fix receptive field setting

* Fix mish arch

* Pass more options from parser

* Support more tok2vec options in pretrain

* Require thinc 7.3

* Add docs [ci skip]

* Require thinc 7.3.0.dev0 to run CI

* Run black

* Fix typo

* Update Thinc version


Co-authored-by: Ines Montani <ines@ines.io>
2019-10-28 15:16:33 +01:00
Ines Montani
c5e41247e8 Tidy up and auto-format 2019-10-28 12:43:55 +01:00
Matthw Honnibal
46eecdcb70 Remove print 2019-10-27 22:24:19 +01:00
Matthw Honnibal
165e378082 Fix tok2vec arch after refactor 2019-10-27 22:19:10 +01:00
Matthew Honnibal
406eb95a47
Refactor Tok2Vec to use architecture registry (#4518)
* Add refactored tok2vec, using register_architecture

* Refactor Tok2Vec

* Fix ml

* Fix new tok2vec

* Move make_layer to util

* Add wire

* Fix missing import
2019-10-25 22:28:20 +02:00