Zheng-Bicheng commented on PR #16653: URL: https://github.com/apache/tvm/pull/16653#issuecomment-1984932842
> Generally speaking, I suggest adding tests that will exercise the path being proposed here, that is from PaddlePaddle to CMSIS-NN, including at least one softmax operator. Does that make sense? Are you suggesting using a PaddlePaddle model with softmax parameters and specifying the runtime as CMSIS-NN, then validating whether the output results of CMSIS-NN match those of the PaddlePaddle model? This approach might not be feasible at the moment. I've already highlighted the potential issues with this method in TVM Pull Request 16651. In simple terms, in the current version of TVM, when a quantized PaddlePaddle model is converted to a TVM model, there are discrepancies in the model's computation results. I'm confident this isn't an issue with my porting efforts because the same problem exists with ONNX models. You can review my detailed test code in [TVM Pull Request 16651](https://github.com/apache/tvm/pull/16651), where I convert the quantized Paddle model to a TVM model and specify the target to be llvm running on CPU. -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected]
