3 ms·

Simple Deep Learning Benchmark

• aizvorski 10y ago
  Here is something to tide you over while waiting for DeepMark... A simple benchmark of TensorFlow, Theano (using Keras), Torch, Caffe and Neon, running the VGG-16 network. Includes results on a GTX 1080. This is also useful as a basic example of how to use each framework. More results and improvements welcome :)
  • visarga 10y ago
    Maybe you can get people with other cards to run the benchmark and add their results to your page.
    • aizvorski 10y ago
      Yes, I hope they do! The benchmark is easy to run if you have CUDA+cuDNN installed.
• TheAlchemist 10y ago
  There is another nice one here: https://github.com/soumith/convnet-benchmarks https://github.com/soumith/convnet-benchmarks
  • aizvorski 10y ago
    Yes, soumith's is a much more comprehensive benchmark. I wanted a new one for a few reasons: - run training including weights updates, not just forward+backward - run on Pascal architecture (latest and greatest) - use cuDNN 5 So mine doesn't try to cover nearly as much ground, but is a closer match to the most common use case of running training on latest hardware right now. When soumith's next project DeepMark is done, that will be the new best benchmark, it looks really good.