Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog and this project adheres to Semantic Versioning.

[2.12.0] - 2021-04-29

Added

  • function support for Pipeline

Changed

Deprecated

Removed

Fixed

Security

[2.11.0] - 2020-12-11

Added

  • Add new name for layers.Sequential (now layers.DAG for Directed Acyclic Graph), kept old name for legacy reasons

  • Add new argument options to layers.DenseIndex (reuse, trainable and initializer)

Changed

  • Remove shortened imports import deepr as dpr as it is useless.

Deprecated

Removed

  • Unnecessary logging in metrics tensors’ names resolution

Fixed

  • Incorrect numpy requirement (incompatible with Tensorflow requirements)

Security

[2.10.0] - 2020-10-28

Added

  • Log yarn launcher app id with mlflow.

[2.9.1] - 2020-10-01

Fixed

  • move tf-yarn dependency to gpu / cpu additional packages.

[2.9.0] - 2020-10-01

Added

  • SVD jobs and configs to MovieLens

  • VAE model to MovieLens

  • NDCG@K and Recall@K metrics to MovieLens

  • CSVReader to MovieLens

  • AddWithWeight, DenseIndex, MultiLogLikelihoodCSS, MultiLogLikelihood layers

  • Lambda layer to layers.base

  • preds option to Trainer

Changed

[2.8.2] - 2020-09-07

Fixed

  • Convert tuple of FileSystems into a List for skein (not cast by cluster_pack).

[2.8.1] - 2020-09-03

Fixed

  • breaking change in grad norms in Optimizer (store by query name, not variable name)

[2.8.0] - 2020-09-03

Added

  • Accuracy and AccuracyAtK metrics

  • EvaluateJob

  • LSTM layer (using FusedOp)

  • Recursive .glob() in TFRecordReader for nested directory structures

  • Experimental Keras Trainer (converts tf.keras.Model into tf.estimator.Estimator)

  • Add run_id parameter to MlFlow macro (to restart a run)

  • Refactor to_example logic (make arrays_to_example importable to convert NumPy arrays to tf.Example)

  • Add missing macros module import

  • Add batch_shape to Field

Changed

  • __iter__ method of Reader (remove context manager to avoid issue with Keras)

  • use skein_launcher instead of in-house implementation using skein.

Fixed

  • Use skip_steps when computing gradient norms in Optimizer

  • Path copy_file (to support local <> HDFS copies)

[2.7.0] - 2020-08-04

Added

  • Add encoding support to Path / HDFSFile

  • Movielens example

  • TripletPrecision layer

Changed

  • Changed default initializer for embeddings in the embeddings layer

Deprecated

Removed

Fixed

Security

[2.6.0] - 2020-07-02

Added

  • Add table support in ProtoPredict and ProtoExport

[2.5.1] - 2020-07-01

Fixed

  • wrong attribute in CopyDir

[2.5.0] - 2020-07-01

Added

  • Batched support for FromExample (use tf.train.parse_example instead of parse_single_example)

  • Support num_shards_embeddings=None in utils.save_checkpoint

  • Top1, Top1Max, BPRMax, NCE losses

Changed

  • Various versions of requirements in requirements.txt

Deprecated

Removed

Fixed

Security

[2.4.2] - 2020-06-19

Added

  • Allow user to ignore cpu packages shipped in the pex to Yarn

  • URL formatter for MLFlow Macro

  • More checks to vocab.write (check type, newlines)

[2.4.1] - 2020-06-15

Fixed

  • Fix incorrect inputs / outputs resolution in deepr.layers.Select

  • Fix edge cases of ToExample and FromExample (on scalars, tensors with dynamic shapes with ndims > 2)

[2.4.0] - 2020-06-10

Added

  • Utilities to metrics.base: get_tensors, keep_scalars, get_scalars

  • New metrics for variables VariableValue that returns value / global norm of a variable

Changed

  • LastValue metric does not store tensor values in a special metric variable.

[2.3.0] - 2020-06-10

Added

  • writers module, TFRecordWriter

  • ToExample prepro to convert a dataset to tf.Example

  • Field now has a to_feature(value) method

  • iter utils: progress (logs progress every n seconds), chunks to return chunks from an iterable

  • SaveDataset job to write a dataset to tfrecords.

Changed

  • TFRecordSequenceExample renamed FromExample (but old name still available).

  • Field method as_feature renamed feature_specs to avoid confusion with to_feature.

Deprecated

Removed

  • Removed Field methods (leading to incorrect uses): has_var_len, as_feature, has_fixed_len

Fixed

  • Incorrect shuffle argument use in TFRecordReader

Security

[2.2.0] - 2020-06-08

Added

ExportXlaModelMetadata job is added : make it possible to export metadata for xla models Changed ~~~~~~~ Deprecated ~~~~~~~~~~ Removed ~~~~~~~ Fixed ~~~~~ Security ~~~~~~~~

[2.1.1] - 2020-06-05

Added

  • Predictors also yield inputs when applied on a tf.data.Dataset

Changed

Deprecated

Removed

Fixed

Security

[2.1.0] - 2020-06-04

Added

  • Add predictors

  • Add new example job PredictSavedModel

Changed

  • Example job Predict renamed into PredictProto

Deprecated

Removed

Fixed

Security

[2.0.0] - 2020-06-03

Added

  • Doctest run in CI.

Changed

  • copy_dir job will now overwrite the target by default

  • Nested support for prepros.Serial

  • Context manager TableContext for tables reuse

  • Automatic table context creation in prepro.__call__

  • Prepro TableInitializer to run table_initializer_fn before map transforms

  • Vocabulary utilities (read, write, size)

  • Reverse lookup table function index_to_string_table_from_file and associated layer LookupIndexToString

  • Layer combinator ActiveMode to apply layer only on given modes

  • Layer ToFloat

  • Config evaluation modes: skip -> None, instance -> call

  • New evaluation mode for config dictionary partial

  • Remove __post_init__ for YarnTrainer and YarnLauncher to avoid unexpected non-laziness

Deprecated

Removed

  • Use of prepro and layer decorator on constructors

  • Lazy behavior for prepro and layer decorator

Fixed

  • Doctests were fixed.

  • Add metric name sanitizer, especially needed to sanitize keras built variable names.

  • Typo in example predict (feedable / fetchable)

Security

[1.2.1] - 2020-05-27

Added

Changed

Deprecated

Removed

Fixed

Avoid mkdir for HDFS path_model for permissions reasons

Security

[1.2.0] - 2020-05-26

Added

  • OptimizeSavedModel now supports multiple fetches

  • new graph utils, import_graph_def, get_feedable_tensors, get_fetchable_tensors

Changed

  • example.jobs.Predict arguments (path_model and graph_name instead of path_model_pb, fetch instead of fetches for consistency with OptimizeSavedModel).

Deprecated

Removed

Fixed

  • wrong arguments in YarnConfig for upload_zip_to_hdfs.

Security

[1.1.0] - 2020-05-25

Added

  • Remove some kwargs for cleaner error stacks

  • Make example more complex, add advanced notebook

  • Track missing macro

  • Update doc of logging tensor (change prefix to name)

  • Add helper to debug class building from config

[1.0.0] - 2020-05-19

Added

  • Public Release