Show HN: Syna – Minimal ML and RL Framework Built from Scratch with NumPy

2 weeks ago 1

Syna is a lightweight machine learning framework inspired by DeZero. Built from scratch using only NumPy, it follows a define-by-run (dynamic computation graph) approach and includes a basic reinforcement learning framework.

Unlike most frameworks that implement reinforcement learning as a separate library, Syna provides everything in a single library.

Designed for beginners and researchers, Syna helps you learn the fundamentals of machine learning and the inner workings of frameworks like PyTorch. Performance is not the focus, and GPU support is intentionally omitted to keep the code simple and easy to understand.

Get the Syna Source

git clone https://github.com/sql-hkr/syna.git cd syna uv venv source .venv/bin/activate uv sync

Or, from PyPI:

Important

To visualize the computation graph, you need to install Graphviz.

brew install graphviz # macOS sudo apt install graphviz # Linux

Visualize the computation graph for the fifth derivative of tanh(x) with respect to x.

import syna import syna.functions as F from syna.utils import plot_dot_graph x = syna.tensor(1.0) y = F.tanh(x) x.name = "x" y.name = "y" y.backward(create_graph=True) iters = 4 for i in range(iters): gx = x.grad x.cleargrad() gx.backward(create_graph=True) gx = x.grad gx.name = "gx" + str(iters + 1) plot_dot_graph(gx, verbose=False, to_file="tanh.svg")

The output graph is shown below.

Solve CartPole-v1 using the DQN algorithm.

from syna.algo.dqn import DQNAgent from syna.rl import Trainer trainer = Trainer( env_name="CartPole-v1", num_episodes=300, agent=DQNAgent( lr=3e-4, ), ) trainer.train()