NumPy Neural Network Module

[](https://numpy.org/ | https://scipy.org/ | https://matplotlib.org/)

A neural network built from scratch using only NumPy and SciPy. This has convolutional, dense, and recurrent layers, with support for mixing them. It also includes a selection of commonly used activation and cost functions, as well as popular optimization algorithms. I tried to make it easily extensible, allowing you to implement their your own custom layers and optimizers.

Features

• Layer Types:
  • Convolutional Layers (Also supports convolutions of 4D+ dimensions)
  • Dense (Fully Connected) Layers
  • Recurrent Layers (e.g., SimpleRNN, LSTM)
  • Utility (Pooling, Reshape, Flatten, Permutate) Layers (May raise errors)
• Layer Mixing: Ability to combine different layer types to create complex architectures.
• Activation Functions:
  • Sigmoid
  • ReLU
  • Tanh
  • Softmax
  • (and more...)
• Cost Functions:
  • Mean Squared Error (MSE)
  • Binary Cross-Entropy
  • Categorical Cross-Entropy
  • (and more...)
• Optimizers:
  • Adam
  • AdamW
  • AmsGrad
  • SGD
  • SGD with Nesterov Momentum
  • RMSprop
• Extensibility:
  • Base classes for creating custom layer types.
  • Base classes for implementing custom optimization algorithms.
• Dependencies: Only relies on NumPy, SciPy and Numba, but for running the test files matplotlib is needed