Overview

The computation results on the Ascend AI Processor of Huawei proprietary operators may be different from those of third-party equivalents (for example, from Caffe and ONNXTensorFlow).

• Model optimization performed during conversion, including operator elimination, fusion, and partition, may cause the computation results of Huawei proprietary operators to be different from those of third-party equivalents (for example, from Caffe, TensorFlow, and ONNX).
• Original networks can be migrated to the Atlas training products for training, but in this case, the computation results of Huawei proprietary operators may be different from those of third-party equivalents (for example, from TensorFlow).

Definition: To quickly resolve the operator accuracy issues, a tool must be designed to compare the accuracy difference of computation results of Huawei proprietary operators and third-party equivalents.

Function: Model Accuracy Analyzer provides the following tensor comparison methods: cosine similarity, Euclidean relative distance, absolute errors (maximum absolute error, mean absolute error, and root mean squared error), relative errors (maximum relative error, mean relative error, and cumulative relative error), Kullback-Leibler divergence, and standard deviation.