4 ms·
Funny implementation of a neural net from the author. I guess it teaches you how it works. https://github.com/perborgen/NeuralNetworkNoob/blob/master/net.py ht
by visarga 11y ago
Funny implementation of a neural net from the author. I guess it teaches you how it works.
https://github.com/perborgen/NeuralNetworkNoob/blob/master/net.py https://github.com/perborgen/NeuralNetworkNoob/blob/master/n...
To add something to the topic: I find that I usually don't use algorithms that I don't understand even if they are available. So, before firing up that convolutional neural network in mxnet, I need to be really comfortable with the theory.
- mrborgen 11y agoYeah, that was my very first one. I have since started with vectorized implementations instead of that clumsy way I did it the first time. However I found vectorization a bit tricky at first. I'm actually planning to try and learn convolutional nets soon, and will try to setup a small study group online (basically just a repo with materials and a Slack/Gitter chat room), where people can help each other with learning it. Do you think that would help you to learn it?
- ratsimihah 11y agoInteresting how you haven't used any machine learning framework :) Did you do it purposely to get a better understanding of neural nets? Are you going to move to frameworks to speed up your iteration speed? Or are you more interested in maybe refactoring your code into a reusable framework? :)
- mrborgen 11y agoYeah, I wanted to understand how they worked, so I stayed away from any framework. I have since moved over to using Theano, Lasagne & nolearn. However, I almost find it just as hard, because it's so easy to step wrong and get errors, and so hard to debug. In addition, Lasagne & nolearn doesn't exactly have great documentation of tutorials available (they have some, but I like to try out many tutorials to get perspective and get a proper understanding). I'm far from good enough to turn my implementations into frameworks :p
- Xcelerate 11y agoI don't think searching for and showcasing someone's old, exploratory code really demonstrates much. A lot of my initial code for studying new concepts is atrocious — inefficient, memory-wasting, overly verbose; the point is to just get something that works and then later it can be refined, packaged up into modules, decoupled, optimized, etc. Particularly for numerical and scientific work, there is a huge tradeoff between code cleanliness/readability and performance. I normally start by just copying the "demo algorithm" from a research paper straight into whatever language I'm using. The goal is to simply get the code working properly, and then I begin tweaking it for speed. Once it's optimized though, the result usually looks nothing like the original algorithm and it is difficult to tell what the code is doing just by looking at it.
- coldtea 11y ago>To add something to the topic: I find that I usually don't use algorithms that I don't understand even if they are available. On the contrary, as everybody else does, you too do it all the time. There are thousands of algorithms hidden behind higher level APIs you use, which you don't understand.