8000 Commits · sdtblck/mesh · GitHub

More Web Proxy on the site http://driver.im/

Commits

Commits on Apr 1, 2021

Allow for not scaling certain parameters updates by its norm in Adafactor. Also add a parameter to allow for changing the Adafactor decay rate.
Mesh TensorFlow Team
committed

Commits on Mar 31, 2021

no public changes

nshazeer
authored and
Mesh TensorFlow Team
committed
Operation to linearly anneal dropout rate between start_step and end_step

William Fedus
authored and
Mesh TensorFlow Team
committed
Adding switch_max routing that, when overflow occurs, will zero out the lowest probability tokens according to the router instead of the last tokens in the group.
Mesh TensorFlow Team
committed

Commits on Mar 24, 2021

Adafactor accepts a list of dimensions not to factor or reduce.

nshazeer
authored and
Mesh TensorFlow Team
committed
Open prediction inputs file for read to fix T5 Colab. Bump version number for release. Fixes https://github.com/google-research/text-to-text-transfer-transformer/issues/760.

adarob
authored and
Mesh TensorFlow Team
committed

Commits on Mar 21, 2021

Change default checkpoint saving dtype to float32 instead of bfloat16. Saving
Mesh TensorFlow Team
committed

Commits on Mar 18, 2021

Change default checkpoint saving dtype to float32 instead of bfloat16. Saving
Mesh TensorFlow Team
committed

Commits on Mar 5, 2021

[MTF] Maybe context.length_dim doesn't exist outside context.train?

afrozenator
authored and
Mesh TensorFlow Team
committed
[Mesh-TF] Add is_training as an arg to mtf.dropout

afrozenator
authored and
Mesh TensorFlow Team
committed

Commits on Mar 4, 2021

Merge pull request #295 from lucidrains:master
Mesh TensorFlow Team
committed
[T5/MTF] Fix predict_fn flow.

afrozenator
authored and
Mesh TensorFlow Team
committed

Commits on Mar 3, 2021

add scale norm
lucidrains
committed

Commits on Mar 2, 2021

Add option to only load from the specific checkpoints requested, instead of trying to find the nearest ones.

daphnei
authored and
Mesh TensorFlow Team
committed

Commits on Feb 23, 2021

Add optional checkpointing of input pipeline during training.

adarob
authored and
Mesh TensorFlow Team
committed

Commits on Feb 5, 2021

This quick change folds the additional pf dimension in the topology into the 1st mesh tf dimension.
Mesh TensorFlow Team
committed

Commits on Feb 3, 2021

Activation gradient clipping

William Fedus
authored and
Mesh TensorFlow Team
committed

Commits on Feb 2, 2021

no public changes

nshazeer
authored and
Mesh TensorFlow Team
committed

Commits on Jan 30, 2021

Enable doing inference where the priming sequences come from a Task rather than a file.

daphnei
authored and
Mesh TensorFlow Team
committed

Commits on Jan 29, 2021

Allow distillation to start later in training, after the student model is pretrained.
Mesh TensorFlow Team
committed
Change the interface to decode to a generator.
Mesh TensorFlow Team
committed

Commits on Jan 22, 2021

Log sequence lengths of targets during evaluation.

katelee168
authored and
Mesh TensorFlow Team
committed

Commits on Jan 19, 2021

Clarify names within Mesh Tensorflow

William Fedus
authored and
Mesh TensorFlow Team
committed

Commits on Jan 15, 2021

Make train_model no longer gin configurable.

katelee168
authored and
Mesh TensorFlow Team
committed
add ParallelLayer

nshazeer
authored and
Mesh TensorFlow Team
committed

Commits on Jan 8, 2021

Fix bug in T5 model export for `eval_with_score` mode -- handle case where `vocabulary` is tuple.
Mesh TensorFlow Team
committed

Commits on Jan 7, 2021

Add parameter `start_step` to `train_model` to only skip `recovered_step - start_step` steps.

katelee168
authored and
Mesh TensorFlow Team
committed

Commits on Jan 5, 2021

Fix decoding bug for Funnel Transformer

hwchung27
authored and
Mesh TensorFlow Team
committed

Commits on Dec 30, 2020

Fixing name error and more detailed summaries for experts code.

William Fedus
authored and
Mesh TensorFlow Team
committed

Commits on Dec 29, 2020

Internal
Mesh TensorFlow Team
committed

Commits on Dec 21, 2020

Add support for non-string inputs to tokenize preprocessor.

cghawthorne
authored and
Mesh TensorFlow Team
committed
Fix packing bug leading to the loss of *_segmentation and *_position when custom ops are being used. Add test.

adarob
authored and
Mesh TensorFlow Team
committed

Commits on Dec 17, 2020

Bump mesh tensorflow version for release.

adarob
authored and
Mesh TensorFlow Team
committed

Commits on Dec 15, 2020

Add parameter for the total_train_steps (separate from train_steps) for training jobs that are split across multiple launches.

katelee168
authored and
Mesh TensorFlow Team
committed
AT5 - Fix Bug with Weighted Transformer
Mesh TensorFlow Team
committed

0