Authors:

Anusha Sarraf

Date Created:

Not specified

Course Title:

Professor:

Not specified

About Paper:

Practically implementing a Machine Learning (ML) model involve developing, deploying, using, and reusing models. Deep Learning (DL) model converters facilitate the use of models by moving them from frameworks to runtime environments. The conversion process is however often accompanied by errors which results in degraded model quality and disruption of deployment. Previous work attempts to analyze failures in model conversion to ONNX: Open Neural Network Exchange representation. The location, symptoms, causes, and trends over time of failures are studied but expanding on the causes remains essential. This project mainly focuses on finding bugs responsible for those errors when converting models to ONNX. We have collected data from Stack Overflow and GitHub to analyze the popularity of different conversion representations and investigated the type of issues that happen specifically with the most popular one: ONNX. Our further work involves building a framework for testing ONNX converters that we can use to find bugs. We intend on testing torch.onnx and tf2onnx converters using our framework. When the framework is completed, we plan on using NN-Smith to generate artificial ML models, convert them to ONNX representation, and then compare the outputs of the original and converted models. In cases of discrepancies, we plan to inspect the converted model layer by layer to find bugs. We expect the bugs we find to be noticed and addressed by software developers. We also hope that the framework we built is utilized by other ONNX users to find and report bugs.

Source:

Purdue University / 2023

Topics:

No topics listed

Co-authors:

Anusha Sarraf