ssd-mobilenet 300x300 onnx model converted from tensorflow can be optimized.

mlcommons / inference

Reference implementations of MLPerf™ inference benchmarks

https://mlcommons.org/en/groups/inference

Apache License 2.0

1.24k stars 536 forks source link

ssd-mobilenet 300x300 onnx model converted from tensorflow can be optimized. #1182

Open hwangdeyu opened 2 years ago

hwangdeyu commented 2 years ago

ssd-mobilenet 300x300 from https://github.com/mlcommons/inference/tree/r1.1/vision/classification_and_detection.

The ZerosLike inserts Casts before Mul issue can be optimized refer: https://github.com/onnx/tensorflow-onnx/issues/1997 which can improve the inference performance.

Optimized model opset 11 and opset 15. ssd_mobilenet_v1_coco_2018_01_28_opset11.zip ssd_mobilenet_v1_coco_2018_01_28_opset15.zip

arjunsuresh commented 2 years ago

Hi @hwangdeyu , thank you for the new model files. Do you have the performance comparison of the models? ssd-mobilenet is removed from inference in closed division but we can still use it to submit results to open division.