SparseZoo v1.3.0

New Features:

• BERT models added for the GoEmotions multi-label dataset.
• BERT models added for SQuAD 2.0 dataset.
• oBERTa base models added for GLUE datasets.
• YOLOv5 and YOLOv5p6 models added for transfer learning.

Changes:

• Minimum Python version changed to 3.7.
• Benchmarking and accuracy metrics for a model propagated to the root Python class.

Resolved Issues:

• None

Known Issues:

• None

SparseZoo v1.2.0

New Features:

• SparseZoo ONNX analysis API added to enable easy model analysis for sparsity, quantization, flops, parameters, and more.
• BERT document classification models added, trained on the IMDB dataset.

Changes:

• Tokenizer_config.json added as required file for transformers models.
• Minimum Python version changed to 3.7 as 3.6 as reached EOL.

Resolved Issues:

• SparseZoo README updated to reflect new APIs and flows that were released with 1.1 release.

Known Issues:

• None

SparseZoo v1.1.0

New Features:

• Python 3.10 supported.
• SparseZoo APIs refactored, focusing on better and easier-to-use functionality. The model class is now the core interface and works the same for both SparseZoo stubs as well as local folders and enables storing models from the SparseZoo in any local folder.
• Model Analysis API added, enabling detailed layer and operator information for operations, pruning, quantization, parameter counts, and more.
• BERT-base-cased models added.
• BioBERT models added.
• Quantized BERT-base MLM models added for support for Information Retrieval pipelines.
• Compound (structured pruning, unstructured pruning, quantization) ResNet-50 models added.

Changes:

• None

Resolved Issues:

• None

Known Issues:

• None

SparseZoo v1.0.0

New Features:

• None

Changes:

• Download progress bar changed to use byte units.
• Click as a root dependency added as the new preferred route for CLI invocation and arg management.

Resolved Issues:

• Protobuf version pinned to prevent issues with ONNX dependencies.

Known Issues:

• None

SparseZoo v0.12.1 Patch Release

This is a patch release for 0.12.0 that contains the following changes:

• Protobuf is restricted to version < 4.0 as the newer version breaks ONNX.

SparseZoo v0.12.0

New Features:

• DistilBERT models added.

Changes:

• None

Resolved Issues:

• None

Known Issues:

• None

SparseZoo v0.11.0

New Features:

• New ResNet-50 models added with improved performance and recovery.
• BERT 12-layer token classification recipes added.
• BERT 3-layer and 6-layer models added.

Changes:

• Documentation updated to align with UI.

Resolved Issues:

• The package version check is now wrapped in an exception so that the background thread for checking does not error out when outbound network restrictions are in place.

Known Issues:

• None

SparseZoo v0.10.0

New Features:

• sparsezoo.download CLI added to easily download models from the SparseZoo by their stub.

Changes:

• Docs updated to reflect latest state of the SparseZoo UI.
• Version check API updated to use POST rather than GET methods to prevent potential caching issues.

Resolved Issues:

• Wrapped tests addressed that were failing intermittently with flaky to prevent improper test failures.

Known Issues:

• None

SparseZoo v0.9.0

New Features:

• Up-to-date version check implemented for SparseZoo.
• Sparse and sparse-quantized YOLACT models deployed.

Changes:

• API docs recompiled.

Resolved Issues:

• None

Known Issues:

• None

SparseZoo v0.8.0

New Features:

• Support added for more recipe types and names in preparation for BERT sparse transfer learning.
• Sparse BERT mask language modeling models pushed to the SparseZoo with example recipes for transferring to other downstream datasets.

Changes:

• None

Resolved Issues:

• None

Known Issues:

• None