Non-Subkernel Convolution
[PyTorch, Pandas, CNN, Python]
A custom Conv2D layer as an alternative to standard convolution, reducing trainable parameters by 85% while maintaining competitive accuracy on a 3-class image classification task.
Designed and implemented a custom Conv2D layer as an alternative to standard convolution, conducting comparative analysis across multiple model configurations.
- Reduced trainable parameters by 85% while maintaining competitive accuracy on a 3-class image classification task (Dog vs. Cat vs. Bird).