Maxine

Named after Matrix Chain (Matrix Multiplication in series), after all thats what Neural Nets are.

This project is a fun implementation of Neural Network to understand the under-the-hood working.

• Optim class -- Optimizer
• Model class -- Neural Network model
• MLOPS -- Operators required for Deep learning, like Sigmoid, ReLU etc

Here are the steps to follow:

1. Write a Model
2. define Loss function
3. Create an Optimizer such as SGD or Adam
4. Load the data to tensor in a Data Loader
5. write a Training Loop
6. train it
7. Train MNIST subset- for eg, 3 vs 7
8. Train full MNIST
9. Utilities-- Saving Loading pickel file
10. Expand

Libraries used-

• Numpy
• Pytorch's Tensor, only for calculating Gradient
• tqdm

A much later goal in this project-

• remove pytorch dependency (gradient)
• Create tensor module

Here is a demo training mnist dataset

NN Module

import maxine.nn as nn

x = torch.randn(60000, 1, 28, 28)
x = x.view(x.size(0), -1)

ih = nn.Linear(784, 10)
ih.forward(x)

Losses

TODO

Metrics

from maxine.metrics import accuracy

a = torch.randn(3,5).normal_(0,1)
b = torch.Tensor([2, 1, 5])
accuracy(a, b)
# output:
# tensor(0.3333)

Installation:

pip install -r requirements.txt

pip install -r requirements-dev.txt

Install the project:

pip install -e .