3 ms·
Comparative Study of Caffe, Neon, Theano, and Torch for Deep Learning
- p1esk 11y agoIt's a bummer they published this just before TensorFlow release...
- mindcrime 11y agoTrue. But new "stuff" is coming out so fast in ML these days, that pretty much whenever you publish a comparison, a month later something new is going to drop.
- frik 11y agoTorch, the scientific computing framework for LuaJIT, is based on the great Lua language. This makes it a good choice - LuaJIT per se is a faster than any Python interpreter/VM. Caffe is C++, and the rest are Python (Neon, Theano and TensorFlow).
- argonaut 11y agoThis is a bit misleading. No framework is doing computation in the language you write your model in. The code you write in any of these frameworks just calls wrappers around the underlying C/C++/Cuda implementation.
- frik 11y agoYou are right, my statement was a bit misleading. These Lua and Python based frameworks call the high performance implementation which is coded usually in C/C++/Cuda/Fortran/etc. Nevertheless the Lua <-> C interface is lightweight and it certainly helps if the overhead is low. I know a guy who wrote a fast SAT solver in C, though called it with various slow bash scripts. I asked him why he uses bash, he mentioned it doesn't matter. In the end Microsoft's SAT solver beat him, exactly because they cared about such details too.
- argonaut 11y agoFor those interested in learning about the field: focus on the concepts, not the frameworks. This is true especially of machine learning, which is highly conceptual. I don't think these frameworks are suitable for people not familiar with the concepts. For example, I think it's a good exercise for beginners to implement a basic (no convolution) two layer neural net in plain Python using matrix operations in numpy. Regardless, these frameworks are more similar than different (one exception is Caffe, which is very focused on computer vision and makes it really easy to do stuff in that area, but makes it really hard to do anything else). In the course of exploring the area, you will probably touch all of them at some point (when you're trying to use a researcher's code).
- mziel 11y agoThis very much. Understanding the underlying concepts is crucial. Many people use out-of-the-box tools just plainly wrong, because they didn't bother with fundamentals. I know a guy who reinvented Naive Bayes and Laplace smoothing. Kudos for him, but he wasted about 3-4 months of company's time on it. That said, I would start with Python based frameworks: - Caffe is more for production, and Theano and family for experimentation. - You can play with Theano or Tensorflow using Keras which abstracts them - Python scientific stack makes it easier to visualize the results later
- emcq 11y agoOn that note, one of the best books I've read for teaching the concepts is Learning From Data by Yaser Abu-Mostafa et al. I highly recommend it as a way to refresh or introduce the fundamentals: http://amlbook.com/ http://amlbook.com/