Accuracy Debugging
Training DevelopmentInference Development
Offers debugging features including data dump, overflow/underflow detection, and accuracy comparison, outputting detailed accuracy diagnostics to help you quickly isolate and identify the root causes of model accuracy issues.
Data dump obtains accuracy debugging data, including the inputs and outputs of the model intermediate layer.
Overflow/Underflow detection detects overflows/underflows during model running and checks for INF/NAN values.
Accuracy comparison compares the model's intermediate layer data with GPU/CPU results and outputs accuracy evaluation metrics.




