Accuracy Tuning Scenarios

If the accuracy does not meet the expectation during model porting from GPU/CPU to NPU for training, or during the iteration of the NPU training version, for example, the loss curve or validation accuracy does not meet the expectation, you can refer to this section to tune the accuracy and find the faulty operators or components.

There are four types of scenarios where the accuracy does not meet the expectation:

• The accuracy does not meet the expectation during model porting from GPU/CPU to NPU for training. You can refer to Comparison Between the GPU/CPU and NPU Networks to tune the accuracy.
• The accuracy does not meet the expectation after the software version or configuration changes during the continuous iterative training of the model on the NPU. You can refer to NPU vs. NPU Network to tune the accuracy.
• The output value of the model is NaN during model training on the NPU. You can refer to NaN Overflow Locating to tune the accuracy.
• The random accuracy error occurs during model training on the NPU. That is, the accuracy may not meet the expectation in some training. You can refer to Random Error Location to tune the accuracy.

If you encounter any problem during the tuning or have any difficulty in performing operations, you are advised to visit the GitCode community to submit feedback or seek help.