文档

Description and Restrictions

Description

  • Former comparison modes will not be updated. New comparison modes are recommended.
  • Former comparison modes support only dump files after Protobuf serialization. Dump files in custom formats are not supported.
  • The network-wide and single-operator tensor comparisons are supported. You can select a comparison mode as required.

Restrictions

  • Dump data of the offline model and the .npy file or dump data of the third-party model (from Caffe or TensorFlow) to be compared should be obtained from the counterpart models. Otherwise, the comparison results of only the counterpart operators are displayed.
  • For the Fast R-CNN network, the comparison result is subject to the accuracy of the FSRDetectionOutput operators. It is justifiable that the ProposalD operator and its downstream operators offer low accuracy.
  • During graph build, if some operators of the graph are fused, the outputs of these operators can no longer be found in the built model. As a result, the comparison of these operators is unavailable.
  • During graph build, if the structure of a graph is modified (such as stride, L1 fusion, and L2 fusion), the comparison of the inputs or outputs of the operators is unavailable.
  • If counterpart operators require different shapes (for example, the offline model operator requires a reduced shape), or format conversion is not supported, the comparison of these operators is unavailable.
  • If data is preprocessed (for example, the input of the data operators is set to YUV in AIPP), the input format of the data operators may be different from that of the original model, leading to an unreliable comparison result.
  • If the corresponding fusion pattern is enabled, a quantized operator will be fused with its upstream operator. As a result, the comparison result of the operator output is unreliable.
  • In a quantized model, the comparison of quantized operators is available only after they are dequantized. Comparison of non-quantized operators in a quantized model is available. For example, in a quantized model, the comparison of the output of the AscendQuant operator is not available.
Parent topic: Tensor Comparison