jacobian

a basic keras-like neural network library for c++/python
Log | Files | Refs | README

readme.md (2625B)

      1   <!-- readme.md -->
      2   <!-- Jacobian -->
      3 
      4   <!-- Markdown has the worst comment syntax I've ever seen. Seriously, what is this. -->
      5   <!-- TODO: Migrate this README to Org Mode. -->
      6 
      7 ![Banner](./pictures/banner.png)
      8 
      9 ## About
     10 Jacobian is a work-in-progress machine learning library written in C++
     11 designed to run as fast as possible while still being simple to
     12 use. Jacobian is accessible via Python and enables you to write models that
     13 train faster with the same amount of code. As of now, Jacobian supports
     14 feedforward neural networks and has partial support for convolutional
     15 neural networks.
     16 
     17 ## Usage
     18 
     19 Initializing and training a neural network with Jacobian takes just 8 lines
     20 of code!
     21 
     22 ```python
     23 import jacobian as jcb
     24 net = jcb.Network("./data_banknote_authentication.txt", 10, 0.0155, 0.03, jcb.L2, 1, 0.9)
     25 net.add_layer(4, jcb.activations.linear, jcb.activations.linear_deriv)
     26 net.add_layer(5, jcb.activations.sigmoid, jcb.activations.sigmoid_deriv)
     27 net.add_layer(2, jcb.activations.linear, jcb.activations.linear_deriv)
     28 # Optional: net.init_optimizer(jcb.optimizers.momentum(0.1))
     29 # Optional: net.init_decay(jcb.decays.exponential(1, 0.5))
     30 net.initialize()
     31 for i in range(50):
     32   net.train()
     33 ```
     34 ## Examples
     35 
     36 See `example.cpp` for an example of using Jacobian from C++, and
     37 `example.py` for an example of using Jacobian from Python.
     38 
     39 ## Building
     40 
     41 ### Dependencies
     42 Eigen 3 is the only dependency, although building the python library
     43 requires `pybind11`.
     44 
     45 ### Building with CMake
     46 
     47 **Don't forget to delete `CMakeCache.txt` after each compilation if you
     48 plan on switching things up!**
     49 
     50 There are two target languages, five main build configurations, and a
     51 number of toggleable build 'attributes'. The preferred target language can
     52 be specified by setting a CMake variable from the command-line: `-DCXX=ON`
     53 or `-DPYTHON=ON`.
     54 
     55 The five main configurations correspond to differing levels of optimization.
     56 
     57 - `cmake .`: No compiler optimizations.
     58 - `cmake . -DFAST=ON`: Enables the O3 optimization layer in the compiler.
     59 - `cmake . -DFASTER=ON`: Enables O3 as well as extra individual flags.
     60 - `cmake . -DTRADEOFFS=ON`: All previous optimizations as well as ones that
     61   sacrifice precision.
     62 - `cmake . -DRECKLESS=ON`: Like `TRADEOFFS`, but defines the RECKLESS macro
     63   (and NDEBUG) which skips all checks within the code.
     64 
     65 One you've selected a main optimization level, extra configurations can be
     66 passed in.
     67 
     68 - `-DDEBUG=ON` enables debugging features in the compiler (and shows
     69   warnings).
     70 
     71 A sample build process would look like this:
     72 
     73 ```sh
     74 rm CMakeCache.txt && cmake . -DPYTHON=ON -DFASTER=ON -DDEBUG=ON && make
     75 ```