Are you able to reproduce the end-to-end latency in the official repo?

[mtli]

Hi, thank you for writing this nice PyTorch implementation of EfficientDet! The official repo says D0's end-to-end latency (CNN + pre- and post-processing) is 10ms without TensorRT. I wonder if your repo can produce similar results?

[rwightman]

@mtli batch size 1 latency will likely never be comparable without some significant work or new optimization features in PyTorch. I think around batch 16-32 and higher they are comparable.

[mtli]

I am not very familiar with TensorFlow, are you suggesting that PyTorch is slower than TensorFlow in this aspect?