This project is a simple fully connected neural network written from scratch in C.
The network is trained to learn facial image patterns and generate new 28×28 RGB face images. It is a small experimental model created for learning and exploring the fundamentals of neural networks and image generation.
It is not intended to produce hyper-realistic images. The generated faces are simple and low-resolution, but the model demonstrates how a neural network can learn visual patterns from training data and generate new outputs.
- Architecture: Fully Connected Autoencoder (Encoder + Decoder)
- Network:
2352 → 256 → 32 → 256 → 2352 - Training Dataset: 8,000 images
- Training Image Size: 28×28 RGB
- Input: 28×28 RGB image
- Output: 28×28 RGB image
- Hidden Layer 1: 256 neurons
- Latent Layer: 32 neurons
- Hidden Layer 2: 256 neurons
- Total Parameters: 1,223,504
- Total Weights: 1,220,608
- Total Biases: 2,896
- Encoder Parameters: 610,592
- Decoder Parameters: 612,912
- Model Size: ~4.67 MiB
- Data Type: 32-bit floating point (
float) - Activation: Sigmoid
- Model File:
model.bin - Implementation: C
- Framework: None
Compile with GCC:
cc model.c -o model -lm
