This repository provides an implementation of a MaxPool2D (2D MaxPooling layer) from scratch using NumPy.
- 2D maxpooling layer implementation
- Support for both single-channel and multi-channel images/feature maps
- Customizable filter size, stride, and padding
- Efficient computation using NumPy
-
Python 3.12.3
-
Install requirements:
pip install -r requirements.txt
def maxpool2d(
image: np.ndarray,
kernel_size,
stride=None,
padding=0,
) -> np.ndarray:
"""
Perform a 2D MaxPool operation
Args:
image (np.ndarray): Input image
kernel_size (int or tuple[int, int]): Size of the maxpooling window
stride (int, optional): Stride value for the maxpooling operation. Default value is kernel_size.
padding (int, optional): Padding value for the input image. Default is 0.
Returns:
np.ndarray: Resulting output of the maxpooling operation.
Raises:
TypeError: If `image` is not of type `numpy.ndarray`.
ValueError: If `kernel_size` is invalid.
"""| Terms | Explainations | Variables |
|---|---|---|
| input | An image of size (height, width, channels) represents a single instance of an image. It can be thought of as a collection of channels 2D matrices, each of size (height, width), stacked together. |
in_channel=channels |
| padding | Technique of adding extra border elements to the input data before applying a convolution operation. It helps preserve spatial dimensions and prevents the output from being smaller than the input. | padding |
| kernel | TODO | kernel_size, stride |
| maxpool | TODO | |
| ouput | TODO | out_channels |
output_height = (input_height - kernel_height + 2 * padding) / stride + 1
output_width = (input_width - kernel_width + 2 * padding) / stride + 12 maxpool2d layers |
3 maxpool2d layers |
|---|---|
kernel_size = 2, stride = 2, padding = 0 |
kernel_size = 2, stride = 2, padding = 1 |
 |
 |