Can I build and use nvcaffe on TX2?

I suppose the best way of finding out is just to try :-)

However, I notice that there is “jetson-inference” in the dusty-nv github.
It uses TensorRT.
I would prefer to re-use existing C++ code I already have that uses caffe.

I have a trained caffe model (from digits.)
I want to run real-time inference using this model on the Jetson.
I’m not afraid of C++.

Is there any reason why I should use jetson-inference/TensorRT, instead of nvcaffe with cuDNN?

some links on the matters:!forum/digits-users

Thanks. There’s a little bit of performance data in that blog post, although it mostly talks about the training half. I’m not patient enough to run training on the Jetson :-)

nvcaffe building is actually a part of jetson-inference:

My main question is: Why does the main README of jetson-inference do all the surgery work of translating a caffe model to tensorRT to run on the TX2? (This is a problem as not all Caffe layer types are supported in tensorRT)
Is there an actual performance benefit, compared to using 16-bit nvcaffe?


TensorRT compile model based on GPU architecture and lead to better performance.
Here is some comparison between Caffe and tensorRT:

Surgery is to check if all the used layers is available on tensorRT.
(We optimize performance based on the operation a layer targeting for. So not all the layers is supported in tensorRT.)
For supported layer information, please find the document located at ‘/usr/share/doc/’ for details.

I need the crop and upscale layers, as they are used inside the network as part of inference.
Are there plans to implement those layers? If so, when?
Are all the parts of tensorRT open source, so could I, in theory, implement them myself?


Thanks for your feedback.
Please find supported layers in the document located at ‘/usr/share/doc/gie/’

Currently, TensorRT is not open source and can’t implement custom layers.
But in tensorRT-3.0, which is not available yet, you can implement custom layer with custom layer API.

Sorry for that we can’t disclosure any schedule plan.
Please pay attention to our announcement and update.

Thanks and sorry for the inconvenience.

Thank you!

Also, I hope that TensorRT 3.0 implements the crop layer. (It should be reasonably easy :-)