ONNX Converters SIG Chin Huang IBM Guenther Schmuelling

ONNX Converters SIG Chin Huang (IBM) Guenther Schmuelling (Microsoft) Workshop 8/23/2019

Quick update • Operations • • Biweekly meeting, every other Wed Alternate times, 9 am and 5 pm Pacific Time Open discussion, https: //gitter. im/onnx/converters Issues, notes, documentation on https: //github. com/onnx/sigs • SIG goals • Provide guidelines and recommendations for all converters to implement to promote ü a. consistent quality and clarity across ONNX converters ü b. ease of understanding and use for ONNX users • Gather community input and feedback to prioritize converter requirements and features • List of converters • https: //github. com/onnx/sigs/blob/master/converters/docs/Converters. List. md • Roadmap, common issues and features…

Breakout session agenda • Roadmap • Inference • Training • Quantization • Pain points • Review of current issues, investigations, and recommendations • Discussions on new issues, common features • Identify action items

Roadmap • Inference • more models • new framework releases like Tensorflow 2. 0 • Training (long term) • figure out how to support the training proposal • get multiple converters to evaluate • Quantization • how to support quantization • Do we need to have testing for quantization in converters?

Pain points: Current issues, investigations, and recommendations • • • What is the role of ONNXMLTools? Do we want standardized frontend test? Clearly define Converter and provide list of converters #13 Evaluate ONNX training proposal #12 Where/how to see current state (coverage, usage, etc) for all converters #15 How to handle unsupported data types in converters? #14

Pain points: New issues and common features • • • How to handle deprecated operators? #18 How to support and verify dynamic shape and rank in inputs? #20 How to support and verify onnx-ml operators? Add custom attributes to converters (Graphcore) How to handle python loops and dicts? (Preferred Networks) • Others…

Backup

PR in onnx-tensorflow: https: //github. com/onnx-tensorflow/pull/484 8

Unsupported data types • Mod • ONNX: uint 8, uint 16, uint 32, uint 64, int 8, int 16, int 32, int 64, float 16, float, double • Tensorflow: int 32, int 64, bfloat 16, half, float 32, float 64 • Input type: int 16 • no optional arguments: exception • --autocast upcast: working • --autocast upcast --logging_level debug: working with console log • Sample API call • config_params = { "logging_level" : logging_level, "autocast_level" : autocast_level } • prepare(onnx_model, **config_params)