🔮 centigrad

Autograd engine and neural network library based on numpy.

Inspired by Andrej Karpathy's micrograd.

Unlike micrograd which uses scalars, centigrad uses vectors.

Installation

git clone https://github.com/dinesh-GDK/centigrad.git
pip install -e centigrad

Features

centigrad has the basic building blocks to construct Fully Connected Neural Networks and 2D Convolution Neural Networks.

Layers

• Flatten
• Fully Connected
• 2D Convolution
• 2D Max Pooling
• 2D Dropout
• 2D Batch Normalization

Activations

• ReLu
• Tanh
• Softmax

Losses

• Mean Square Error
• Cross Entropy

Optimizers

• Gradient Descent

Example

Here is an example of how a model is defined in centigrad

class MnistNet(Model):
    def __init__(self):
        super().__init__()
        self.layerc1 = Conv2d(1, 2)
        self.maxpool = MaxPool2d()
        self.dropout = Dropout2d()
        self.batchnorm = BatchNorm2d(2)
        self.flatten = Flatten()
        self.layer1 = FullyConnected(338, 10)

    def forward(self, x):
        x = relu(self.layerc1(x))
        x = self.maxpool(x)
        x = self.dropout(x)
        x = self.batchnorm(x)
        x = self.flatten(x)
        x = softmax(self.layer1(x))
        return x

See demo notebook for more details

References

• micrograd