embedml 0.1.3

Minimal autograd engine using numpy with minimal implementation of pyTorch or Jax like library for traning NNs.

embedML

pytorch like machine learning framework from scratch I started the repo with building a simple compiler for embeded platfroms, but later decided to make a self contained abstract baby machine learning framework.

Could be used mostly for the Educational purpose as well. Easy and full python implemtation using numpy. Smaller sibling or framworks like pytorch or jax :) and useful to see how the behind the scene of those bigger platfrom.

and it is only less than 200 lines of code!

Instal

pip install embedml

Example tarinig simiar to pytorch:

y = model(x)
loss = criterion(y, t)
loss.backward()   
opt.step()
opt.zero_grad()

or like this implement the gradient decent:

params = model.get_parameters()
for param in params:
  param -= param.grad * self.lr

Example of differentiation of a funtion

form Jax example:

from jax import grad
import jax.numpy as jnp

def tanh(x):  # Define a function
  y = jnp.exp(-2.0 * x)
  return (1.0 - y) / (1.0 + y)

grad_tanh = grad(tanh)  # Obtain its gradient function
print(grad_tanh(1.0))   # Evaluate it at x = 1.0
# prints 0.4199743

and similar (backward gradient) with embedML:

import numpy as np
from embedml.tensor import Tensor

def tanh(x):  # Define a function
    y = (-2 * x).exp()
    return (1 - y).div((1 + y))

x = Tensor(np.array(1))
y = tanh(x)
y.backward()

print(x.grad)