Commits

Branch selector

 master

User selector

All users

Datepicker

All time

Commit History

Commits on Nov 2, 2018

• Prop quantization hyperparams to depthwise sub-layer of the SeparableConv2D layer.

  Show description for e9d0e64
  

  laurenzo

  authored and

  Copybara-Service

  committed
  e9d0e64

  Copy full SHA for e9d0e64

Commits on Nov 1, 2018

• Add residual quant activation tracking for StackedFRNNLayerByLayer.

  Show description for 61238ef
  

  laurenzo

  authored and

  Copybara-Service

  committed
  61238ef

  Copy full SHA for 61238ef

• Add qsubtract alias function.

  Show description for 02d58be
  

  laurenzo

  authored and

  Copybara-Service

  committed
  02d58be

  Copy full SHA for 02d58be

• Propagate static shape in ProjectionLayer.

  Show description for 25bf644
  

  jonathanasdf

  authored and

  Copybara-Service

  committed
  25bf644

  Copy full SHA for 25bf644

Commits on Oct 31, 2018

• Fix RNNT decoder memory usage issue.

  Show description for 10dcea6
  

  yonghui

  authored and

  Copybara-Service

  committed
  10dcea6

  Copy full SHA for 10dcea6

• Apply gradient clipping before writing the histogram to avoid "Nan in summary histogram" errors which cause the controller to crash.

  Show description for 4845ef9
  

  lingvo-bot

  authored and

  Copybara-Service

  committed
  4845ef9

  Copy full SHA for 4845ef9

• Add a function to change the dtype of params recursively.

  Show description for b89bc7c
  

  rohan-anil

  authored and

  Copybara-Service

  committed
  b89bc7c

  Copy full SHA for b89bc7c

Commits on Oct 30, 2018

• Add speech synthesis papers to README.

  Show description for 006c92d
  

  jonathanasdf

  authored and

  Copybara-Service

  committed
  006c92d

  Copy full SHA for 006c92d

• Fix a bug in rnn_cell.py where _FPropDtype should be migrated to

  Show description for 5e7e331
  

  yonghui

  authored and

  Copybara-Service

  committed
  5e7e331

  Copy full SHA for 5e7e331

• Support bfloat16 for HypsFromBeamSearchOuts.

  Show description for fa58d4d
  

  rohan-anil

  authored and

  Copybara-Service

  committed
  fa58d4d

  Copy full SHA for fa58d4d

• Use same type as input for clipping cap.

  Show description for 9c446ff
  

  rohan-anil

  authored and

  Copybara-Service

  committed
  9c446ff

  Copy full SHA for 9c446ff

• Add padding_algorithm and quantizable support to PoolingLayer.

  Show description for 2eac768
  

  laurenzo

  authored and

  Copybara-Service

  committed
  2eac768

  Copy full SHA for 2eac768

• Drop Mix Matmul-FakeQuant for SimpleLstmCell with Bias

  Show description for a4c5027
  

  rsuderman

  authored and

  Copybara-Service

  committed
  a4c5027

  Copy full SHA for a4c5027

• Adds target_sequence_sampler, a helper library to sample a sequence from a decoder and decoder inputs.

  Show description for 36da238
  

  rpang

  authored and

  Copybara-Service

  committed
  36da238

  Copy full SHA for 36da238

• Fix variance computation.

  Show description for 60dbfdc
  

  drpng

  authored and

  Copybara-Service

  committed
  60dbfdc

  Copy full SHA for 60dbfdc

• Moves creation of 'beam_search' to BaseBeamSearchDecoder.

  Show description for 40d9be0
  

  rpang

  authored and

  Copybara-Service

  committed
  40d9be0

  Copy full SHA for 40d9be0

Commits on Oct 29, 2018

• external_max_length can be a Tensor in tokenizer inputs.

  Show description for 8c11898
  

  jonathanasdf

  authored and

  Copybara-Service

  committed
  8c11898

  Copy full SHA for 8c11898

• Added quantization to the simple embedding layer.

