This repository has been archived by the owner on Jul 7, 2023. It is now read-only.

v1.3#449

Merged

lukaszkaiser merged 33 commits intotensorflow:masterfrom rsepassi:push

Nov 29, 2017

+2,656-2,105

Commits on Nov 29, 2017

More tiny sketch fixes
Ryan Sepassi
committed
Add Modality.targets_weights_fn
Ryan Sepassi
committed
Update layout optimizer config code. pruning and constfold are enabled by default now (no need to explicitly specify them any more).

T2T Team
authored and
Ryan Sepassi
committed
Preserve static shape info where possible with shape_list and enable constant ones_matrix_band_part
Ryan Sepassi
committed
Fix weights_fn calls
Ryan Sepassi
committed
Refactor rev_block into a RevBlock class that exposes forward and backward
Ryan Sepassi
committed
Clean up revblock code with tf.contrib.framework.nest
Ryan Sepassi
committed
Add ImageImagenet224; resnet50 now fits batch size of 128
Ryan Sepassi
committed
Added an option to omit quotes from the subtoken voc format.

T2T Team
authored and
Ryan Sepassi
committed
Bug fix, make class lab a VarLen feature to be compatible with img2img datasets that don't have labels.

Niki Parmar
authored and
Ryan Sepassi
committed
Update tpu_trainer to use T2TModel.model_fn and upate RunConfig for non-tpu use
Ryan Sepassi
committed
Fix SymbolModality weights fn
Ryan Sepassi
committed
Add support for multi-gpu training with tpu_trainer
Ryan Sepassi
committed
CHECKPOINT BREAKING: make T2TModel a subclass of Layer so it can be called; all variables are now in model-name scope.

Lukasz Kaiser
authored and
Ryan Sepassi
committed
Remove unused functions from transformer_vae.

Lukasz Kaiser
authored and
Ryan Sepassi
committed
Update distance computation for k-nearest neighbours to be more efficient, by computing the norms separately.

T2T Team
authored and
Ryan Sepassi
committed
Pass the modified hparams to the modalities, so they can know the mode.

nshazeer
authored and
Ryan Sepassi
committed
This change breaks previous checkpoints. Make Transformer fast on TPU.

nshazeer
authored and
Ryan Sepassi
committed
Changes to make t2t tf.eager compatible

Niki Parmar
authored and
Ryan Sepassi
committed
This change breaks previous checkpoints. Make Transformer fast on TPU.

nshazeer
authored and
Ryan Sepassi
committed
Versions of problems with combined examples. Good for TPU training.

nshazeer
authored and
Ryan Sepassi
committed
Small bug fix.

Ashish Vaswani
authored and
Ryan Sepassi
committed
When using eager, use slow decoding for transformer models.

Niki Parmar
authored and
Ryan Sepassi
committed
merge PRs.

Lukasz Kaiser
authored and
Ryan Sepassi
committed
Small updates to T2TModel.__call__ callers
Ryan Sepassi
committed
Make daisy_chain_variables and hparam instead of flag and unset to allow LSTM to train in distributed mode.

Lukasz Kaiser
authored and
Ryan Sepassi
committed
Move Estimator input_fn and model_fn construction into Problem and T2TModel, respectively, which allows subclassing
Ryan Sepassi
committed
This change breaks previous checkpoints. Make Transformer fast on TPU.

nshazeer
authored and
Ryan Sepassi
committed
Correction for eager-mode decoding scopes (to use with pre-trained checkpoints).

Lukasz Kaiser
authored and
Ryan Sepassi
committed
Make default Parallelism in T2TModel reusing (removes "parallel_0" from variable names in colab).

Lukasz Kaiser
authored and
Ryan Sepassi
committed
fix bug with bpe32k - prepend <pad> and <EOS> to vocab.

nshazeer
authored and
Ryan Sepassi
committed
Disable image summary in eager mode (tf.Eager doesn't like it for now).

Lukasz Kaiser
authored and
Ryan Sepassi
committed
v1.3
Ryan Sepassi
committed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

v1.3#449

Commits on Nov 29, 2017

More tiny sketch fixes

Add Modality.targets_weights_fn

Update layout optimizer config code. pruning and constfold are enabled by default now (no need to explicitly specify them any more).

Preserve static shape info where possible with shape_list and enable constant ones_matrix_band_part

Fix weights_fn calls

Refactor rev_block into a RevBlock class that exposes forward and backward

Clean up revblock code with tf.contrib.framework.nest

Add ImageImagenet224; resnet50 now fits batch size of 128

Added an option to omit quotes from the subtoken voc format.

Bug fix, make class lab a VarLen feature to be compatible with img2img datasets that don't have labels.

Update tpu_trainer to use T2TModel.model_fn and upate RunConfig for non-tpu use

Fix SymbolModality weights fn

Add support for multi-gpu training with tpu_trainer

CHECKPOINT BREAKING: make T2TModel a subclass of Layer so it can be called; all variables are now in model-name scope.

Remove unused functions from transformer_vae.

Update distance computation for k-nearest neighbours to be more efficient, by computing the norms separately.

Pass the modified hparams to the modalities, so they can know the mode.

This change breaks previous checkpoints. Make Transformer fast on TPU.

Changes to make t2t tf.eager compatible

This change breaks previous checkpoints. Make Transformer fast on TPU.

Versions of problems with combined examples. Good for TPU training.

Small bug fix.

When using eager, use slow decoding for transformer models.

merge PRs.

Small updates to T2TModel.call callers

Make daisy_chain_variables and hparam instead of flag and unset to allow LSTM to train in distributed mode.

Move Estimator input_fn and model_fn construction into Problem and T2TModel, respectively, which allows subclassing

This change breaks previous checkpoints. Make Transformer fast on TPU.

Correction for eager-mode decoding scopes (to use with pre-trained checkpoints).

Make default Parallelism in T2TModel reusing (removes "parallel_0" from variable names in colab).

fix bug with bpe32k - prepend <pad> and <EOS> to vocab.

Disable image summary in eager mode (tf.Eager doesn't like it for now).

v1.3