NanoTorch

NanoTorch is a deep learning library (micro-framework) inspired by the PyTorch framework, which I created using only Math and Numpy :). My purpose here is not to create a powerful deep learning framework (maybe in the future), but solely to understand how deep learning frameworks like PyTorch and TensorFlow work behind the scenes.

Neural Networks:

Let's explore an example of building a simple neural network (essentially a Linear Regression model) with NanoTorch:

import nanotorch as nnt
import nanotorch.nn as nn 

Let's build a simple model:

class MLPModel(nn.Module):

    def __init__(self):

        self.fc = nn.Sequential(
            nn.Linear(3, 3),
            nn.Sigmoid(),
            nn.Linear(3, 5),
            nn.Sigmoid(), 
            nn.Linear(5, 1)
        )

    def forward(self, input: nnt.Tensor) -> nnt.Tensor:
        return self.fc(input)

Let's generate a simple dataset, using the nn.rand function and the nnt.dot operation:

X = nnt.rand(100, 3)
y = nnt.dot(X, nnt.Tensor([1, -2, 3]).T)    

Now, let's create an instance of MlpModel

model = MLPModel()

We are dealing with regression task. So, the nn.MSELoss is chosen

mse = nn.MSELoss(model.layers())

Let's define the stochastic gradient descent optimizer

opt = nnt.optim.SGD(model.layers(), lr=0.001)

Finally, The training loop

for epoch in range(30):

    for xi, yi in zip(X, y):

        opt.zero_grad()

        y_predi = model(nnt.Tensor(xi))

        loss = mse(y_predi, nnt.Tensor(yi))

        loss.backward()

        opt.step()

print(model.layers()[0].parameter)

The output is as follows:

[[1.00772484]
 [1.98651816]
 [3.04503581]]