A from scratch implementation of matrix autograd along with a simple proof of concept implementation of a deep neural network that achieves 82% accuracy on MNIST handwritten digit classification.

Clone the repo:

git clone [email protected]:jagdeepsb/autograd.git

Either install Python dependencies with conda:

conda env create -f environment.yml
conda activate autograd

or with pip:

pip install -r requirements.txt

To see a simple example of matrix autograd, you can run:

python simple_computation.py

For a simple training script of the MNIST digit classifier, run:

python train_example.py

The autograd Tensor class supports many common operations. Let a and b be two tensors of the same shape. Let c be a constant (float). This implementation supports the following:

• result = a + b
• result = a + c
• result = c + a

• result = a - b
• result = a - c
• result = c - a
• result = -a

• result = a * b
• result = a * c
• result = c * a

• result = a / c

• result = a ** c

Let W be a Tensor with shape (m x n) and x be a Tensor with shape (n x 1):

• result = W @ x

• Transpose: result = a.T
• Sum: result = a.sum
• Sigmoid: result = a.sigmoid
• ReLU: result = a.relu

Feel free to contribute by making a pull request to the repo! You will need to install the following pip packages: mypy, pytest, pytest-pylint, docformatter, yapf and make sure your contributions pass the following styling guidelines:

mypy . --config-file mypy.ini
python -m py.test  --pylint -m pylint --pylint-rcfile=.cbt_pylintrc

Additionally, run the following to format your code:

docformatter -i -r .
yapf -r --style .style.yapf .