GitLab

https://github.com/tensorflow/models/issues/10074

PiperOrigin-RevId: 463195812

PiperOrigin-RevId: 463166393

Move `output_range` argument to call(). After evaluating several use cases, `output_range` is better to be specified at `call` time which allows us the train with full sequence and serve with certain tokens.

PiperOrigin-RevId: 462885854

Fix attention dropout rate bug. The output_dropout rate was used for attention dropout by mistake.

PiperOrigin-RevId: 462470287

PiperOrigin-RevId: 461063108

PiperOrigin-RevId: 460912624

PiperOrigin-RevId: 460756776

Adding FAVOR++ mechanism from the paper: "Chefs' Random Tables: Non-Trigonometric Random Features" to TF MODEL GARDEN.

PiperOrigin-RevId: 460625574

PiperOrigin-RevId: 458268967

PiperOrigin-RevId: 458012017

PiperOrigin-RevId: 457383160

PiperOrigin-RevId: 456624049

PiperOrigin-RevId: 456569753

PiperOrigin-RevId: 456159384

PiperOrigin-RevId: 456100154

PiperOrigin-RevId: 455384659

report mrpc with f1 by default.
Remove todos.

PiperOrigin-RevId: 454653695

PiperOrigin-RevId: 454382820

PiperOrigin-RevId: 453982185

PiperOrigin-RevId: 453749903

PiperOrigin-RevId: 453490507

PiperOrigin-RevId: 453487987

PiperOrigin-RevId: 452839394

PiperOrigin-RevId: 450785872

PiperOrigin-RevId: 450124488

https://github.com/tensorflow/models/issues/10638

PiperOrigin-RevId: 448816131

PiperOrigin-RevId: 448695305

PiperOrigin-RevId: 446536595

PiperOrigin-RevId: 446242527

PiperOrigin-RevId: 446219773

PiperOrigin-RevId: 444158657

Also, refactor uvit builder.

PiperOrigin-RevId: 443105970

Update nlp/README.md according to the update in vision/.

PiperOrigin-RevId: 442428156

PiperOrigin-RevId: 441337748

Clean up: move the legacy AdamW implementation to optimization/ and add a warning to guide users consider the choice in Keras package.

PiperOrigin-RevId: 440383490

PiperOrigin-RevId: 440245619

PiperOrigin-RevId: 439095919

PiperOrigin-RevId: 438209905

PiperOrigin-RevId: 438078675