Sign in to your account. Urgency This bug is hindering the development of training ops on the DnnlExecutionProvider. Download the mnist model and training data: Mnist 10K data set. Screenshots If applicable, add screenshots to help explain your problem. I have not had a chance to run a git bisect yet to identify the first commit that has the issue. The text was updated successfully, but these errors were encountered:. Sorry, something went wrong. I'll take a look. I used git bisect to find the first instance of the bug and narrowed it down to the following commit.
The error I got was: This is an invalid model. It looks like the graph generation for Gemm may not be properly handling the optional input C. I don't know if the graph generation would also affect the gradient op as well but I suspect it is possible. Since I am not completely sure this issue is directly related to this issue I will make a new issue.
What was the ONNX opset version for the model? For older opsets the 3rd Gemm input was mandatory. For Gemm, the 3rd input is optional only since opset PR assumes "The bias argument to Gemm is optional and defaults to zero", so it implicitly assumes that opset is greater or equal than Not yet. Differential dataflow. Download BibTex. View Publication Projects Naiad. In Proceedings of Big Learning Steno: automatic optimization of declarative queries Derek G. Murray, Michael Isard, and Yuan Yu.
In Proceedings of PLDi Non-deterministic parallelism considered useful Derek G. Murray and Steven Hand. Ciel: a universal execution engine for distributed data-flow computing Derek G. The case for crowd computing Derek G. In Proceedings of MobiHeld
0コメント