Symptom

Unsatisfactory inference results refer to the following two situations:

• Wrong inference result. For example, the mean average precision (mAP) result of an object detection network is all 0s, or the cosine similarity between the inference result of your OM and that of the benchmark network is 0.
• Inference accuracy drop. For example, the cosine similarity between the inference result of your OM and that of the benchmark network is over 95%, but an accuracy gap is noticed:
  • Classification task: OM (top 1: 0.90, top 5: 0.70) vs benchmark network (top 1: 0.92, top 5: 0.71)
  • Detection task: offline model (MAP: 0.54) vs benchmark network (MAP: 0.55)