micrograd2023 0.1.2

micrograd2023 was developed based on Andrej Karpathy micrograd with added documentations using nbdev for teachning purposes

micrograd2023

Literate Programming

flowchart LR
  A(Andrej's micrograd) --> C((Combination))
  B(Jeremy's nbdev) --> C
  C -->|Literate Programming| D(micrograd2023)

Disclaimers

micrograd2023, an automatic differentiation software, was developed based on Andrej Karpathy’s micrograd.

Andrej is the man who needs no introduction in the field of Deep Learning and Computer Vision. He released a series of lectures called Neural Network from Zero to Hero, which I found extremely educational and practical. I am reviewing the lectures and creating notes for myself and for teaching purposes.

mirograd2023 was written using nbdev, which was developed by Jeremy Howard, the man who needs no introduction in the field of Deep Learning. Jeremy also created fastai Deep Learning software library and Courses that are extremely influential. I highly recommend fastai if you are interested in starting your journey and learning with ML and DL.

nbdev is a powerful tool that can be used to efficiently develop, build, test, document, and distribute software packages all in one place, Jupyter Notebook (I used Jupyter Notebooks in VS Code). In this tutorial, you will learn how to use nbdev to develop software micrograd2023.

Features

Compared to Andrej’s micrograd, micrograd2023 has many extensions such as:

• Adding more and extensive unit and integration tests.

• Adding more methods for Value object such as tanh(), exp(), and log(). In principle, any method/function with known derivative or can be broken into primitive operations can be added to the Value object. Examples are sin(), sigmoid(), cos(), etc., which I left as exercises 😄.

• Refactoring Andrej’s demo code make it easier to demonstrate many fundamental concepts and/or best engineering practices when training neural network. The concepts/best-practices are listed below. Some concepts were demonstrated while the rest are left as exercises 😄.

  • Always implemented a simplest and most intuitive solution as a baseline to compare with whatever fancy implementations we want to achieve

  • Data preparation - train, validation, and test sets are disjointed

  • Over-fitting

  • Gradient Descent vs. Stochastic Gradient Descent (SGD)

  • Develop and experiment with different optimizations i.e. SGD, SGD with momentum, rmsProp, Adam, etc.

  • SGD with momentum

  • Non-Optimal learning rate

  • How to find the optimal learning rate

  • Learning rate decay and learning rate schedule

  • Role of nonlinearity

  • Linear separable and non-separable data

  • Out of distribution shift

  • Under-fitting

  • The importance and trade-off between width and depth of the MLP

  • Over-fitting a single-batch

  • Hyperparameter tuning and optimizing

  • Weights initialization

  • Inspect and visualize statistics of weights, gradients, gradient to data ratios, and update to data ratios

  • Forward and backward dynamics of shallow and deep linear and non-linear Neural Network

  • etc.

If you study lectures by Andrej and Jeremy you will probably notice that they are both great educators and utilize both top-down and bottom-up approaches in their teaching, but Andrej predominantly uses bottom-up approach while Jeremy predominantly uses top-down one. I personally fascinated by both educators and found values from both of them and hope you are too!

Related Projects

Below are a few of my projects related to optimization and Deep Learning:

• Diploma Research on Crystal Structure using Gradient-based Optimization SLIDES

• Deep Convolution Neural Network (DCNN) for MRI image segmentation with uncertainty quantification and controllable tradeoff between False Positive and False Negative. Journal Paper PDF and Conference Talk SLIDES

• Deep Learning-based Denoising for quantitative MRI. Conference Talk SLIDES

• Besides technical projects, I had an opportunity to contribute and engage in the whole process of 510(k) FDA clinical validation of Deep Learning-based MRI Reconstruction resulting the worlds-first fully integrated Deep Learning-based Reconstruction Technology to receive Food and Drug Administration (FDA) 510(k)-clearance for use in clinical environment. Product Page, Whitepaper HTMLs, Whitepaper PDF, and Whitepaper PDF2

  • AiCE Challenge 1: 1.5T MRI with Deep Learning Reconstruction (DLR) vs. 3T MRI

  • AiCE Challenge 2: 1.5T MRI with DLR vs. 3T MRI - beyond knee and brain MRI

  • AiCE Challenge 3: Faster and Higher Resolution MRI with DLR

  • AiCE Challenge 4: Faster MRI with DLR

How to install

The micrograd2023 package was uploaded to PyPI and can be easily installed using the below command.

pip install micrograd2023

Developer install

If you want to develop micrograd2023 yourself, please use an editable installation.

git clone https://github.com/hdocmsu/micrograd2023.git

pip install -e "micrograd2023[dev]"

You also need to use an editable installation of nbdev, fastcore, and execnb.

Happy Coding!!!

How to use

Here are examples of using micrograd2023.

# import necessary objects and functions
from micrograd2023.engine import Value
from micrograd2023.nn import Neuron, Layer, MLP
from micrograd2023.utils import draw_dot
import random

# inputs xs, weights ws, and bias b
w1 = Value(1.1)
x1 = Value(0.5)
w2 = Value(0.12)
x2 = Value(1.7)
b = Value(0.34)

# pre-activation
s = w1*x1 + x2*w2 + b

# activation
y = s.tanh()

# automatic differentiation
y.backward()

# show the computation graph of the perceptron
draw_dot(y)

# added random seed for reproducibility
random.seed(1234)
n = Neuron(3)
x = [Value(0.15), Value(-0.21), Value(-0.91) ]
y = n(x)
y.backward()
draw_dot(y)

You can use micrograd2023 to train a MLP and learn fundamental concepts such as overfilling, optimal learning rate, etc.

Good training

Overfitting

Testings

To perform unit testing, using terminal to navigate to the directory, which contains tests folder, then simply type python -m pytest on the terminal. Note that, PyTorch is needed for the test to run since derivatives calculated using micrograd2023 are compared against those calculated using PyTorch as references.

python -m pytest