microdist

Note: this is still WIP, don't expect anything just yet.

The tiny dog from micrograd, combined with Ray, makes up the mascot for microdist.

microdist is a tiny distributed deep learning framework (with a bite, and a stare).

It slightly modifies the ~150 lines of code of micrograd,
by keeping the core engine untouched, and adding another ~100 LOC to the neural network layer
to introduce basic parameter servers using Ray.
The changes to the networks itself are just needed to get and send parameters to a server, resp. set them again after a training step.
We introduce one single dependency (Ray) and otherwise use plain Python, which all comes together at around ~250 LOC.
This might be useful for demo purposes or to teach the fundamentals of distributed training.

Installation

git clone https://github.com/maxpumperla/microdist
cd microdist
python setup.py install

Example usage

You can create some sample data, define a tiny micrograd model with two input neurons, two hidden layers of length 16 each and an output layer with a single neuron, and train this model distributed across 4 workers, communicating with a "master"
model via a parameter server by calling:

from microdist import Model
from sklearn.datasets import make_moons

X, y = make_moons(n_samples=100, noise=0.1)
y = y*2 - 1

model = Model([2, 16, 16, 1])

model.dist_fit(X, y, steps=10, num_workers=4)

The notebook demo.ipynb provides a full demo of training an
2-layer neural network (MLP) binary classifier.

Running tests

To run the unit tests you will have to install PyTorch,
which the tests use as a reference for verifying the correctness of the calculated
gradients. Then simply:

python -m pytest