Re: Keras (abstraction layer for Theano and TensorFlow) seeks for an adopter

[lumin <cdluminate@gmail.com>]

Hello everyone,
I'd like to provide some information about this.

Bengio (an important guy behind Theano) declared the end of Theano's
development. So we should not pay too much time on it. Tensorflow,
the computation graph can be considered as a successor of Theano
the symbolic graph engine.

Here is a slide from Stanford's Computer Vision class:
http://cs231n.stanford.edu/slides/2017/cs231n_2017_lecture8.pdf
>From the slides, especially pp. 151-152, we know that
(1) Caffe/Caffe2, Theano/Tensorflow, Torch/Pytorch are what
    these specialists recommend to the students.
(2) Caffe, Caffe2, Tensorflow are recommended for production.
(3) Pytorch is recommended for research.

When linked against cuDNN[1], these deep learning software
will run many times faster than that in CPU mode.
e.g.
  Caffe (CUDA w/o cuDNN) >= 10x * Caffe (CPU)
  Caffe (CUDA with cuDNN) >= 4x * Caffe (CUDA w/o cuDNN)
  This is my "caffe time" test result on Titan X Pascal card,
  with the alexnet deploy version shipped by caffe source.

* Caffe and Torch are currently maintained by me. And I indeed
  used them in my research work.
* Caffe2 is designed for production but its code base is not
  that stable.
* My pytorch packaging is basically done. It just need some
  tweaks and a patch that disables the SIMD instruction sets.
* cuDNN is valuable to package. I have some working packaging
  script for it. Currently I'm contacting nvidia to sort out
  some legal issues...
* Tensorflow the hotspot is certainly valuable to package, but
  it is somewhat difficult. A Debian developer Paul Liu tried
  to work on it (incl. bazel) but things seem to be not that easy.
  Apart from the bazel build, Tensorflow also has a CMake build.
  However that CMake build files need an amount of patches
  to avoid violating the policy. Tensorboard is a noticeable
  tool, which is also very valuable to package. I'm now a fan
  of Pytorch and I currently have no plan to use tensorflow
  in my research work, so I don't want to touch it...

Best.

[1] Magically optimized CUDA deep neural network library.