Neograd

A Deep Learning framework created from scratch with Python and NumPy

Get started

Installation

pip install neograd

PyPI

https://pypi.org/project/neograd/

Documentation

https://neograd.readthedocs.io/

Explore on Colab

https://colab.research.google.com/drive/1D4JgBwKgnNQ8Q5DpninB6rdFUidRbjwM?usp=sharing https://colab.research.google.com/drive/184916aB5alIyM_xCa0qWnZAL35fDa43L?usp=sharing

Motivation

I firmly believe that in order to understand something completely, you have to build it on your own from scratch. I used to do gradient calculation analytically, and thought that autograd was some kind of magic. So this was initially built to understand autograd but later on its scope was extended. You might be wondering, there are already many frameworks like TensorFlow and PyTorch that are very popular, and why did I have to create another one? The answer is that these have very complex codebases that are difficult to grasp. So I intend that this repository be used as an educational tool in order to understand how things work under the hood in these giant frameworks, with code that is intuitive and easily readable.

Features

Automatic Differentiation

autograd offers automatic differentiation, implemented for the most commonly required operations for vectors of any dimension, with broadcasting capabilities python import neograd as ng a = ng.tensor(3, requires_grad=True) b = ng.tensor([1,2,3], requires_grad=True) c = a+b c.backward([1,1,1]) print(a.grad) print(b.grad)

Custom autograd operations

If you wanted a custom operation to have autograd capabilities, those can be defined with very simple interface each having a forward method and a backward method python class Custom(Operation): def forward(self): pass def backward(self): pass

Gradient Checking

Debug your models/functions with Gradient Checking, to ensure that the gradients are getting propagated correctly

Highly customizable

Create your own custom layers, optimizers, loss functions which provides more flexibility to create anything you desire

PyTorch like API

PyTorch's API is one of the best and one the most elegant API designs, so we've leveraged the same

Neural Network module

nn contains some of the most commonly used optimizers, activations and loss functions required to train a Neural Network

Save and Load weights, model

Trained a model already? Then save the weights onto a file and load them whenever required or save the entire model, onto a file

Checkpoints

Let's say you're training a model and your computer runs out of juice and if you'd waited until training was finished, to save the weights, then you'd lose all the weights. To prevent this, checkpoint your model with various sessions to save the weights during regular intervals with additional supporting data

Example

```python import neograd as ng from neograd import nn import numpy as np from neograd.nn.loss import BCE from neograd.nn.optim import Adam from neograd.autograd.utils import gradcheck from sklearn.datasets import makecircles from sklearn.modelselection import traintestsplit from sklearn.metrics import classificationreport, accuracy_score

load dataset (binary classification problem)

X, y = makecircles(nsamples=1000, noise=0.05, randomstate=100) Xtrain, Xtest, ytrain, ytest = traintest_split(X,y)

numtrain = 750 # number of train examples numtest = 250 # number of test examples num_iter = 50 # number of training iterations

convert data into tensors

Xtrain, Xtest = ng.tensor(Xtrain[:numtrain,:]), ng.tensor(Xtest[:numtest,:]) ytrain, ytest = ng.tensor(ytrain[:numtrain].reshape(numtrain,1)), ng.tensor(ytest[:numtest].reshape(numtest,1))

define the structure of your neural net

class NN(nn.Model): def init(self): self.stack = nn.Sequential( nn.Linear(2,100), nn.ReLU(), nn.Linear(100,1), nn.Sigmoid() )

def forward(self, inputs): return self.stack(inputs)

model = NN() # initialize a model loss_fn = BCE() # initialize a loss function (Binary Cross Entropy) optim = Adam(model.parameters(), 0.05) # initialize an optimizer

training loop

for i in range(numiter): optim.zerograd() # zero out the gradients in the tensors outputs = model(Xtrain) # get the outputs by passing the training data to your model loss = lossfn(outputs, ytrain) # calculate the loss loss.backward() # initiate the backward pass to calculate the gradients optim.step() # update the parameters print(f"iter {i+1}/{numiter}\nloss: {loss}\n")

with model.eval(): # put the model in evaluation mode testoutputs = model(Xtest) # get the outputs of the model on test data preds = np.where(test_outputs.data>=0.5, 1, 0) # make predictions

print(classificationreport(ytest.data.astype(int).flatten(), preds.flatten())) print(accuracyscore(ytest.data.astype(int).flatten(), preds.flatten()))

gradcheck(model, Xtrain, ytrain, lossfn) # perform gradient checking in your model ```

How is this any different from

• Andrej Karpathy's micrograd
  Natively only supports scalar values for computation, whereas we support scalars, vectors, matrices all compatible with NumPy broadcasting
• George Hotz's tinygrad
  Has an obligation to be under 1000 lines of code leading to cramped up code, therefore our implementation is so much more readable and easily understandable. Also, no dealing with C/C++ code used in tinygrad for GPU acceleration
• pytorch, tensorflow, etc
  Large messy codebases written mostly in C/C++ for efficiency making it impossible to find you're way around and understand stuff. We've a pure Python implementation making it easy to get started and understand what's going on under the hood

Resources

• A big thank you to Andrej Karpathy for his CS231n lecture on Backpropagation which was instrumental in helping me gain a good grasp of the basic mechanisms of autograd
• Thanks to Terance Parr and Jeremy Howard for their paper The Matrix Calculus You Need For Deep Learning which helped me get rid of my fear for matrix calculus, that is beautifully written starting from the very fundamentals and slowly transitioning into advanced topics