Closed Hrayo712 closed 4 years ago
After evaluating the Frozen graph of the softmax version of the ss_dmobilenetv2 it was found that Optimization on TensorRT at FP16 degrades precision on approximately 2% +/- 1. This will be documented as part of the results. Closing the Issue.
It is necessary to confirm the mAP of the TensorRT Engine is consistent regarding the mAP of the Tensorflow Graph. This will give an insight on the performance degradation (if any), as a tradeoff between inference speed and accuracy of the model. Results should be similar (or identical).