5 Commits

Author SHA1 Message Date
MichelBartels
5b6b0cef77
Add UnlabeledTextProcessor (#2054)
* add UnlabeledTextProcessor

* allow choosing processor when finetuning or distilling

* fix type hint

* Add latest docstring and tutorial changes

* improve segment id computation for UnlabeledTextProcessor

* add text and documentation

* change batch size parameter for intermediate layer distillation

* Add latest docstring and tutorial changes

* fix distillation dim mapping

* remove unnecessary changes

* removed confusing parameter

* Add latest docstring and tutorial changes

Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
2022-01-25 14:54:34 +01:00
MichelBartels
0cca2b97cd
distinguish intermediate layer & prediction layer distillation phases with different parameters (#2001)
* add parameters to allow for different hyperparameters in stage 1 and 2 of tinybert distillation

* Add latest docstring and tutorial changes

* improve default parameters

* Add latest docstring and tutorial changes

* split up distillation method

* Add latest docstring and tutorial changes

Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
2022-01-14 20:40:38 +01:00
MichelBartels
f33c2b987a
Adding distillation loss functions from TinyBERT (#1879)
* initial tinybertdistill commit

* add tinybert distill loss

* remove teacher caching for tinybert

* add tinybert to distil_from method

* Add latest docstring and tutorial changes

* add dim mapping and fix type hints

* fix type hints

* fix dummy input

* fix dim mapping for tinybert loss and add comments/doc strings

* add test for tinybert loss

* Add latest docstring and tutorial changes

* add comment

* fix BERT forward parameters

* add doc string to AdaptiveModel forward method

* remove unnecessary data silo

* fix farm import

Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
2021-12-23 14:54:02 +01:00
tstadel
fc8df2163d
Fix Windows CI OOM (#1878)
* set fixture scope to "function"

* run FARMReader without multiprocessing

* dispose off ray after tests

* run most expensive tasks first in test files

* run expensive tests first

* run garbage collector between tests

Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
2021-12-22 17:20:23 +01:00
MichelBartels
84147edcca
Model Distillation (#1758)
* initial commit

* Add latest docstring and tutorial changes

* added comments and fixed bug

* fixed bugs, added benchmark and added documentation

* Add latest docstring and tutorial changes

* fix type: ignore comment

* fix logging in benchmark

* fixed distillation config

* Add latest docstring and tutorial changes

* added type annotations

* fixed distillation loss calculation

* added type annotations

* fixed distillation mse loss

* improved model distillation benchmark config loading

* added temperature for model distillation

* removed uncessary imports, added comments, added named parameter calls

* Add latest docstring and tutorial changes

* added some more comments

* added distillation test

* fixed distillation test

* removed unnecessary import

* fix softmax dimension

* add grid search

* improved model distillation benchmark config

* fixed model distillation hyperparameter search

* added doc strings and type hints for model distillation

* Add latest docstring and tutorial changes

* fixed type hints

* fixed type hints

* fixed type hints

* wrote out params instead of kwargs in DistillationDataSilo initializer

* fixed type hints

* fixed typo

* fixed typo

Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
2021-11-26 18:49:30 +01:00