search

chenzhi1992 / TensorRT-SSD

Use TensorRT API to implement Caffe-SSD， SSD（channel pruning）， Mobilenet-SSD
251 stars 84 forks source link

Related to performance gain #23

Open jagadeesh09 opened 6 years ago

jagadeesh09 commented 6 years ago

Hi,

I would like to know the performance gain for SSD using Tensorrt.

chenzhi1992 commented 6 years ago

ssd+vgg16： inference time 80-100ms ssd+Mobilenetv1： inference time 15-20ms

jagadeesh09 commented 6 years ago

Hi,

Thanks for the reply! I am working with Tesla K80 GPU and x86 architecture. I have experimented with Tensorrt by using different network architectures. For faster r-cnn ( from the samples), there is performance gain by just deploying it using Tensorrt. Whereas, for SSD architecture, I couldn't observe any such performance gain over the caffe deployment.

And I have tested Tensorrt for VGG architecture. Here also, I couldn't observe any performance gain. Time taken for Tensorrt and Caffe deployments for VGG architecture is almost same i.e 25 ms.

If possible, can you please comment with your numbers i.e not by using tensorrt for SSD on your machine?

Thanks, Jagadeesh

chenzhi1992 commented 6 years ago

Hi, In my jetson-tx2, I got twice the acceleration by using TensorRT on the ssd-vgg16 architecture. Maybe you have a problem with the implementation of your TensorRT code.

Optimus1072 commented 6 years ago

@jagadeesh09 Will it be possible for you to give some pointers to implement SSD using TensorRT