Hi, thanks for checking out mikerograd! This is a Julia version of Andrej Karpathy's original micrograd Python library, with an additional Tensor object-type added.
I started this project as a learning exercise for myself, since I'm a beginner at Julia and machine learning. Figured other people might find it useful too, so I'm putting it up on my Github.
Basically the goal here is to make some basic gradient-tracking objects, with code that's simple enough to read, understand, and edit.
The file mikerograd.jl contains 2 object-types (basically classes):
Value
- basically a Julia version of the Value class in Karpathy's original micrograd.
- scalar-valued gradient tracking (meaning operations are done on single numbers, not arrays)
- great tool for learning the basics of backpropagation
- possibly useful for simple examples like fitting a line to data by gradient descent.
Tensor
- something new that I decided to add myself
- tensor-valued gradient tracking (arrays and matrices, rather than single numbers)
- faster than the Value class, and a bit more similar to how real ML libraries actually work.
- still simple enough to read and understand the code
- useful for some basic neural net applications, like MNIST
Please see the demo.ipynb notebook for an introduction on how to use these objects. It includes a basic introduction, as well as line-fitting and MNIST examples.
Anyway, I hope some people will find this helpful as a learning tool for Julia and machine learning! If you have any questions or run into problems with the code, please feel free to email me at [email protected].
Citations (and great sources for learning the basics of neural nets in Python):