1

I've look at the questions on here regarding the different python libraries around for deep learning and neural nets. They include:

  • Keras
  • Caffe
  • Lasagne
  • PyLearn2
  • Deepy
  • Theano
  • Torch

My understanding is that Keras and Lasagne require the user to have varying degrees of interaction with Theano.

I'm looking for the one that would be the easiest to start testing an algorithm I have already developed an apply deep learning on it in a packaged way. I am not looking to extend, modify or tweak anything. I certainly won't be doing any research in the field with it.

The only advanced feature I would need would be GPU support.

In short, what would be the best "dummy" library to test recursive convolutional neural nets?

Elliott Rogasik
  • 9
  • 1
  • None of the libraries will auto-select a network architecture for you. At the least, you will be wanting to test with variations in number of layers, types of layers (convolutional, pooling, dropout etc), hidden layer sizes, and choosing between several other hyper-params and/or variations on training. The easiest way to "auto-pick" those kinds of things is to follow a tutorial on a related problem to the one you want to solve, and keep things as similar as possible so you don't need to think about those things for now . . . – Neil Slater Jul 30 '15 at 06:53
  • . . . so it might be worth giving a brief summary of what kind of problem you are trying to solve (e.g. image classifying). A library with good tutorials and sample code related to your problem may be the best choice. – Neil Slater Jul 30 '15 at 06:56
  • @NeilSlater I am exploring the possibility of using DL on futures contracts in the financial markets. I will be using minute by minute price data, ratios, etc. All the tutorials I can find are for text recognition or image classifying, as you mentioned. I haven't been able to find one related to financial "prediction"... – Elliott Rogasik Jul 30 '15 at 13:17
  • CNN is not synonymous with Deep Learning, it is just one of the more successful designs. Do you have a plan for the convolution parts? If not, an RNN (deep or not) may be better suited. Not that this changes your question much, just you are more likely to find what you want looking up RNN and time-series than CNN with anything – Neil Slater Jul 30 '15 at 13:25
  • @NeilSlater Sorry if my question was poorly formulated. I have been planning to use RNN, specifically a variant with LTSM. Thanks for all the insights. – Elliott Rogasik Jul 30 '15 at 13:30

0 Answers0