DPU Utilities

This contains a set of utilities used across projects of the DPU team.

Python

Stored in the python subdirectory, published as the dpu-utils package.

Installation

pip install dpu-utils

OR via the community-maintained Conda recipe:

conda install -c conda-forge dpu-utils

Overview

Below you can find an overview of the utilities included. Detailed documentation is provided at the docstring of each class.

Generic Utilities dpu_utils.utils

• ChunkWriter provides a convenient API for writing output in multiple parts (chunks).
• RichPath an API that abstract local and Azure Blob paths in your code.
• *Iterator Wrappers that can parallelize and shuffle iterators.
• {load,save}_json[l]_gz convenience API for loading and writing .json[l].gz files.
• git_tag_run tags the current working directory git the state of the code.
• run_and_debug when an exception happens, start a debug session. Usually a wrapper of __main__.

General Machine Learning Utilities dpu_utils.mlutils

• Vocabulary map elements into unique integer ids and back. Commonly used in machine learning models that work over discrete data (e.g. words in NLP). Contains methods for converting an list of tokens into their "tensorized" for of integer ids.
• BpeVocabulary a vocabulary for machine learning models that employs BPE (via sentencepiece).
• CharTensorizer convert character sequences into into tensors, commonly used in machine learning models whose input is a list of characters.

Code-related Utilities dpu_utils.codeutils

• split_identifier_into_parts() split identifiers into subtokens on CamelCase and snake_case.
• Lattice, CSharpLattice represent lattices and useful operations on lattices in Python.
• get_language_keywords() an API to retrieve the keyword tokens for many programming languages.
• language_candidates_from_suffix() a function to retrieve the candidate language given the file suffix.
• deduplication.DuplicateDetector API to detects (near)duplicates in codebases. See also here for a command line tool.
• treesitter.parser_for get Tree-sitter parser by language name.

TensorFlow 1.x Utilities dpu_utils.tfutils

• get_activation retrieve activations function by name.
• GradRatioLoggingOptimizer a wrapper around optimizers that logs the ratios of grad norms to parameter norms.
• TFVariableSaver save TF variables in an object that can be pickled.

Unsorted segment operations following TensorFlow's unsorted_segment_sum operations:

• unsorted_segment_logsumexp
• unsorted_segment_log_softmax
• unsorted_segment_softmax

TensorFlow 2.x Utilities dpu_utils.tf2utils

• get_activation_function_by_name retrieve activation functions by name.
• gelu The GeLU activation function.
• MLP An MLP layer.

Unsorted segment operations following TensorFlow's unsorted_segment_sum operations:

• unsorted_segment_logsumexp
• unsorted_segment_log_softmax
• unsorted_segment_softmax

TensorFlow Models dpu_utils.tfmodels

• SparseGGNN a sparse GGNN implementation.
• AsyncGGNN an asynchronous GGNN implementation.

These models have not been tested with TF 2.0.

PyTorch Utilities dpu_utils.ptutils

• BaseComponent a wrapper abstract class around nn.Module that takes care of essential elements of most neural network components.
• ComponentTrainer a training loop for BaseComponents.

Command-line tools

Approximate Duplicate Code Detection

You can use the deduplicationcli command to detect duplicates in pre-processed source code, by invoking

deduplicationcli DATA_PATH OUT_JSON

where DATA_PATH is a file containing tokenized .jsonl.gz files and OUT_JSON is the target output file. For more options look at --help.

An exact (but usually slower) version of this can be found here along with code to tokenize Java, C#, Python and JavaScript into the relevant formats.

Tests

Run the unit tests

python setup.py test

Generate code coverage reports

# pip install coverage
coverage run --source dpu_utils/ setup.py test && \
  coverage html

The resulting HTML file will be in htmlcov/index.html.

.NET

Stored in the dotnet subdirectory.

Generic Utilities:

• Microsoft.Research.DPU.Utils.RichPath: a convenient way of using both paths and Azure paths in your code.

Code-related Utilities:

• Microsoft.Research.DPU.CSharpSourceGraphExtraction: infrastructure to extract Program Graphs from C# projects.

Contributing

This project welcomes contributions and suggestions. Most contributions require you to agree to a Contributor License Agreement (CLA) declaring that you have the right to, and actually do, grant us the rights to use your contribution. For details, visit https://cla.microsoft.com.

When you submit a pull request, a CLA-bot will automatically determine whether you need to provide a CLA and decorate the PR appropriately (e.g., label, comment). Simply follow the instructions provided by the bot. You will only need to do this once across all repos using our CLA.

This project has adopted the Microsoft Open Source Code of Conduct. For more information see the Code of Conduct FAQ or contact opencode@microsoft.com with any additional questions or comments.