Commits

Breadcrumbs

History for

onmaster

User selector

All users

Datepicker

All time

Commit History

Commits on May 27, 2021

• Add missing typing.Optional type annotations to function parameters.

  Show description for 0566f08
  

  rchen152

  authored and

  copybara-github

  committed
  0566f08

Commits on May 26, 2021

• Set prefer_int8_to_int32_dot to True by default.

  Show description for 08e9fcd
  

  shivaniag

  authored and

  copybara-github

  committed
  08e9fcd

Commits on May 25, 2021

• Internal changes only.

  Show description for c9509dc
  

  shivaniag

  authored and

  copybara-github

  committed
  c9509dc

• Fix typo

  Show description for 44a7cbc
  

  reinerp

  authored and

  copybara-github

  committed
  44a7cbc

• Removes prepare_and_save_report; we do have functionality save report yet.

  Show description for 0344f02
  

  shivaniag

  authored and

  copybara-github

  committed
  0344f02

• Adds dacite package to the list of requirements, is not pre-installed on the standard TPU VM.

  Show description for c73ec37
  

  shivaniag

  authored and

  copybara-github

  committed
  c73ec37

Commits on May 24, 2021

• Adds training script to train imagenet.

  Show description for 8344c56
  

  shivaniag

  authored and

  copybara-github

  committed
  8344c56

• Merge of PR google-research#691

  Show description for 79b5796
  

  clee1994

  authored and

  copybara-github

  committed
  79b5796

Commits on May 20, 2021

• Threshold the scaling value to the smallest representable positive value in quantized layer normalization to prevent scaling by 0 when the number of features is large and precision is low.

  Show description for 12256bc
  

  dmmolitor

  authored and

  copybara-github

  committed
  12256bc

• BatchNorm update bug fix. Replaced mutable=True in model.apply() during training with mutable=['batch_stats', 'get_bounds'], otherwise BN statistics would not get updated during training due …

  Show description for ae9d07f
  

  lisa-1010

  authored and

  copybara-github

  committed
  ae9d07f

Commits on May 3, 2021

• Remove instances of uint32 that were unnecessary or otherwise being immediately cast away.

  Show description for cc996ef
  

  phoenix-meadowlark

  authored and

  copybara-github

  committed
  cc996ef

Commits on Apr 28, 2021

• Adds Resnet model and input pipeline for ImageNet data. Also adds config schema and configs to optionally quantize Resnet model.

  Show description for fda69c3
  

  shivaniag

  authored and

  copybara-github

  committed
  fda69c3

• Adds config_schema_utils to utils; utility for adding config schema to models.

  Show description for d18752b
  

  shivaniag

  authored and

  copybara-github

  committed
  d18752b

Commits on Apr 27, 2021

• Open-sourcing util functions for compute cost analysis and HLO parsing of neural networks. Also open-sourcing hlo_pb2 py_proto as we require it for HLO analysis.

  Show description for 32731b9
  

  lisa-1010

  authored and

  copybara-github

  committed
  32731b9

Commits on Apr 23, 2021

• Disallows casting of activations with positive distribution to int8; bug fix.

  Show description for 0782952
  

  shivaniag

  authored and

  copybara-github

  committed
  0782952

• Move hparams_utils to third_party.

  Show description for 412e0ef
  

  shivaniag

  authored and

  copybara-github

  committed
  412e0ef

Commits on Apr 22, 2021

• Move all utilities files in folder aqt/utils.

  Show description for b12b575
  

  shivaniag

  authored and

  copybara-github

  committed
  b12b575

• Move fp_cast file to JAX folder; contains down-casting functionality for JAX.

  Show description for 123dc82
  

  shivaniag

  authored and

  copybara-github

  committed
  123dc82

Commits on Apr 17, 2021

• Open-sourcing util functions necessary for model training.

  Show description for d154ffb
  

  lisa-1010

  authored and

  copybara-github

  committed
  d154ffb

• Adds AQT (Accurate Quantized Training) libraries for jax and linen layers; AQT is what you serve is what you train quantization library.

  Show description for 0ab2247
  

  shivaniag

  authored and

  copybara-github

  committed
  0ab2247

Commits on Mar 12, 2021

• Wrapper functions for common pandas dataframe operations.

  Show description for 45aff62
  

  lisa-1010

  authored and

  copybara-github

  committed
  45aff62

Commits on Mar 10, 2021

• Update gradient for floating-point downcast.

  Show description for 5e50235
  

  malmaud

  authored and

  copybara-github

  committed
  5e50235

Commits on Feb 27, 2021

• Docstring fix.

  Show description for 243c923
  

  lisa-1010

  authored and

  copybara-github

  committed
  243c923

Commits on Feb 20, 2021

• Added experiment analysis helper functions.

  Show description for dbdf119
  

  lisa-1010

  authored and

  copybara-github

  committed
  dbdf119

Commits on Feb 19, 2021

• Provided additional usage example in README.

  Show description for 94cae50
  

  lisa-1010

  authored and

  copybara-github

  committed
  94cae50

• Small improvements in documentation.

  Show description for a7a58c3
  

  lisa-1010

  authored and

  copybara-github

  committed
  a7a58c3

Commits on Feb 17, 2021

• Add reporting tool, for creating experiment reports from TfEvent files.

  Show description for b62b212
  

  lisa-1010

  authored and

  copybara-github

  committed
  b62b212

Commits on Feb 12, 2021

• Add util functions to parse TFEvent files.

  Show description for d4d85f5
  

  lisa-1010

  authored and

  copybara-github

  committed
  d4d85f5

Commits on Jan 22, 2021

• Fix gradient of floating-point quantization function.

  Show description for e23e4f6
  

  malmaud

  authored and

  copybara-github

  committed
  e23e4f6

Commits on Jan 15, 2021

• Colab example taking gradients wrt flax variables inside a flax module.

  Show description for a2b80f5
  

  lisa-1010

  authored and

  copybara-github

  committed
  a2b80f5

Commits on Jan 13, 2021

• type casting for quantization aware training

  Show description for 8f53eae
  

  rybakov

  authored and

  copybara-github

  committed
  8f53eae

Loading