This package provides an interface to experiment with the design of Gradient descent algorithms:
- Gradient descent
- Stochastic gradient descent
- Line search
- Conjugate gradient descent
- Momentum
- Nesterov accelerated gradient
- AdaGrad
- AdaDelta
- ...
All algorithms can be easily vizualized as 2d animations.
This package relies on Numpy & Scipy for computations and on Matplotlib for visualization.
To load the real toy-example datasets, this package uses Pandas and Scikit-learn, but both dependencies are only required if the datasets are loaded.
Run pip install -r requirements.txt to install the dependencies.
To run the tutorial demos, run python anim.py --demo N where N can be an integer from 1 to 10.
Note that the demos have been tested on Mac OS X. The backend used by matplotlib might need to be changed to work as expected on other platforms.