Performance Benchmarking hctorch using CNN and ConvNet benchmarks

[Former user]

Hi Neelakandan, Thank you for coding hctorch and hcnn packages. Like many other machine learning practitioners, I am also looking forward to a strong alternative to NVidia cuDNN from AMD.

 I would like to request you to perform a benchmark and share how the library compares to cuDNN and fbnn.  Can you please run these two benchmarks with hctorch and share the results?

1. Soumith Chintala's benchmark: https://github.com/soumith/convnet-benchmarks
2. Justin Johnson's benchmark: https://github.com/jcjohnson/cnn-benchmarks

Thank you!

[Kyzer F]

Would definitely love to see this!

[Neelakandan Ramachandran]

Over ROCM 1.0 and Fiji I was able to currently run just alexnet and Overfeat. VGG and Oxford couldn't be run. I was hitting mem alloc issues. May be fiji is too small to accomodate VGG and Oxford Below are the numbers for Alexnet and Overfeat nets. There is still lot of stuffs to be exploited.

currently hctorch just exploits one command queue and there isn' any kernel data transfer overlap. the next step will be to get this

ModelType: AlexNet Kernels: nn Input shape: 128x3x224x224
nn :updateOutput(): 358.44
nn :updateGradInput(): 153.80
nn :accGradParameters(): 372.74
nn :Forward: 358.44
nn :Backward: 526.54
nn :TOTAL: 884.98
ModelType: OverFeat[fast] Kernels: nn Input shape: 128x3x231x231
nn :updateOutput(): 1058.56
nn :updateGradInput(): 483.62
nn :accGradParameters(): 634.47
nn :Forward: 1058.56
nn :Backward: 1118.08
nn :TOTAL: 2176.65

[Kyzer F]

Thanks for the update Neelakandan!

Is there a feature roadmap and estimated release date for hctorch?