HybridBackend 1.0.0

Objectives:

• Memory-efficient loading of categorical data
• Communication-efficient training and evaluation at scale
• Easy to use with existing AI workflows

Features:

1. Performance:
   - Support ORC format in data loading.
   - Support data deduplication.
   - Improve performance of data transfer.
   - Improve performance of loading and shuffling string data.
   - Support workers with unbalanced training data via SyncReplicasDataset.
   - Support pipeline-based semi-synchronous training.
   - Support a hierarchical embedding lookup.

2. Usability
   - Support standalone evaluation and prediction APIs of estimator and keras.

3. Bugfixes:
   - Fix shape calculation of tf.feature_column.shared_embeddings

HybridBackend 0.8.0

Objectives:

1. Memory-efficient loading of categorical data
2. Communication-efficient training and evaluation at scale

Features:

1. Performance
   - Support of automatic embedding fusion on PAI DLC / PAI DSW
   - Support of row-wise shuffling
   - Improves data transfer prefetching

2. Usability
   - Support of embedding_lookup_* API
   - Support of new composable Dataset API

HybridBackend v0.7.0

Objectives:

1. Memory-efficient loading of categorical data
2. GPU-efficient orchestration of embedding layers
3. Communication-efficient training and evaluation at scale
4. Easy to use with existing AI workflows

Features:

1. Performance
   - Support of data transfer prefetching

2. Usability
   - Support of Keras Model API
   - Support direct pip install via Pypi

HybridBackend 0.5.4

Objectives:

• Easy to use with existing AI workflows

Features：

• Support fixed length list in ParquetDataset
• Support schema parsing in ParquetDataset
• Provide validation tools for parquet files

Bug Fixes:

• Fixes indices calculation in rebatching

HybridBackend v0.6.0

Objectives:

1. Communication-efficient training and evaluation at scale
2. Easy to use with existing AI workflows

Features:

1. Data-Parallel Training and Evaluation
   - Bucketized Gradients Aggregation using AllReduce
   - Global Metric Operations
   - Out-Of-Range Coordination

2. Hybrid-Parallel Embedding Learning
   - Bucketized Embedding Exchanging using AllToAllv
   - Fusion and Quantization of AllToAllv
   - Fusion of Partitioning and Stitching

3. Usability
   - Support of MonitoredSession and Estimator
   - Declarative API for Model Definition

4. Compatibility
   - Support of NVIDIA TensorFlow and DeepRec

5. Interoperability
   - Inference Pipeline Needs No Change
   - Support of SavedModel
   - Support of Variable, XDL HashTable and PAI Embedding Variable

Bug Fixes:

[#46] Fixes rebatching in ParquetDataset.

HybridBackend v0.5.3

Objectives:

• Easy to use with existing AI workflows

Features:

• Support working with GPU
• Support building on macOS

HybridBackend v0.5.2

Objectives:

• Memory-efficient loading of categorical data
• Easy to use with existing AI workflows

Features:

1. Parquet Dataset
   - Reading batch of tensors from numeric fields in zero-copy way
   - Reading batch of sparse tensors from numeric list fields in zero-copy way
   - Support of string fields
   - Support of local filesystem, HDFS, S3 and OSS

2. Data Pipeline Functions
   - Resizing batch of tensors and ragged tensors
   - Converting ragged tensors to sparse tensors
   - Objective: "Easy to use with existing AI workflows"

3. Compatibility
   - Support of TensorFlow 1.15 and Tensorflow 1.14
   - GitHub actions for uploading wheels to PyPI

Bug Fixes:

• [#11][#12][#13] Supports manylinux_2_24 platform.