FCN Alexnet model pruning

Hi,

I have been looking at the example in the jetson-inference repo using TensorRT.
The segnet example using the Nvidia FPV aerial dataset model is pruned.

I found this paper this paper [1510.00149] Deep Compression: Compressing Deep Neural Networks with Pruning, Trained Quantization and Huffman Coding on how to prune a model.
I would like to do the same thing on my custom model. Is it possible to do it directly in DIGITS ? And How ?

Best regards,

Hi Austriker, the process isn’t integrated with DIGITS. At present you would need to implement it from the paper or find an open-source tool or similar.

Hi Dusty,

Thank you for you answer. Does TensorRT 2.1 supports Sparse Matrix ? or does the Nvidia FPV model use another pruning technique which make is usable with tensorRT.

Best regards,

To my knowledge it’s still operating in dense mode, however if you’d like you may be able to define a custom layer plugin for TensorRT2 that invokes the cuSPARSE CSRMV kernel that’s mentioned in the arXiv paper. Even so the performance is not bad.