  Show description for 48af5c4
  

  rsuderman

  authored and

  Copybara-Service

  committed
  48af5c4

  Copy full SHA for 48af5c4

• Adds a debug helper that saves intermediate tensors to separate files.

  Show description for 5309119
  

  zhifengc

  authored and

  Copybara-Service

  committed
  5309119

  Copy full SHA for 5309119

• Verify that sequence_length is set on RNN layers.

  Show description for 3181ef3
  

  lingvo-bot

  authored and

  Copybara-Service

  committed
  3181ef3

  Copy full SHA for 3181ef3

• Moves definition of 'target_seq_len' to BaseBeamSearchDecoder.Params().

  Show description for f60cf0e
  

  rpang

  authored and

  Copybara-Service

  committed
  f60cf0e

  Copy full SHA for f60cf0e

• Adds accuray5 summary.

  Show description for a97343a
  

  zhifengc

  authored and

  Copybara-Service

  committed
  a97343a

  Copy full SHA for a97343a

• Adds FPropMeta method to DropoutLayer.

  Show description for 224ab7c
  

  zhifengc

  authored and

  Copybara-Service

  committed
  224ab7c

  Copy full SHA for 224ab7c

Commits on Oct 27, 2018

• Add DepthwiseConv2DLayer, SeparableConv2DLayer and plumbing to make ConvSetLayer handle fake_quant feeding into its concat properly.

  Show description for a464893
  

  laurenzo

  authored and

  Copybara-Service

  committed
  a464893

  Copy full SHA for a464893

• Add option to ensure a full beam in beam search

  Show description for 87b6f3b
  

  anjuli

  authored and

  Copybara-Service

  committed
  87b6f3b

  Copy full SHA for 87b6f3b

Commits on Oct 26, 2018

• Retain shape information if static rank is known in PadSequenceDimension.

  Show description for 5708984
  

  jonathanasdf

  authored and

  Copybara-Service

  committed
  5708984

  Copy full SHA for 5708984

• Extend ConvLayer to support folded batch norm and fake_quant for compatibility with TFLite quantized convolution kernels.

  Show description for f1dd51c
  

  laurenzo

  authored and

  Copybara-Service

  committed
  f1dd51c

  Copy full SHA for f1dd51c

• Adds 'theta' to beam search callbacks.

  Show description for 5a146cb
  

  rpang

  authored and

  Copybara-Service

  committed
  5a146cb

  Copy full SHA for 5a146cb

Commits on Oct 24, 2018

• Add LayerNormalizedLSTMCellLean which is more performant on TPU.

  Show description for 6c0fc6e
  

  yonghui

  authored and

  Copybara-Service

  committed
  6c0fc6e

  Copy full SHA for 6c0fc6e

Commits on Oct 23, 2018

• Implement FPROP_META for identity layer.

  Show description for 69c7db1
  

  bignamehyp

  authored and

  Copybara-Service

  committed
  69c7db1

  Copy full SHA for 69c7db1

• Added a GetQTensorRange and GetTensorRange for QLayer/QDomain

  Show description for 5d92545
  

  rsuderman

  authored and

  Copybara-Service

  committed
  5d92545

  Copy full SHA for 5d92545

• Adds py_utils.RunOnTpuHost(), a helper function to run TPU-incompatible logic on host CPU.

  Show description for 25c57fc
  

  rpang

  authored and

  Copybara-Service

  committed
  25c57fc

  Copy full SHA for 25c57fc

• Remove ComputeLoss function from MTDecoderV1 since it's the same as the overridden function.

  Show description for 273c2c8
  

  lingvo-bot

  authored and

  Copybara-Service

  committed
  273c2c8

  Copy full SHA for 273c2c8

Commits on Oct 22, 2018

• Internal change

  Show description for 6c89b10
  

  lingvo-bot

  authored and

  Copybara-Service

  committed
  6c89b10

  Copy full SHA for 6c89b10

• Fix flaky record_yielder_test.

  Show description for 69e697a
  

  jonathanasdf

  authored and

  Copybara-Service

  committed
  69e697a

  Copy full SHA for 69e697a

Pagination

Previous

Next