1D CNN From scratch (NumPy)

My OOP implementation of a 1D CNN (NumPy) with forward propagation and backpropagation (gradients computed by chain-rule).

Note: This is not a PyTorch or TensorFlow implementation — all layers, activations, loss functions, and gradient computations are manually coded for educational purpose.

---

🚀 Features

• 1D Convolution (Conv1D)
• Max & Average Pooling (MaxPool1D, AvgPool1D)
• Fully Connected Layers (FCNN)
• Activations: ReLU, LeakyReLU, Sigmoid, Swish
• Loss Functions: Squared Error Loss (MSE), Binary Cross-Entropy (planned)
• Forward and Backward Pass: gradients computed by chain rule
• Flatten Layer: Reshapes feature maps for FCNN
• Utility Function: Padding, logging
• Future additions:
  • Batching
  • Support for multiple channels
  • Xavier/He initialization
  • Gradient clipping & numerical checks
  • Experiment tracking (W&B)
  • Vectorization
  • Tensor implementation wth CUDA support

---

📁 Repository Structure

.
├── activations                 #nn activations
│   ├── LeakyReLU.py
│   ├── ReLU.py
│   ├── Sigmoid.py
│   └── Swish.py
├── config.py                   #project-wide config
├── layers                      #nn layers
│   ├── AvgPool1D.py
│   ├── Conv1D.py
│   ├── FCNN.py
│   ├── Flatten.py
│   ├── MaxPool1D.py
├── loss                        #loss functions
│   ├── BCELoss.py
│   ├── MSELoss.py
│   └── SquaredErrorLoss.py
├── main.py
├── readme.md
├── simple_cnn.py               #1d cnn example
└── utils
    ├── logging_helper.py       #helper function for logging
    ├── pad_input.py
    └── plot_loss_curve.py

⚡ Usage Example

A working example is provided in simple_cnn.py

Installation

1. Clone this repo: git clone https://github.com/rizanB/from_scratch.git
2. Install dependencies with pip or conda: numpy matplotlib timeit