Deki is a library is to create beautiful ✨interative✨ visuals to understand your optimizer better, because seeing is believing!
And the best part, it is extremely simple to use! Just a few steps and you'll find your visuals ready to be played with.
$ pip install dekiDeki works by taking in torch.optim.Optimizer classes and a configuration dictionary for the hyperparameters and other keyword arguments, like in the example below 🔻
from torch.optim import SGD
from deki import Rosenbrock
optimiser = SGD
optimiser_state = {"lr":1E-3}
steps = 1000
Rosenbrock.plot((optimiser, optimiser_state), steps)
Just like its mentioned above, any PyTorch optimizer class can be used as long as its required keyword args can be passed in a dict. In conclusion, Deki can handle your custom optimizers, so build away!
from torch.optim import Optimizer
from deki import Rosenbrock
class MySpecialOptimizer(Optimizer):
...
optimizer = MySpecialOptimizer
optimiser_state = {"lr":1E-3}
steps = 1000
Rosenbrock.plot((optimiser, optimiser_state), steps)Not only this, Deki also manages to make its Visualizer general so you can implement your custom test functions super easily, just like in the following manner 🔻
import torch
from torch.optim import SGD
from deki import Visualizer
# you can define any function as long as it takes a X
# wherein X will be torch Tensor tuple of x,y
def my_special_test_function (X : Tuple[torch.Tensor, torch.Tensor]):
x, y = X
# as an example we will return the sphere function
# use only torch ops to maintain differentiability
return torch.pow(x, 2) + torch.pow(y, 2)
# Define the object for the Visualizer class
# For supported functions, such objects come pre-defined in deki
visualizer = Visualizer( my_special_test_function,
x_range = (-2,2),
y_range = (-2,2),
minima = (0, 0),
start = (-1.5, 1)
)
# Now call the plot function with the optimizer to get the plot
visualizer.plot((SGD, {"lr":1E-3}), steps=1000)
| Function | Example |
|---|---|
| Rosenbrock | |
| Sphere |
A special thanks to all the contributors on this project for making it what it is today! ❤️
If you found this repository helpful, please leave a star! ⭐
If you are using this repository for research, please cite it in the following manner:
@misc{minhas_2023,
url={https://github.com/Dawn-Of-Eve/deki},
title={Deki: Library for Visualizing Optimization Algorithms},
publisher={Dawn Of Eve},
author={Minhas, Bhavnick Singh},
year={2023},
month={Feb}}