A lightweight autograd engine with a deep learning library and a Torch-like API, built from scratch.

Python ≥ 3.9

pip install synapgrad

This project implements a completely functional engine for tracking operations between Tensors, by dynamically building a Directed Acyclic Graph (DAG), and an automatic gradient calculation and backpropagation algorithm (reverse-mode autodiff) over this DAG.

Built on top of the engine, the deep learning library implements the most common functions, layers, losses and optimizers in order to create AI models able to solve real problems. This library mimics PyTorch in a simplified way, with the same API style and function names.

The objective of this project is to develop a lightweight deep learning library entirely from scratch, without relying on any existing frameworks (such as Keras, PyTorch, TensorFlow, scikit-learn, etc.), using only the numpy library.

At the moment, supporting GPU execution is not within the scope of the project

Automatic gradient calculation and backpropagation algorithm

numpy>=1.23.5 # Core
graphviz>=0.20.1 # (Optional) Visualize DAG (https://graphviz.org/download/)

In the examples/visualize_graph.ipynb notebook there is an example of how to display the graph that synapgrad creates in the background as operations are chained:

import synapgrad

with synapgrad.retain_grads():
    x1 = synapgrad.tensor([[-5.0, 3.0], [2.0, -4.0]], requires_grad=True)
    x2 = synapgrad.tensor([[6.0, 0.4], [1.9,  2.0]], requires_grad=True)
    x3 = synapgrad.tensor(3.0, requires_grad=True)
    y = synapgrad.addmm(x3, x1, x2) # x3 + x1 @ x2
    z = x2.sqrt() @ y
    z.backward(synapgrad.ones(z.shape))
    
# graphviz is required to draw the graph
z.draw_graph()

Built on top of the engine, synapgrad has a deep learning library that implements the following features:

• Weight initialization: Xavier Glorot uniform, Xavier Glorot normal, He Kaiming uniform, He Kaiming normal
• Activations: ReLU, LeakyReLU, SELU, Tanh, Sigmoid, Softmax, LogSoftmax
• Layers: Linear, Unfold, Fold, Flatten, Dropout
• Convolutions: MaxPool1d, MaxPool2d, AvgPool1d, AvgPool2d, Conv1d, Conv2d
• Normalizations: BatchNorm1d, BatchNorm2d
• Optimizers: SGD, Adam, AdamW
• Losses: MSELoss, NLLLoss, BCELoss, BCEWithLogitsLoss, CrossEntropyLoss

This project includes three Jupyter notebooks (located in examples/) that tackle three beginner-level AI problems:

• 1. Basic MLP for binary classification (sklearn 'make_moons' toy dataset)
• 2. MLP for handwritten digits classification (MNIST dataset)
• 3. CNN for handwritten digits classification (MNIST dataset)

To use the SynapGrad training utilities, install the optional requirements by running:

pip install synapgrad[train]

Additionally to run the notebooks in /examples you will need to install:

torchvision
# **** torch required only if you want to use torch engine instead of synapgrad's *****
torch>=1.12.1 # Install following the instructions in https://pytorch.org/

To measure the efficiency of synapgrad, all three examples were compared to PyTorch. All training sessions were conducted on a laptop with an Intel Core i7 10th generation processor and 16 GB of RAM.

To run the unit tests you will have to install Pytest and PyTorch, which is used to check whether the gradients are calculated correctly

pytest==7.3.1
torch>=1.12.1 # Install following the instructions in https://pytorch.org/

Run all the tests:

python -m pytest

To compare the speed of torch and synapgrad in each operation, run the command below:

python -m pytest ./tests/test_ops.py -s