Thanks to visit codestin.com
Credit goes to github.com

Commits

Commits on Apr 12, 2022

Allow to propagate `TunerManagedError` to the caller of base_tuner

lingvo-bot
authored and
copybara-github
committed
Softmax logit cap.

lingvo-bot
authored and
copybara-github
committed
Remove the init_op parameter from Checkpointer constructor. Based on existing

aaroey
authored and
copybara-github
committed

Commits on Apr 11, 2022

Support adapter in ConformerLayer

protoget
authored and
copybara-github
committed
Import `MutableMapping` from `collection.abc` instead of `collections`.

xldrx
authored and
copybara-github
committed

Commits on Apr 9, 2022

internal change

lingvo-bot
authored and
copybara-github
committed

Commits on Apr 8, 2022

Write trainer params in proto format in addition to text.

lingvo-bot
authored and
copybara-github
committed
More general unrolling in pipeline layer

lingvo-bot
authored and
copybara-github
committed

Commits on Apr 6, 2022

fix the bug under third_party/py/lingvo/tasks/mt/input_generator.py that if self.params.mass_layer is None, self.mass_layer does not exist and complains 'mass_layer' lookup failure.

lingvo-bot
authored and
copybara-github
committed

Commits on Apr 5, 2022

Refactor lr schedule so that it can be determined from optimizer state, e.g. number of parameter update steps instead of only global_step.

lingvo-bot
authored and
copybara-github
committed

Commits on Apr 4, 2022

Pass the weight of current step's gradient into optimizer.Apply, so that accumulated gradients from different steps can be normalized by these weights; Add a DynamicGradientAccumulator

lingvo-bot
authored and
copybara-github
committed

Commits on Apr 1, 2022

Commits on Mar 31, 2022

Commits on Mar 30, 2022

Add scaling options for input embeddings and decoder outputs.

lingvo-bot
authored and
copybara-github
committed
Add layer norm implemented in the same way as Lingvo Jax.

lingvo-bot
authored and
copybara-github
committed
Fixes chunkwise attention selection logic.

WeiHan3
authored and
copybara-github
committed
Internal changes only; it defined wrapped matmul/conv/einsum op for optional quantization, but do not use these anywhere external.

shivaniag
authored and
copybara-github
committed
Don't restore inside the loop, it's not needed.

aaroey
authored and
copybara-github
committed
When TPU embedding retrieve ops are used:

aaroey
authored and
copybara-github
committed
Let Saver able to wait for (async) saving to finish before exiting the

aaroey
authored and
copybara-github
committed
pass domain embedding in conformer layer.

lingvo-bot
authored and
copybara-github
committed

Commits on Mar 29, 2022

Add option to allow using truncated sequences in TextPackedInput.

Ankur Bapna
authored and
copybara-github
committed
Adding camera labels.

lingvo-bot
authored and
copybara-github
committed
[XLA] Fix copy insertion bug where copies may be inserted at wrong while level in nested whiles

lingvo-bot
authored and
copybara-github
committed

Commits on Mar 28, 2022

Remove name check for variables that have non-static shape (CuDNN support was

aaroey
authored and
copybara-github
committed
Avoid setting self.enqueue_op when TPU embedding layer is used, so we can run

aaroey
authored and
copybara-github
committed

Commits on Mar 23, 2022

The `GenericInputV2` ops rely on resource handlers for creation/reuse control. We introduced a global cache to store these resource handlers in Python. To accommodate this a new kwarg is introduced…

lingvo-bot
authored and
copybara-github
committed

Commits on Mar 22, 2022

Enable custom saver by default.

aaroey
authored and
copybara-github
committed
Add option to enable bidirectional relative attention for decoders.

lingvo-bot
authored and
copybara-github
committed
Update docker build with TensorStore and list it as a dependency.

laurentes
authored and
copybara-github
committed